**The Association between Utilization of Media Information and Current Health Anxiety Among the Fukushima Daiichi Nuclear Disaster Evacuees**

**Masatsugu Orui 1,2,\* , Chihiro Nakayama <sup>1</sup> , Yujiro Kuroda <sup>3</sup> , Nobuaki Moriyama <sup>1</sup> , Hajime Iwasa 1,4, Teruko Horiuchi <sup>1</sup> , Takeo Nakayama <sup>5</sup> , Minoru Sugita <sup>6</sup> and Seiji Yasumura <sup>1</sup>**


Received: 19 May 2020; Accepted: 29 May 2020; Published: 1 June 2020

**Abstract:** The 2011 nuclear disaster in Fukushima was not only a health disaster, but also an information disaster. Although media can promote health communication following disasters, studies have revealed associations between media information and negative psychological reactions. To clarify the relationship between media utilization and current health anxiety due to radiation exposure, a cross-sectional questionnaire survey was conducted in Fukushima. We selected 2000 subjects from evacuation (i.e., 500) and non-evacuation (i.e., 1500) areas by two-stage stratified random sampling. As the independent variable, participants were asked about current health anxiety due to radiation exposure at the time of answering the questionnaire. For utilization of media about radiation exposure, local media, national media, Internet media, public broadcasts, and public relations information from local government were set as the dependent variables. Questionnaire data were analyzed by evacuation type (i.e., forced/voluntary). In a multivariate logistic regression analysis, the use of public relations information was significantly associated with lower anxiety for the forced evacuees (odds ratio: 0.72; 95% confidence interval: 0.56–0.93). Our findings highlight the importance of public relations information from local government in terms of it being associated with lower current health anxiety, and this could potentially aid in preparing for future disasters.

**Keywords:** Fukushima nuclear accident; mass media; Internet; public health practice; community mental health services

#### **1. Introduction**

The Great East Japan Earthquake, which occurred on 11 March 2011, was the largest earthquake ever recorded in Japan's history. The earthquake (magnitude 9.0) generated a massive tsunami that caused enormous damage to the Pacific Coast. This was followed by a separate tsunami, which hit the Fukushima Daiichi Nuclear Power Plant operated by the Tokyo Electric Power Company, causing radiation disasters in Fukushima Prefecture and requiring the long-term evacuation of residents from many surrounding municipalities. Due to this triple disaster, more than 92,000 residents who lived in an area designated by the national government as an evacuation area were forced to leave their homes (as of May 2016) [1]. Moreover, some residents decided to evacuate voluntarily to avoid the effects of the nuclear disaster, even residents who lived in non-evacuation areas.

The nuclear accident at the Fukushima Daiichi Nuclear Power Station caused multiple public health problems, including increased anxiety and mental health issues due to perceived risk among the evacuees and residents of Fukushima. In addition, Yamashita, who supported the nuclear accident response on-site as a radiation specialist, have argued that the Fukushima event was not only a health disaster, but also an information disaster [2], because the accident was an unprecedented experience for evacuees, and their perceived radiation exposure risk may have been related to the mass media. Consequently, their disaster-related stress and/or psychological distress levels may have been affected [3]. Indeed, newspaper coverage of the accident focused mainly on the crisis response relating to immediate issues, actions, and decisions in the aftermath of the accident (e.g., information of on-site actions undertaken, communications about the INES (International Nuclear Event Scale), food restrictions, cost, and number of people affected and being evacuated) [4].

One of the recommendations of the Chernobyl Forum report was to address the lack of accurate information available to local populations on the health risks as a result of the disaster itself, as well as wider health risks such as non-communicable diseases [5]. Moreover, the United Nations Sendai Framework for Disaster Risk Reduction aims to understand disaster risk while sharing non-sensitive information and appropriate communications, and to strengthen the utilization of media, including social media and traditional media [5]. In fact, the media functioned as a form of interpersonal communication with others, or as a channel for local government and other organizations during the immediate aftermath of the Great East Japan Earthquake [6]. However, the media are not always helpful. Several studies have examined disaster-related television viewing in the context of terrorism and have explored a range of outcomes, including post-traumatic stress disorder (PTSD), depression, anxiety, stress reactions, and substance use [7]. One study reported a significant association between the consumption of television and Internet coverage of the 2011 Great East Japan Earthquake and Tsunami, and post-traumatic reactions [8]. This suggests that the media (including Internet and television) can trigger negative psychological responses in evacuees and residents who use it.

Against this backdrop, the present study aims to clarify the association between media utilization (e.g., Internet media, public relations information from local government, and other traditional media) and current strong health anxiety at the time of the survey in the context of the Fukushima nuclear disaster, in order to consider effective modes of disaster communication among evacuees. These findings will likely be useful for future disaster risk reduction and management.

#### **2. Materials and Methods**

#### *2.1. Participants*

This cross-sectional questionnaire survey targeted 2000 residents of Fukushima Prefecture aged 20–79 years. Participant selection was based on two-stage stratified random sampling (stage one, survey of the region; stage two, survey of individuals). A random selection occurred of 33–34 individuals per point from municipal resident registration files to obtain 2000 representative participants. Of the 2000 subjects, 500 were from the three types of evacuation area that the Japanese government designated according to spatial radiation dose rates, as follows: (1) difficult-to-return areas, with a radiation dose rate ≥50 millisieverts (mSv) per year; (2) residence restriction areas, with a radiation dose rate ≥20 and <50 mSv per year; and (3) areas where evacuation orders were ready to be lifted as of 22 April 2011. The remaining 1500 people lived in the non-evacuation areas of Fukushima Prefecture (500 people were selected from each of the three areas of Hama-Dori, Naka-Dori, and Aizu) (Figure 1). We sent an anonymous, self-reporting postal questionnaire to participants between August and October 2016. The survey was approved by the ethics review committee of Fukushima Medical University on 12 April 2016 (approval number: 2699).

**Figure 1.** Evacuation and non-evacuation areas in Fukushima. Regions colored in dark gray correspond to the municipalities where evacuation orders were issued. Hama-Dori, Naka-Dori, and Aizu were the non-evacuation areas.

#### *2.2. Survey Variables*

For the independent variable, i.e., current anxiety regarding perceived radiation health risks, participants were asked to subjectively rate at the time of answering the questionnaire "Your current level of anxiety about the effects of radiation on your health due to the nuclear disaster" on a five-point scale: "Not at all," "Only a little," "Somewhat," "Very," and "Extremely." "Very" and "Extremely" were categorized as the "current strong anxiety group," with the other levels of anxiety as the "no or weak anxiety group." This questionnaire was investigator-designed.

For utilization of media about radiation, respondents selected up to three items from the following 13 options: local newspapers, national newspapers, NHK (Nippon Hoso Kyokai) television (public broadcast television, both national and local), private local broadcast television, private national broadcast television, radio, Internet news, Internet sites/blogs, social network services (SNS), magazines/books, public relations information from local government, word of mouth, and none of the above. To assess the association between media utilization and current strong health anxiety, we categorized "any local media (local newspapers and broadcasting)," "any national media (national newspapers and broadcasting)," "public broadcasting (NHK)," "any Internet media (Internet news, Internet sites/blogs, SNS)," and "public relations information from local government" as dependent variables, since these types of media were utilized by a relatively large number of respondents.

Regarding current health anxiety due to radiation exposure, participants were asked about: (1) anxiety related to delayed effects (e.g., severe diseases) with the statement "I am worried I might suffer from serious diseases due to the influence of radiation in the future"; (2) anxiety related to unhealthy status with the statement "Every time my condition gets worse, I become anxious about radiation exposure"; (3) anxiety related to genetic effects with the statement "I am worried that the influence of radiation will be inherited by the next generation, such as my children and grandchildren"; and (4) anxiety relating to broadcasting about nuclear issues with the statement "Looking at reports on nuclear power plant accidents, I become very anxious." These four single-item questions were part of a reliable questionnaire regarding radiation anxiety (i.e., the 7-item Radiation Anxiety Scale developed by Umeda et al. [9] and presented by Fukasawa et al. [10]). The Cronbach's alpha coefficient of the scale has been reported as 0.81, and in the present study sample, it was 0.84.

The other questionnaire than the 7-item Radiation Anxiety Scale was investigator-designed. All questionnaire items were shown in a previous report presented by Nakayama et al. [11].

#### *2.3. Statistical Analysis*

Data were also analyzed by evacuation type: (1) forced evacuation (or forced evacuees), which refers to evacuation due to living in an area designated by the national government as an evacuation area, as of 11 March 2011; and (2) voluntary evacuation (or voluntary evacuees), which refers to voluntary evacuation to avoid the effects of the nuclear disaster, even among residents living in non-evacuation areas, as of 11 March 2011. The chi-square test and multivariate logistic regression analysis were used to examine the association between media utilization and current strong health anxiety due to the nuclear disaster, as well as the characteristics of current strong health anxiety among evacuees by evacuation type. Statistical significance was evaluated using two-sided, design-based tests with a 5% level of significance. All statistical analyses were performed using SPSS 23.0 (IBM Corp., Armonk, NY, USA).

#### **3. Results**

#### *3.1. Participants*

We sent out 1985 questionnaires (excluding those returned to the sender due to no one residing at the address) and received 916 responses from August to December 2016 (response rate, 46.1%). After excluding 55 respondents who failed to provide information regarding sex or age, as well as 636 respondents who were not evacuees or did not answer a question about relocation due to nuclear disaster, the final study population consisted of 225 respondents who were either forced (*n* = 156) or voluntary (*n* = 69) evacuees (Figure 2).

**Figure 2.** Sample selection in the evacuation and non-evacuation areas. The analyzed subjects included 156 forced evacuees and 69 voluntary evacuees.

### *3.2. Respondent Characteristics*

The proportions of respondents aged 65 years and older, of respondents with a junior/senior high school education, and of respondents who were unemployed were higher among forced evacuees compared to voluntary evacuees (Table 1).


**Table 1.** Basic characteristics of the participants (forced/voluntary evacuees).

### *3.3. Utilization of Media Relating to Nuclear Exposure*

The type of media with the highest utilization rate was any local media (69.8%), followed by public broadcasting (NHK) (45.3%), and then public relations information from local government (44.0%). There was no significant difference in the utilization of local, national, or public broadcasting (NHK) between forced and voluntary evacuees. In contrast, the utilization rate of Internet media and public relations information from local governments differed significantly between forced and voluntary evacuees (Table 2). Moreover, the characteristics of the users of media relating to nuclear exposure are shown in Supplementary Tables S1 and S2. The Internet media users in this study tended to be of a younger generation and of a higher educational level than users of the other types of media. Furthermore, the proportion of those who utilized any Internet media among voluntary evacuees was higher in comparison to forced evacuees.



#### *3.4. Specifics of Current Strong Anxiety*

The proportion of respondents with current strong health anxiety due to radiation exposure at the time of answering the questionnaire was 20.3% (43/223). Among evacuees who expressed current health anxiety at the time of answering the questionnaire, most were concerned about the delayed effects (92.9%), the genetic effects (92.9%), and the broadcasting about nuclear issues (95.3%). Only anxiety about unhealthy status was of relatively low concern among evacuees (63.4%). The proportion of evacuees with these concerns was significantly higher among those who expressed current strong health anxiety compared to those who did not. There was no significant difference between forced and voluntary evacuees (Table 3).

**Table 3.** Characteristics of current anxiety (forced/voluntary evacuees).


#### *3.5. Association between Current Strong Health Anxiety and Utilization of Media Information*

Among all evacuees, a significant negative association was observed between utilization of public relations information from local government and current strong health anxiety at the time of answering the questionnaire. Among the voluntary evacuees, there was a non-significant trend between utilization of Internet media and current strong health anxiety (Table 4).


**Table 4.** Association between current strong anxiety and media utilization (forced/voluntary evacuees).

In the multivariate logistic regression analysis, utilization of public relations information from local government was significantly associated with lower current strong health anxiety at the time of answering the questionnaire among all evacuees (odds ratio (OR): 0.76; 95% confidence interval (CI): 0.61–0.94) and among forced evacuees (OR: 0.72; 95% CI: 0.56–0.93). However, public broadcasting (NHK) showed a non-significant relation between utilization and lower current health anxiety (OR: 0.85; 95% CI: 0.69–1.04). Moreover, there was a non-significant trend between utilization of Internet media and current strong health anxiety (OR: 1.56; 95% CI: 0.99–2.43) (Table 5).

**Table 5.** Multivariate logistic regression analysis with utilized media and current strong anxiety (forced/voluntary evacuees).


Model 1: Adjusted for gender, age, education, and evacuation type. Model 2: Adjusted for gender, age, and education OR, odds ratio; CI, confidence interval.

#### **4. Discussion**

The present study aimed to clarify the association between media utilization (e.g., Internet media, public relations information from local government, and other traditional media) and current strong health anxiety at the time of answering the questionnaire in the context of the Fukushima nuclear disaster. As per the results, the present study found a significant association between the use of public

relations information from local government and lower current health anxiety at the time of answering the questionnaire.

#### *4.1. Utilization of Media Information*

In a previous study that assessed media consumption after the Fukushima Daiichi nuclear disaster, over 95% of participants answered that they used television news as a common media source, whereas Internet news and personal Internet websites were used by less than 50% (39% and 14%, respectively) [12]. Although a simple comparison with our results is not possible due to differences in the survey items and methods, those findings are largely consistent with our present findings. On the other hand, the rate of use of any Internet media among voluntary evacuees was significantly higher than that among forced evacuees. In fact, the rate of Internet media usage was the highest among the media sources in the voluntary evacuees, while it was the lowest in the forced evacuees. When considering age, the Internet media utilization rate in the voluntary evacuees was 81.3% (13/16 respondents) among those aged 20–39 years (forced: 43.8%; χ 2 test, *p* = 0.03), 47.1% (16/34 respondents) among those aged 40–64 years (forced: 21.1%; χ 2 test, *p* = 0.01), and 5.3% (1/19 respondents) among those aged ≥65 years (forced: 1.6%; χ 2 test, *p* = 0.36) (Table 1 and Supplementary Tables S1 and S2). This suggests that younger generations use Internet media to the greatest extent among age groups, particularly among voluntary evacuees.

The utilization rate of public relations information from local government among forced evacuees was higher than that of voluntary evacuees. Compared by age group, the utilization rate among forced evacuees aged 40–64 years was significantly higher than that of the corresponding age group of voluntary evacuees (forced: 53.9%; voluntary: 17.6%; χ 2 test, *p* < 0.01), whereas the rates in those aged ≥65 years were similar between forced and voluntary evacuees (forced: 57.8%; voluntary: 52.6%; χ 2 test, *p* = 0.69) (Table 1 and Supplementary Tables S1 and S2). This result might be explained by differences in the age group composition of forced and voluntary evacuees.

#### *4.2. Specific Aspects of Current Strong Health Anxiety*

The proportion of respondents with current strong health anxiety due to radiation exposure at the time of answering the questionnaire was 20.3%. Among evacuees who experienced current strong health anxiety at the time of answering the questionnaire, more than 90% were concerned about the delayed effects, the genetic effects, and broadcasting about nuclear issues. In previous studies, specific anxiety was associated with the effects of radiation on the development of thyroid cancer [13,14], on the workplace environment [15], on expectant mothers and children [16], on the estimated occurrence of acute radiation syndrome (an acute illness caused by irradiation of the entire body by a high dose of radiation in a short period of time) [17], and on the reluctance to eat foods grown in the evacuation area [18]. Moreover, among the Fukushima nuclear disaster evacuees, concerns about radiation risks were associated with psychological distress [19]. Although risk perception or anxiety regarding the delayed and genetic effects due to radiation exposure decreased from 2012 to 2015 (delayed effects: 48.1% in 2012 to 42.8% in 2015; genetic effects: 60.2% in 2012 to 37.6% in 2015) [18,19], these rates of risk perception and anxiety were still over 30% among all evacuees, even four years after the disaster. Therefore, despite the gradual decrease in the risk perception of radiation exposure, anxiety regarding the delayed and genetic effects due to exposure was associated with current strong health anxiety.

During the nuclear emergency in Fukushima, the traditional media were found to provide a broad context, including frequent comparisons with previous nuclear accidents; however, the experts' technical vocabulary concerning radiation appeared incompletely translated for public understanding [20]. Therefore, our findings may show that, among evacuees who experienced current strong health anxiety at the time of answering the questionnaire, more than 90% had concerns about broadcasting regarding nuclear issues.

#### *4.3. Association between Current Strong Health Anxiety and Utilization of Media Information: Considerations for E*ff*ective Disaster Communication*

Among the responders, the proportion with current strong health anxiety due to radiation exposure at the time of answering the questionnaire was 20.3%. Current strong health anxiety at the time of answering the questionnaire was significantly lower among those who utilized public relations information from local government. The Public Relations Society of America (PRSA) stated that "Public relations is a strategic communication process that builds mutually beneficial relationships between organizations and their publics. It is thought that not providing information unilaterally but providing information promoting mutual-communication or useful information about lots of variety of consultation will lead to mutually beneficial relationships" [21]. In a previous report related to the Fukushima Daiichi Nuclear Power Plant accident, a unified approach was found no longer to be sufficient to address personal problems and anxiety as diverse information became available and people's perceptions developed. This led to the need for one-to-one or small-group communication [22]. Another study reported that attending radiation information seminars or programs helped to reduce anxiety and psychological distress in a post-Fukushima disaster setting [12,23]. The public relations information from local government in evacuation areas included several articles regarding the health effects of radiation exposure, as well as the maximum annual exposure dose. Other posted articles included records of decontamination processing, discussion records of the health risk communication promotion committee in evacuation areas, and articles providing information such as general health consultations and dialogue among evacuees and experts, as well as information including education regarding stress reactions and coping [24]. Therefore, utilizing public relations information from local government may be associated with lower current health anxiety at the time of answering the questionnaire.

Users of Internet media tended to feel anxiety toward perceived radiation health risks, but not to a significant degree. Several studies have examined the correlation between risk perception and anxiety and media/information after the Fukushima nuclear accident. Murakami et al. revealed that dread risk perception was greater among people who trusted direct information from online researchers or others than those who did not, but was lower among people who trusted central governmental information than among those who did not [25]. Sugimoto et al. surveyed 1560 residents of Soma City in July 2011 and found that health anxiety was high among those who relied on word-of-mouth or rumors as a means to obtain information [12]. Baseless rumors and conspiracy theories spread very quickly on the Internet, which may also explain the high levels of anxiety among those who mainly used the Internet as their information source. Any Internet media usage in this study did not include solely Internet news, but also the use of personal websites such as social networking sites (SNS), and information from these sources likely include word-of-mouth or rumors. This suggests a potential association between evacuees with strong anxiety and the use of any type of Internet media.

#### *4.4. Limitations and Strengths*

This study has several limitations. First, due to its cross-sectional design, causality could not be established. Second, our primary outcome, i.e., current anxiety regarding perceived radiation health risks, was a subjective response This could certainly be the case with actual and perceived health risks, leading to important differences in anxiety levels, without validated measures and reliability statistics such as a test–retest correlation, which is a critical limitation. Therefore, it would hardly be applicable to different settings without reliability. Further studies are needed to confirm the validity and reliability regarding current anxiety regarding perceived radiation health risks. Additionally, our definition of "current strong health anxiety" may be included as a bias. Those who responded to "Somewhat" (47.1% in both the forced and the voluntary evacuees, see Supplementary Table S3) as current health anxiety at the time of answering the questionnaire due to radiation exposure were categorized as the "no or weak health anxiety" group, although this could be in either group. However, this categorization was comparable to that of the Fukushima Health Management Survey report [26], which focused

on high and extreme anxiety of health effects due to radiation exposure. The third limitation relates to sampling from non-evacuation areas. We could not grasp detailed enough information about the number of voluntary evacuations in advance. As a result, many residents had not experienced evacuation voluntary, and thus more than 600 respondents were later excluded from the analysis. Fourth, because respondents tended to be relatively older, our study population included fewer Internet users, in particular SNS users. Finally, depending on the three types of evacuation areas according to the spatial radiation dose rates, there may have been different perceived health risks, leading to important differences in anxiety levels. However, due to anonymous sampling, it was impossible to obtain the detailed information on whether subjects were living in one of the three types of evacuation areas. This is because the first-stage sampling was selected by municipality, not by each of the three area types.

Despite these limitations, we were able to show a positive association between the utilization of public relations information from local government and lower health anxiety due to radiation exposure, even after adjusting for age, gender, education, and evacuation type. Although we examined the association between utilization of media and current health anxiety after the 2011 nuclear disaster, further studies in other settings, such as that of the novel coronavirus pandemic, are needed.

#### **5. Conclusions**

The 2011 nuclear disaster in Fukushima was not only a health disaster, but also an information disaster. Our findings highlight the importance of public relations information from local government in terms of it being associated with lower current health anxiety at the time of answering the questionnaire related to disaster situations, and this could potentially aid in preparing for future disasters.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/1660-4601/17/11/3921/s1, Table S1: Characteristics of media users relating to nuclear exposure (forced evacuees). Table S2: Characteristics of media users relating to nuclear exposure (voluntary evacuees). Table S3: Prevalence of current health anxiety due to radiation exposure.

**Author Contributions:** S.Y. designed the framed study and acquired funding. S.Y., T.N., M.S., C.N., and Y.K. contributed to designing the questionnaire. M.O., Y.K., N.M., H.I., T.H., and S.Y. investigated and administrated the survey. M.O. conducted data analysis and wrote the draft. All authors have read and agreed to the published version of the manuscript.

**Funding:** This study was supported by a grant from KAKENHI, Japan Society for the Promotion of Science (JSPS), as a Grant-in-Aid for Scientific Research (C) research (JSPS KAKENHI Grant Number: 15K08810).

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

International Journal of *Environmental Research and Public Health*

### *Article* **Factors Associated with Urban Risk-Taking Behaviour during 2018 Typhoon Mangkhut: A Cross Sectional Study**

**Evan Su Wei Shang 1,2, Eugene Siu Kai Lo <sup>1</sup> , Zhe Huang 1,2, Kevin Kei Ching Hung 1,3 and Emily Ying Yang Chan 1,2,3,4,\***


Received: 18 May 2020; Accepted: 8 June 2020; Published: 10 June 2020

**Abstract:** Although much of the health emergency and disaster risk management (Health-EDRM) literature evaluates methods to protect health assets and mitigate health risks from disasters, there is a lack of research into those who have taken high-risk behaviour during extreme events. The study's main objective is to examine the association between engaging in high-risk behaviour and factors including sociodemographic characteristics, disaster risk perception and household preparedness during a super typhoon. A computerized randomized digit dialling cross-sectional household survey was conducted in Hong Kong, an urban metropolis, two weeks after the landing of Typhoon Mangkhut. Telephone interviews were conducted in Cantonese with adult residents. The response rate was 23.8% and the sample was representative of the Hong Kong population. Multivariable logistic regressions of 521 respondents adjusted with age and gender found education, income, risk perception and disaster preparedness were insignificantly associated with risk-taking behaviour during typhoons. This suggests that other factors may be involved in driving this behaviour, such as a general tendency to underestimate risk or sensation seeking. Further Health-EDRM research into risk-taking and sensation seeking behaviour during extreme events is needed to identify policy measures.

**Keywords:** typhoon; hurricane; cyclone; strong wind levels; natural disaster; Health-EDRM; urban; risk-taking behaviour; sensation seeking

#### **1. Introduction**

Asia is particularly at risk of tropical cyclones, also known as typhoons in the western Pacific, with around half of worldwide tropical cyclones recorded and more than 90% of cyclone-related deaths being from this region [1]. Typhoon Mangkhut, the fourth supertyphoon in the 2018 Pacific typhoon season, started east of Guam in September 2018 and caused devastating damage to the Philippines [2] and South China [3]. In Hong Kong, tropical cyclone warning signal No. 10 (the highest signal for Hong Kong) was hoisted for 10 h, with the highest wind speed exceeding 150 km/h, just lower than the respective records held by Typhoon York in 1999 and Typhoon Ellen in 1983 [4]. Although there were no deaths reported directly due to the typhoon, more than 60,000 trees had reportedly fallen and around 13,500 households experienced a power outage for more than 24 h in Hong Kong [5]. Record

breaking storm surges coupled with high waves caused flooding in various coastal areas, despite not being at high tide, with hundreds of stranded or damaged vessels. These occurrences highlighted the increasing environmental hazards and potential health risks faced by coastal communities around the world as typhoons become more frequent and severe through climate change [6].

Health emergency and disaster risk management (Health-EDRM) is a field which involves the systematic analysis and management of health risks surrounding emergencies and disasters. By reducing risk and vulnerabilities and improving preparedness, response and recovery measures, the impact of disasters can be minimised [7]. Health-EDRM focuses on increasing the resilience of individuals, households and communities through education interventions, promotion of disaster risk reduction and supporting mechanisms in place to mitigate impacts of disasters [8]. These can include accessible disaster warning and information, protection of key health services and securing the basic needs of the population. By understanding the causes and factors in play regarding disaster risk, Health-EDRM provides evidence to drive future interventions and policy decisions.

A previous study on urban disaster preparedness in Hong Kong discovered that only 20.6% of respondents chose the correct action to take while major disaster warnings were in force (such as staying in a safe place until heavy rain has passed) [9]. However, few studies investigate individuals intentionally performing risk-taking behaviour during natural disasters. Research on 'storm chasers' (those who intercept severe convective storms for sport or for scientific research) [10] in the United States has examined individual perception of recreational storm chasing to dispel its myths [11], operational methods of storm chasing tour groups [12] and sensation seeking traits associated with tour participants [13]. There is a lack of research into such behaviour outside of the United States, especially in urban areas directly impacted by meteorological disasters. Various studies by Zuckerman on sensation seeking have identified its dimensions [14] and relationships to different aspects through the sensation seeking scale (SSS) [15]. A meta-analysis into sex differences in sensation seeking showed that men scored higher than women using Zuckerman's SSS-V and could be explained by evolutionary psychology and through a cultural socialisation perspective [16]. Furthermore, a review of behavioural and biological correlates of sensation seeking also found males significantly outscored females on total sensation seeking in different western countries and sensation seeking typically decreases with increasing age after adolescence [17]. Education and occupation were less associated with sensation seeking, particularly for females. The review also highlighted high sensation seekers perceived risks in the environment as less threatening compared to low sensation seekers and did not perceive engaging in high-risk behaviours would lead to negative consequences.

This article is an extension of the cross-sectional study investigating risk perception, household preparedness, and self-reported short-term impacts of typhoons after Typhoon Mangkhut [18]. Our previous published paper highlighted 16.0% of respondents reportedly left their homes when the warning signal was T8 or above, when the typhoon was at the height of strength. This behaviour will henceforth be referred to as risk-taking behaviour during typhoons (RBDT). Out of those respondents, a majority (74.7%) performed RBDT for non-essential reasons. The previous article also found that men and younger respondents were more likely to execute non-essential RBDT. Using the same dataset, the objectives of the current study are to expand on these findings and investigate other factors that may be related to RBDT, namely (1) to identify the sociodemographic characteristics of those who left their homes when the storm was at its height of strength (i.e., the warning signal was T8 or above); and (2) to explore the associations between sociodemographic factors other than age and gender, risk perception, household preparedness, and RBDT for non-essential purposes. The study findings will offer further insight into risk-taking behaviour during natural disasters to guide future interventions and policy on preventing such unnecessary high-risk behaviour.

#### **2. Materials and Methods**

A computerized digit dialling population-based household telephone survey was conducted from 17 September 2018 to 2 October 2018, right after the date of Typhoon Mangkhut landing in Hong Kong. Random digit dialling and the last birthday method [19] (interviewer would seek the household member whose birthday was the closest to the interview date) were used to ensure randomization in the study. Hong Kong residents who understood Cantonese and were 18 years old or older were interviewed. In total, 2500 landline numbers were called, and 521 respondents were successfully recruited (Figure 1). The response rate was 23.8% (response rate: 521 (final sample size) /2188 (eligible persons)). Please refer to the previous published study for study design [18].

**Figure 1.** The recruitment details in the telephone survey.

This study investigates the associations between non-essential RBDT and the sociodemographic factors of education and income; indicators of risk perception including perception of Hong Kong being susceptible to disasters, perception of the impact of Typhoon Mangkhut compared to expectations and concern for the safety of oneself and family members; indicators of household preparedness including food and water reserves prepared routinely or specifically for Typhoon Mangkhut.

This paper considers the act of going outside during the strongest typhoon winds as risk-taking (RBDT) or high-risk behaviour, regardless of the reasoning. Those who go outside for work or emergency related purposes would be considered as engaging in socially acceptable or understandable risk-taking behaviour. This article focuses on active risk-taking behaviour, where individuals remove themselves from areas of safety to head into areas posing health risks, and does not investigate passive risk-taking behaviour, such as failure to act, evacuate or engage in other safety-seeking behaviour. This is because there are potentially different motivating factors, rationale and mechanisms involved in these two types of behaviour and research has found that passive risks are associated with a lower risk perception than equivalent active risks [20]. Previous literature has investigated passive risks associated with disasters and relevant methods to protect passive risk, but there is a lack of research into active risk-taking behaviour.

To clarify, the respondents were asked whether they left their homes to go outside during Typhoon Mangkhut while the typhoon warning signal T8 or above was in force. Those who left their home to address any perceived urgent and unexpected situations that required immediate action to prevent further deterioration, such as "due to injury or disease", were classified as having 'emergency' reasons. Those who did not have to manage such pressing issues or work-related duties but left their home for other reasons, such as "eating a meal or watching a movie", were categorized as having 'non-emergency' reasons. This paper will hence refer to RBDT due to 'non-emergency', also referred to as non-emergency and non-occupational reasons in our previous published paper, as non-essential reasons. Verbal informed consent was obtained at the beginning of the interview. The ethical approval of this study was obtained from the Survey Behavioural Research Committee at the Chinese University of Hong Kong (SBRE-18-075).

Descriptive chi-square (or X<sup>2</sup> ) tests were used to compare the study population and the respondents who reported going outdoors during Typhoon Mangkhut. We conducted univariate analyses to identify associations between sociodemographic characteristics, risk perception, disaster preparedness factors

and the risk-taking behaviour. Multivariable logistic regression was performed to identify factors related to going outdoors during strong typhoon winds for 'non-emergency' reasons, using variables with at least marginal statistical association in the univariate analysis (*p* < 0.10). Age and gender were covariates for the multivariable model base. All odds ratios (OR) present in this paper were adjusted odds ratios from the multivariable models. Statistical analyses were performed using IBM SPSS 24 (International Business Machines Corporation, Armonk, NY, USA) [21] and statistical significance was set at α = 0.05 two-sided.

#### **3. Results**

Data were collected from 17 September 2018 to 2 October 2018. The final sample size constituted 521 valid respondents (the response rate was 23.8% among eligible people called). The study population were comparable with the Hong Kong 2016 census data, except the study population were proportionally more middle age (age 45–64), more at the post-secondary education level and had higher income. Please refer to the previous paper [18] for further information and detailed analyses.

#### *3.1. Description of the Study Population*

For the descriptive comparison (Table 1), it was found that men were more likely to engage in non-essential RBDT (*p* = 0.006). There were no significant associations between RBDT and marital status, income, and respondents with chronic disease. Of the respondents with occupations which require emergency work while typhoon No. 8 or higher was in force, 72.2% reported staying home during Typhoon Mangkhut when the typhoon was at the height of strength. In addition, respondents who had occupations relating to "Sales and services" and "Elementary occupation" were more likely to leave their homes for emergency or work reasons. Respondents who participated in RBDT for emergency or work reasons were found to be more likely to obtain their weather-related information during this typhoon through television and newspapers and less likely through websites and mobile apps.


#### **Table 1.** Descriptive table of the study population.

#### *3.2. Factors Associated with Non-Essential RBDT*

In the multivariable logistic regressions (Table 2), the relationship between various sociodemographic details, the subject's perception and preparedness for the typhoon and non-essential RBDT was examined. As reported in the previous article, being male and younger were found to have higher odds in performing non-essential RBDT than those that did not go outside. All factors investigated in this article such as education, disaster risk perception or household preparedness were not found to have a statistically significant association with non-essential RBDT after adjusting for age and gender. Food and water reserves (both routine and specifically prepared for Typhoon Mangkhut) were also not found to be related.

**Table 2.** Chi-square comparison and multivariable logistic regressions of the associating factors towards non-essential RBDT.


a is the multivariable regression of age and gender; <sup>b</sup> is the logistic regression adjusted with age and gender.

#### **4. Discussion**

The current study is an extension of the previous published paper [18], which investigated risk perception, household preparedness, and self-reported short-term impacts of typhoons after Typhoon Mangkhut. This study aims to identify the sociodemographic characteristics of those who performed RBDT and investigate correlations between other sociodemographic characteristics apart from age and gender, risk perception and preparedness, and non-essential RBDT. There were no significant associations between RBDT and marital status, income, and respondents with chronic disease. Respondents who participated in RBDT for emergency or work reasons were more likely to watch television and read newspapers to obtain their weather-related information and less likely through websites and mobile apps. Education, income, risk perception and preparedness were found to be insignificantly associated with non-essential RBDT.

#### *4.1. Comparison between Household Preparedness and Risk-Taking Behaviour*

Household preparedness and risk-taking behaviour may be negatively associated, as household preparedness represents active protective behaviour, while non-essential RBDT involves potentially injurious active behaviour. Although this study did not find any significant negative association between higher educational level, higher risk perception and routine emergency preparedness with non-essential RBDT, the previous published paper [18] found a positive association of these factors with individuals who engaged in household preparedness. The results suggest these two behaviours may have different perceived levels of risk and/or involve separate rationale, such as respondents not considering non-essential RBDT as high-risk activity but performing household preparedness to ensure adequate supplies for family members. However, there is a lack of data in this study on the magnitude and severity of risk perceived from the typhoon to support this hypothesis. In addition, this study could not indirectly gauge the relative amount of risk respondents were willing to take for RBDT, which could be independent from perception of general risks of the typhoon and disasters.

#### *4.2. Di*ff*erence between Typhoon Risk Perception and Risk Perception of Non-Essential RBDT*

Although there was no association between perceived impact of Typhoon Mangkhut compared to expectations and non-essential RBDT, 80.9% of all respondents thought the impact of the typhoon was similar or less than expected. This may suggest that fewer people underestimate the risks of typhoons, even after the recent typhoon influence, given that Typhoon Mangkhut was objectively one of the strongest typhoons in Hong Kong to date. The literature on sensation seeking involving demographics [17] has also shown that high sensation seekers have lower risk perception of activities they have not engaged in before and are less likely to perceive the negative consequences of risk-taking behaviour. Therefore, there may be a difference between the perceived risk of typhoons or the resulting impacts and non-essential RBDT which is an active behaviour. High sensation seekers may identify typhoons as events that cause harm but do not perceive the negative consequences of non-essential RBDT. As there is a lack of data directly investigating the reasons people performed non-essential RBDT, this study explores other inferred and documented possible rationale.

#### *4.3. Other Reasons for Non-Essential RBDT*

This study found food and potable water reserves, whether prepared regularly or specifically for this typhoon, were not associated with respondents going outside for non-essential reasons at the height of the storm. In addition to being markers of household preparedness, the reserves and the lack thereof may indicate situations of non-essential character, which could motivate individuals to leave their homes under the rationale of necessity. The results suggest that seeking these two necessities were not primary reasons for non-essential RBDT.

News outlets have documented a case of an elderly man in Hong Kong stranded at sea and requiring rescue after swimming during Typhoon Haima while the typhoon signal No. 8 was in force [22], highlighting an instance of RBDT which would have caused significant harm if rescue operations did not take place. As the man's motives were not interviewed and reported, the behaviour could have been due to an underestimation of the risk involved or due to sensation seeking. Despite warnings from the Hong Kong government [23] and Hong Kong Observatory [24] before and during Typhoon Mangkhut, people were reported participating in sensation seeking behaviour, such as 'experiencing the wind' and engaging in disaster photography [25,26]. High-risk behaviour during extreme weather has also been documented in the news, with 300 firefighters mobilised to rescue scores of 'frost chasers' and other hikers from Hong Kong's highest peak during the city's coldest day in six decades [27]. Although the news reports did not interview the individuals and verify their desire to seek novel and intense experiences, it is clear that the risk-taking behaviours reported were performed directly to experience aspects of the disasters.

#### *4.4. Ethical Concerns in Rescue Operations*

There is also an ethical argument on the duty of firefighters who bear personal risk and danger to help those engaging in sensation seeking behaviour. A fireman in Hong Kong died in 2017 after sustaining a cliff fall while rescuing a pair of hikers [28], highlighting the risk and peril involved even in normal conditions. While rescuers have a duty to respond to emergencies, in cases involving sensation seeking behaviour, this creates a moral issue posing unnecessary risks to the rescuers in such extreme weather events. Rescue operations may also hinder other emergency responses and invoke issues of justice and equitable resource allocation. In total, 160 firemen were involved in a 24 h operation to rescue a pair of hikers who were stuck on a hiking trail during Tropical Storm Pakhar, which struck Hong Kong only days after Typhoon Hato. The entire operation included a total of ten ambulances and thirty-one fire engines, with an estimated total cost of more than 344,000 HKD in staffing costs alone [29].

#### *4.5. What Is the Gap Found in This Study?*

One of the major gaps found in this study is that education, disaster risk perception and disaster preparedness were not found to be associated with whether respondents engaged in unexplained risky behaviour. Common understanding would suggest that those with better understanding of the risks presented and especially those concerned about the safety of themselves and their family members during the typhoon would practice less risk-taking or sensation seeking. This suggests that knowledge-based interventions may not be as effective in deterring individuals from engaging in non-essential RBDT.

#### *4.6. Recommendations*

As younger aged males were found to be more likely to engage in non-essential RBDT, health promotion targeting this group may be more effective, possibly through official government websites/apps or well-placed advertisements in social media platforms. Legislation limiting accessibility to areas of higher risk during typhoons, such as waterfronts and beaches, may also have a beneficial effect on warding off risk-taking behaviour. Numerous provinces and cities in the Philippines have imposed liquor bans during typhoons to prevent inebriated individuals from hindering relief operations [30–32], but the effectiveness of such legislation have not been investigated. Furthermore, legal restrictions of outdoor movement in the interests of public health may be more socially acceptable due to the experiences of social distancing amid COVID-19.

As some respondents were found to engage in RBDT for work purposes, the updated "Code of Practice in Times of Typhoons and Rainstorms", published by the Hong Kong Labour Department [33], is a crucial document in mitigating health risks to essential workers. Although the guidelines are not compulsory by law, employers and employees must discuss and clearly outline methods to protect the health of those working during extreme conditions. Interventions targeting this group may be more effective if broadcasted through television or newspapers. Thus, a multi-stakeholder approach should be adopted to promote work-related safety and reduce RBDT. This also has a global implication on preventable injury risk reduction, with international planning and discussions necessary in the near future before stronger natural disaster occur driven by climate change.

#### *4.7. Limitations*

The main limitation of the analysis is that the results do not demonstrate a causative effect since this study used descriptive analysis and logistic regression as a cross-sectional study. There may also be reporting bias as respondents may present with higher social desirability response bias due to telephone interviews [34]. In addition, analyses involving risk perception asked two weeks after may not correlate directly with risky actions performed during the disaster. This study is also unable to distinguish whether those who executed non-essential RBDT carried out high-risk behaviour due to

perception of low-risk or for sensation seeking purposes. There may also be other reasons forcing individuals to remain outdoors during the typhoon, such as street sleepers unable to reach temporary shelters, unaccounted for due to the study design. Recruiting participants using the last birthday method also limits analysis on household preparedness as the recruited participant may not be the decision-maker for household preparedness [35], and thus, unfamiliar with the measures taken.

#### *4.8. Future Research Directions*

Future studies can directly investigate the reasons or rationale behind RBDT and explore the extent of such behaviour in detail. The sensation seeking trait and relevant factors could also be examined to determine its association and relevance to RBDT by using a modified version of the sensation seeking scale. Research into interpersonal perspectives through injunctive safety norms [36] and motivating factors for adaptation behaviour, such as descriptive norms [37], may also offer further insight into risk-taking behaviour during disasters. Future studies regarding disaster risk perception may also consider quantifying risk perception using balanced rating scales, as the degree of risk perceived is likely important when analysing risk-taking behaviour.

#### **5. Conclusions**

While age and gender were associated with risk-taking behaviour during natural disasters, other sociodemographic characteristics (such as education) and measures of disaster risk perception or disaster preparedness were not found to be correlated with non-essential RBDT. Despite the lack of literature investigating this phenomenon, media outlets in Hong Kong have reported several cases of sensation seeking behaviour during typhoons. Future studies are necessary to investigate the scope of risk-taking behaviour during disasters and the reasons driving this behaviour. Although there have been limited reports of injuries caused by risk-taking behaviour, relevant policy makers should begin to discuss and implement solutions to prevent any accidents and reduce potential extra burden on the emergency response system.

**Author Contributions:** Conceptualization, E.Y.Y.C.; methodology, E.Y.Y.C. and E.S.W.S.; validation, E.S.K.L. and Z.H.; formal analysis, E.S.K.L. and Z.H.; investigation, E.S.W.S., E.S.K.L., Z.H., E.Y.Y.C., K.K.C.H.; data curation, E.S.K.L. and Z.H.; writing—original draft preparation, E.S.W.S., E.S.K.L., Z.H.; writing—review and editing, E.S.W.S., E.S.K.L., K.K.C.H., E.Y.Y.C., Z.H.; visualization, E.S.K.L. and Z.H.; supervision, E.Y.Y.C.; funding acquisition, E.Y.Y.C. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research paper was funded by the CCOUC Development Fund, Faculty of Medicine of The Chinese University of Hong Kong.

**Conflicts of Interest:** The authors declare no conflict of interest.

### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

International Journal of *Environmental Research and Public Health*

### *Article* **Forecasting of Landslide Displacement Using a Probability-Scheme Combination Ensemble Prediction Technique**

**Junwei Ma <sup>1</sup> , Xiao Liu 1,\* , Xiaoxu Niu <sup>1</sup> , Yankun Wang <sup>2</sup> , Tao Wen <sup>3</sup> , Junrong Zhang <sup>2</sup> and Zongxing Zou <sup>1</sup>**


Received: 9 June 2020; Accepted: 30 June 2020; Published: 3 July 2020

**Abstract:** Data-driven models have been extensively employed in landslide displacement prediction. However, predictive uncertainty, which consists of input uncertainty, parameter uncertainty, and model uncertainty, is usually disregarded in deterministic data-driven modeling, and point estimates are separately presented. In this study, a probability-scheme combination ensemble prediction that employs quantile regression neural networks and kernel density estimation (QRNNs-KDE) is proposed for robust and accurate prediction and uncertainty quantification of landslide displacement. In the ensemble model, QRNNs serve as base learning algorithms to generate multiple base learners. Final ensemble prediction is obtained by integration of all base learners through a probability combination scheme based on KDE. The Fanjiaping landslide in the Three Gorges Reservoir area (TGRA) was selected as a case study to explore the performance of the ensemble prediction. Based on long-term (2006–2018) and near real-time monitoring data, a comprehensive analysis of the deformation characteristics was conducted for fully understanding the triggering factors. The experimental results indicate that the QRNNs-KDE approach can perform predictions with perfect performance and outperform the traditional backpropagation (BP), radial basis function (RBF), extreme learning machine (ELM), support vector machine (SVM) methods, bootstrap-extreme learning machine-artificial neural network (bootstrap-ELM-ANN), and Copula-kernel-based support vector machine quantile regression (Copula-KSVMQR). The proposed QRNNs-KDE approach has significant potential in medium-term to long-term horizon forecasting and quantification of uncertainty.

**Keywords:** landslide displacement; predictive uncertainty; ensemble prediction; probability combination scheme; quantile regression neural networks (QRNNs); kernel density estimation (KDE)

#### **1. Introduction**

As one of the most common natural hazards in the world, landslides pose a significant threat to public health and safety. According to statistics, landslides have affected 4.8 million people and caused 18,275 deaths during the period of 2009–2019 [1]. Landslide displacement prediction, which provides the necessary information to determine the extent of ongoing hazard, has proven to be the most cost-saving risk reduction measure [2–4]. However, landslide displacement prediction is complex and remains a key challenge in natural hazard research. This challenge arises because landslides are

nonlinear, dynamic systems, and the associated movements can be induced by different causes, such as geological factors [5], hydrological factors [6,7], morphological factors, and human activities [4,8].

A large number of efforts in the literature have focused on the precise prediction of landslide displacement [9]. Currently, approaches used for landslide displacement prediction are categorized as physical modelling approaches and data-driven approaches [10]. Physical models (also known as white-box models), which rely on detailed descriptions of landslide mechanism processes, can provide clear physical explanations of landslides. The commonly used physical models include the tertiary creep model [11], the Hayashi model [12], and the general creep model [13]. Those physical models require numerous expensive geotechnical characterizations of the materials involved in landslides and therefore may be applicable only in limited cases [14].

Data-driven models differ from physical models because a characterization of the actual landslide mechanism processes is not fully required. Thus, the data-driven models are also known as black-box models. The main advantage of data-driven models is that the trained models can be easily updated on the basis of new and more recent data.

Data-driven models include but are not limited to statistical methods, artificial neural networks (ANNs), support vector machines (SVMs) [15], and extreme learning machines (ELMs) [16]. Owing to their capacity to approximate arbitrary, nonlinear, and dynamic systems with high precision, data-driven models achieve good model performance in the prediction of landslide displacement.

Despite their widespread application, the output of most existing data-driven models is a single estimate for each prediction horizon. These single estimates, which provide deterministic values, are referred to as point predictions [3]. The defining characteristic of a point prediction is its accessibility with regard to understanding and operation. The main drawback of point prediction is that it only provides the prediction error, with no information regarding the associated predictive uncertainties, which limits the use of point prediction in decision-making applications.

The predictive uncertainties consisting primarily of input uncertainty, parameter uncertainty, and model uncertainty could be substantial. It is highly desirable to know the degree of uncertainty that is associated with a particular point prediction and convert the point prediction into informative resources for emergency landslide risk management [3,17]. Only limited studies have examined the quantification of uncertainty associated with landslide displacement prediction by constructing prediction intervals (PIs). The output of a PI is an interval composed of upper and lower bounds, where we expect the predictive value of the series to fall within some (prespecified) probability, which is deemed the PI nominal confidence (PINC). A hybrid approach based on an echo state network and mean-variance estimation was proposed by Yao et al. [18] to measure the uncertainty in landslide deformation prediction and perform interval prediction. A bootstrap-based approach was proposed by Ma et al. [4] to perform interval prediction of landslide displacement. Wang et al. [2] proposed a direct interval prediction using least squares support vector machines or the construction of PIs of landslide displacement. Kernel-based support vector machine quantile regression (KSVMQR) was utilized in [3] for quantification of the predictive uncertainty of landslide displacement.

However, the traditional methods have certain disadvantages in displacement prediction and quantification of predictive uncertainty. For example, the bootstrap-based approach requires significantly high computational costs, especially for large datasets [2]. Additionally, the performances of SVM-based approaches are sensitive to the choice of kernel type and parameter values [19]. Therefore, more efforts still need to be made for the improvement of prediction performance and quantification of the predictive uncertainty.

Ensemble prediction, a state-of-the-art artificial intelligence technique, aims to improve prediction robustness and accuracy and uncertainty quantification [20,21]. Ensemble prediction has been successfully applied in a variety of fields, including prediction performance improvement and uncertainty quantification of remaining useful life [22], bankruptcy [23], shear capacity of reinforced-concrete deep beams [24], residential electricity consumption [25], wind power [26], flood susceptibility [27,28], and landslide susceptibility [29].

In this study, a probability-scheme combination ensemble prediction that employs quantile regression neural networks and kernel density estimation (QRNNs-KDE) was proposed for robust and accurate prediction and uncertainty quantification of landslide displacement. The Fanjiaping landslide with long-term and near real-time monitoring data was selected as a case study to explore the performance of the QRNNs-KDE approach. The deformation characteristics were clarified for fully understanding the triggering factors.

#### **2. Methodology**

#### *2.1. Description of Uncertainty Sources*

Predictive uncertainty in data-driven models consists primarily of input uncertainty, parameter uncertainty, and model uncertainty [30–32].

The input uncertainty is related to the input data uncertainty and the input variable section uncertainty. The input data uncertainty is primarily due to measurement and sampling error and environmental noise. The input variable section uncertainty accounts for uncertainty inherent in the selection of input variables from the candidate data set. For physical models, the required inputs are pre-determined, being consistent with considered rheological models. However, for data-driven models, the selection of input variables is problem-dependent and cannot be determined in advance. Only major and relevant variables are selected as final inputs to train the data-driven model. The selection of the variables to include in a data-driven model from the original data set is inherently uncertain, especially when the input candidate pool is very large. For example, in data-driven models that utilize decomposition algorithms, only a portion of the decomposed sub-components are selected as input variables. The candidate input pool, which consists of sub-components, increases very quickly with the decomposition level and potentially increases the input variable selection uncertainty.

The parameter uncertainty refers to the uncertainty in the model parameter vector and mainly arises from the inability to identify a unique set of best parameters for the model [33].

Model uncertainty arises primarily from the model structure uncertainty and model error. Model structure uncertainty is associated with the specific model setting of learning algorithms, such as the polynomial order in polynomial regression models, the number of hidden nodes in an ANN or ELM, and the type of kernel function in an SVM. The input uncertainty may also account for model structure uncertainty, because different input variables "automatically" produce different model structures. Model error refers to the difference between two model estimates with respect to the corresponding target and is caused by the inability to reproduce the real processes.

#### *2.2. Ensemble Prediction*

Ensemble prediction is not a specific learning algorithm but a strategic combination of multiple predictions into a single output with a model combination process [21]. Based on the selection of the learning algorithm, ensemble prediction models can be further classified into homogeneous and heterogeneous ensemble models (Figure 1). A homogeneous ensemble model generates multiple learners with the same learning algorithm on different training datasets, which are produced by manipulating the original training data (schematic illustrated in Figure 1a). Bootstrap aggregation, also known as bagging for short, is the most straightforward and widely used method of manipulating the training dataset. By contrast, a heterogeneous ensemble model generates multiple learners with different learning algorithms on the same training data set (schematic illustrated in Figure 1b).

**Figure 1.** General framework for ensemble prediction models. (**a**) Homogeneous ensemble model and

The base learner combination is the main step in the ensemble prediction model. Summation and averaging are simple combination schemes. A more general approach involves assigning a weight to each base learner. In the present study, a heterogeneous ensemble model was built based on QRNNs and KDE. QRNNs serve as base learning algorithms to produce multiple base learners, and the probability combination scheme based on KDE is used to combine the base learners into the final ensemble prediction.

( ) Prob( )

#### *2.3. Quantile Regression Neural Network*

(**b**) heterogeneous ensemble model.

#### 0 1 2.3.1. Quantile Regression

( ) inf{ : ( ) } th ( ( ), ( )) 1,2, , 1,2, , Quantile regression is a common statistical technique for conducting inferences concerning conditional quantile functions [34,35]. More formally, any real-valued random variable *Y* may be characterized by its distribution function as follows:

$$F(y) = \text{Prob}(Y \le y) \tag{1}$$

whereas for any 0 < τ < 1,

$$Q(\tau) = \inf\{y : F(y) \ge \tau\}\tag{2}$$

0 1 is called the τth quantile of *Y*.

( )

θ Given a data set (*xi*(*t*),*Y*(*t*)) for *i* = 1, 2, · · · ,*I* and *t* = 1, 2, · · · , *N*, the linear quantile regression can be expressed as follows:

ˆ

( )

 

$$\hat{Y}\_{\pi}(t) = \sum\_{i=1}^{I} \theta\_i \mathbf{x}\_i(t) + b \tag{3}$$

where 0 < τ < 1 is the quantile, and *b* is an error with zero expectation.

if 0

( 1) if 0

*Int. J. Environ. Res. Public Health* **2020**, *17*, 4788

The estimated parameters θ*<sup>i</sup>* can be approximated by minimizing a sum of the asymmetrically weighted absolute residual cost functions, which are expressed as follows:

$$E\_{\tau} = \frac{1}{N} \sum\_{t=1}^{N} \rho\_{\tau}(Y(t) - \hat{Y}\_{\tau}(t))\tag{4}$$

where *Y*(*t*) is the observation at time *t* and ρ<sup>τ</sup> is the check function, which is also known as the pinball loss function and is defined as follows:

$$\rho\_{\mathsf{T}}(\mathsf{x}) = \begin{cases} \mathsf{ }\mathsf{tau} & \text{if } \mathsf{x} \ge 0 \\ (\mathsf{x} - 1)\mathsf{x} & \text{if } \mathsf{x} < 0 \end{cases} \tag{5}$$

#### 2.3.2. Quantile Regression Neural Network

Given inputs *xi*(*t*) and an output *Y*(*t*), the output from a QRNN is calculated as follows:

Consider a hidden-layer transfer function *h*(·); the output from the *j*-th hidden-layer node *gj*(*t*) is given by applying the hidden-layer transfer function to the inner product between *xi*(*t*) and hidden-layer weights *w* (*h*) *ij* plus the hidden-layer bias *b<sup>j</sup>* (*h*) , which can be calculated as follows:

$$\log\_{\hat{l}}(t) = h(\sum\_{i=1}^{I} \mathbf{x}\_{\hat{l}}(t)\mathbf{w}\_{ij}^{(h)} + b\_{j}^{(h)}) \tag{6}$$

An estimate of the conditional τ-quantile *y*ˆτ(*t*) is

$$\hat{Y}\_{\tau}(t) = f(\sum\_{j=1}^{I} g\_j(t) w\_j^{(o)} + b^{(o)}) \tag{7}$$

where *w* (*o*) *j* are the output-layer weights, *b* (*o*) is the output-layer bias, and *f*(·) is the output-layer transfer function. The transfer function *h*(·) and *f*(·) are usually set as the hyperbolic tangent sigmoidal and linear function, respectively [36].

As an alternative method to prevent overfitting, weight delay regularization for the magnitude of the input-hidden layer weight can be applied by setting a penalty with a nonzero value.

#### *2.4. Kernel Density Estimation (KDE)*

Nonparametric density estimation is the process of fitting a parametric density model of a random variable without making the assumption that the density belongs to a particular parametric family [37,38]. Various methods have been proposed for nonparametric density estimation, e.g., k-nearest neighbors method, Parzen windows, histogram, and KDE [38]. In the domain of nonparametric density estimation, the K-nearest neighbors method has a very limited scope of practical applications due to its very poor performance. The Parzen windows method presents slightly better performance but also produces discontinuities (stair-like curves) that are quite annoying in practice [38]. A histogram is a simple form of the nonparametric density estimation. However, it suffers serious and noticeable drawbacks. First, the resulting visualization strongly depends on the choice of binning. Second, the natural feature of the histogram is discontinuity, which causes extreme difficulty if derivatives of the estimates are required.

Fortunately, those abovementioned drawbacks can be easily eliminated by using KDE [38,39]. In fact, KDE has been extensively studied and has become the most popular method in nonparametric

density estimation. Given a random sample *Y*1,*Y*2, · · · ,*Ym*, the value of the density at the point *y* estimated by the KDE method is given by the following: ˆ

1 <sup>5</sup> 1.06 ˆ

1

<sup>5</sup> 2.34 ˆ

$$\hat{f}(y,h) = \frac{1}{mh} \sum\_{i=1}^{m} K(\frac{y-Y\_i}{h}) \tag{8}$$

where *h* is the bandwidth with positive real value and *K*(·) is the kernel function. In this study, the most effective Epanechnikov kernel [38] was adopted and expressed as

$$K(y) = \frac{3}{4}(1 - y^2) \mathbb{R}(|y| \le 1) \tag{9}$$

where R(·) is the indicator function, that is, R(*y* ∈ *A*) = 1 for *y* ∈ *A* and R(*y* ∈ *A*) = 0 for *y* < *A*.

The selection of bandwidth parameter is a crucial issue in KDE. The bandwidth parameter influences the smoothness of the KDE curve and also determines the tradeoff between the bias and variance. In general, the smaller the bandwidth, the smaller the bias, and the larger the variance. A number of methods have been proposed to find the optimal bandwidth, such as Silverman's rule of thumb and the Sheather-Jones method. Silverman's rule of thumb bandwidth with a Gaussian kernel and Epanechnikov kernel can be computed as follows: 1 2 ( ), ( ), , ( ) 1 2 0 1

$$h^{optimal} \approx \mathbf{1.06} \boldsymbol{\partial} n^{-\frac{1}{5}} \tag{10}$$

$$h^{\text{optimal}} \approx 2.34 \text{\\$} \text{m}^{-\frac{1}{5}} \tag{11}$$

where σˆ is the estimation of σ (standard deviation of the input data) [38].

#### *2.5. Ensemble Prediction Employing QRNNs and KDE* = ( ) ( )

The proposed ensemble prediction employing QRNNs and KDE is shown in Figure 2. The QRNNs-KDE approach consists of four stages: (1) data splitting and normalization, (2) QRNN modelling, (3) probability density function (PDF) estimation by KDE, and (4) final ensemble prediction. 1 ( ) ( )

**Figure 2.** The overall flowchart of ensemble prediction based on the quantile regression neural networks and kernel density estimation (QRNNs-KDE) approach.

Data splitting and normalization: The original landslide monitoring dataset is divided into training data and testing data. The training data are used for model construction, and the testing data are used to evaluate the performance of the constructed model. To eliminate the influence of dimensional data, the training data and testing data are first normalized in the range of 0 to 1.

QRNNs modelling: QRNNs serve as base learning algorithms to generate multiple base learners *Y*1(*t*), *Y*2(*t*), · · · , *Ym*(*t*) by applying a finite number of conditional quantities τ<sup>1</sup> ≤ τ<sup>2</sup> ≤ · · · ≤ τ*<sup>m</sup>* within the domain 0 < τ < 1, e.g., τ = 0.01, 0.02, . . . , 0.98, 0.99. The base learners of landslide displacement are obtained after renormalizing the outputs from the QRNNs approach. To avoiding overfitting in QRNNs modelling, a penalty parameter with nonzero value is applied.

PDF estimation by KDE: Multiple base learners from the QRNNs base model are treated as the input for KDE to estimate the probability density function (PDF) of the base learners. The kernel function and bandwidth influence the shape of the KDE curve. An appropriate kernel function and an optimal bandwidth should be chosen to best match the features of the original dataset.

Final ensemble prediction: In the present study, the final ensemble prediction was obtained through a probability combination scheme as follows:

$$u\_t = \sum\_{i=1}^{m} p\_i(t)Y\_i(t) \tag{12}$$

where *pi*(*t*) is the probability value of the *i*-th base learner and *Yi*(*t*) is obtained from the KDE for monitoring period *t*.

#### *2.6. Evaluation Metrics and Uncertainty Quantification*

In this study, five indices—coefficient of determination (*R* 2 ) MSE, RMSE, NRMSE, and MAPE were applied to assess the performance of point prediction. R<sup>2</sup> , MSE, RMSE, NRMSE, and MAPE are defined as

$$R^2 = \left[\frac{\sum\_{t=1}^{N} \left(u\_t - \overline{u}\right)\left(\hat{u}\_t - \overline{u}\right)}{\sqrt{\sum\_{t=1}^{N} \left(u\_t - \overline{u}\right)^2 \left(\hat{u}\_t - \overline{u}\right)^2}}\right]^2\tag{13}$$

$$N \ll N$$

$$MSE = \frac{\stackrel{\leftrightarrow}{\sum} \left(\mathfrak{d}\_{t} - u\_{t}\right)^{2}}{N} \tag{14}$$

$$RMSE = \sqrt{\frac{\sum\_{t=1}^{N} \left(\hat{u}\_t - u\_t\right)^2}{N}} \tag{15}$$

$$\text{NRMSE} = \sqrt{\frac{\sum\_{t=1}^{N} (\hat{u}\_t - u\_t)^2}{\sum\_{t=1}^{N} u\_t^2}} \tag{16}$$

$$MAPE = \frac{1}{N} (\sum\_{t=1}^{N} \left| \frac{\mu\_t - \mu\_t}{\mu\_t} \right|) \times 100\text{\%} \tag{17}$$

where *u*ˆ*<sup>t</sup>* and *u<sup>t</sup>* denote the *t*-th predictive value and observation, respectively, and *u* and *u*ˆ denote the mean of the observation and the mean of the predictive value, respectively.

In the present study, the associated predictive uncertainties were quantified with PIs. After the above procedures, full PDFs of the future landslide displacement were achieved. An interval prediction with a (1 − α) × 100% confidence interval can be obtained from the α/2 and 1 − α/2 quantiles of the obtained PDF. The α level, also called the significance level, ranges from 0 to 1 and is the probability of not capturing the value of the parameter. The predictive values of the α/2 quantity and 1 − α/2 quantity are set as the upper bound (*U*1−<sup>α</sup> *t* ) and lower bound (*L* 1−α *t* ), respectively. For example, a 90% central PI can be obtained from the 0.05 and 0.95 quantiles of the PDF. The upper bound and lower bound of the 90% confidence level correspond to the predictive values of the 0.95 and 0.05 quantiles of the obtained PDF.

The prediction interval coverage probability (PICP), normalized mean PI width (NMPIW), and coverage width-based criterion (CWC) are three indices for evaluating the correctness of the approximated PIs. The PICP reflects the degree of reliability of PIs and is defined as

$$\text{PICP} = \frac{1}{N} \sum\_{t=1}^{N} I\_t^{1-\alpha} \tag{18}$$

where *I* 1−α *t* is defined as follows:

$$I\_t^{1-\alpha} = \begin{cases} 1 & u\_t \in [L\_t^{1-\alpha}, L\_t^{1-\alpha}] \\ 0 & u\_t \notin [L\_t^{1-\alpha}, L\_t^{1-\alpha}] \end{cases} \tag{19}$$

NMPIW measures the width of the PI; it is defined as

$$NMPIV = \frac{1}{N\varepsilon} \sum\_{t=1}^{N} \left( \mathcal{U}\_t^{1-\alpha} - L\_t^{1-\alpha} \right) \tag{20}$$

where ς is the range of the underlying targets.

For high-quality PIs, narrow PIs (smaller NMPIW) with a high coverage probability (large PICP close to 100%) have great value [40,41]. Theoretically, NMPIW and PICP are conflicting. Therefore, CWC, which is a new balance criterion between PICP and NMPIW [42], is proposed to give a comprehensive assessment of PIs. CWC is defined as

$$\text{CWC} = (\text{NMPIW} + \psi)e^{\frac{\gamma(\text{PICP} - \mu)}{2\delta^2}} \tag{21}$$

where ψ is a small positive value within the range of (0.1%, 0.5%), µ corresponds to the nominal confidence level associated with PIs that is usually set to 1 − α, and δ is a small positive value less than 1. γ is set to 1 during the training process; for testing, it is defined by the following step function:

$$\gamma = \begin{cases} 1, & \text{PICP} \ge \mu \\ 0, & \text{PICP} < \mu \end{cases} \tag{22}$$

#### **3. Case Study: Fanjiaping Landslide**

#### *3.1. Features of the Fanjiaping Landslide*

The Fanjiaping landslide is located on the southern bank of the Yangtze River and upstream of the Baishuihe landslide and downstream of the well-known Huangtupo landslide, which is approximately 56 km northwest of the Three Gorges Reservoir Dam (see Figure 3 for location). The Fanjiaping landslide is an ancient landslide [43,44] composed of two blocks: the Muyubao landslide and Fanjiaping landslide. The entire planar area of the landslide is approximately 1.96 million square meters, and the landslide volume is approximately 106 million cubic meters. The thickness of the Fanjiaping landslide ranges from 40 to 139.16 m. The Muyubao landslide is approximately 1500 m long and 1200 m wide. The average thickness of the Muyubao landslide body is approximately 50 m, and its estimated volume is 90 million m<sup>3</sup> .

**Figure 3.** Location of the landslide site.

The Muyubao landslide extends from an elevation of 100 m at the toe to 520 m at the crown (Figure 4a,b). The slope surface consists of alternating gentle and comparatively steep landforms. The sliding direction of the landslide is 20◦ . The Tanjiahe landslide, located on the downstream of the Muyubao landslide, is approximately 1000 m long and 400 m wide. The average thickness of the Tanjiahe landslide body is approximately 40 m, and its estimated volume is 16 million m<sup>3</sup> . The Tanjiahe landslide extends from an elevation of 135 m at the toe to 420 m at the crown (Figure 4c,d). The slope surface consists of alternating gentle and comparatively steep landforms. The sliding direction of the landslide is 345◦ .

The site-specific investigation shows that the landslide materials are arranged in two different layers: a colluvial deposit at the upper surface and highly disturbed sandstone at the lower surface. The cataclastic sandstone is underlaid by sandstone and mudstone of the Jurassic Xiangxi formation (J1x) with an average dip direction of 10–25◦ and a dip angle of 27–36◦ (Figure 4b,d). Soft coal layers are prevalent in the J1x formation, and many landslides have developed along the soft coal layers. The borehole data indicates that the landslide mass of the Muyubao and Tanjiahe landslide slide along a soft coal layer with a thickness ranging from 0.1 to 0.3 m. According to laboratory testing of sliding zone soil obtained from the borehole, the natural moisture content of the soil is 12.6%, and the natural density is 1.9 g/cm<sup>3</sup> .

**Figure 4.** *Cont*.

′ ′ **Figure 4.** Topographic map and geological profile of the Fanjiaping landslide. (**a**) Topographic map of the Muyubao landslide. (**b**) Geological profile of the Muyubao landslide along sections A-A′ , as recorded with monitoring instruments. (**c**) Topographic map of the Tanjiahe landslide. (**d**) Geological profile of the Tanjiahe landslide along sections B-B′ , as recorded with monitoring instruments.

#### *3.2. Input Data*

A total of sixteen GPS beacons were installed on the landslide mass to monitor the landslide movements in September 2006 (see Figure 4 for the GPS locations): four on the Tanjiahe landslide and twelve on the Muyubao landslide. The GPS monuments were manually surveyed once a month. In April 2016, four GPS monitoring points, ZG295, ZG296, ZG297, and ZG298, were updated to near real-time monitoring. At most, thirteen years' worth of monitoring data were obtained. Figure 5 shows the monthly rainfall intensity obtained from the Shazhenxi Meteorological Station near the Fanjiaping landslide, the reservoir water level, and the displacement from GPS survey monuments over the thirteen-year period from October 2006 to March 2018. The available data indicate that the landslide was unstable and continuously deforming during the entire monitoring period. The landslide exhibits a step-like deformation behavior because of the periodic fluctuations in the reservoir water level and heavy precipitation. The monitoring data from both Muyubao and Tanjiahe show that larger displacements occurred in the upper middle part of the landslide mass. From the sequence of the surface cracks and displacement magnitude, we speculate that the movement occurred first at the rear part and progressed downslope. Based on a previous study on the relations between slip-surface geometry, material structures, and deformational structures [45,46], the observed kinematic behaviors are expected independent of the characteristics of the landslide material. However, more work is needed to confirm these findings.

**Figure 5.** Reservoir water level, monthly rainfall intensity, and cumulative displacement from the Fanjiaping landslide area.

#### *3.3. Triggering Factors of the Landslide Movements*

Although the Fanjiaping landslide is one of the largest landslides in the TGRA, very few publications have reported detailed information on the triggering factors of the landslide movements. Fully understanding the triggering factors is critical for landslide mitigation and early warning. In this study, long-term and near real-time monitoring data were used to comprehensively analyze the landslide movements. The cumulative displacement at monitoring point ZG295, monthly rainfall intensity, and reservoir water levels in 2009, 2011, 2012, 2015 are shown in Figure 6a–d. The available data shows the following trends:

**Figure 6.** (**a**–**d**) Cumulative displacement at monitoring point ZG295, monthly rainfall intensity, and reservoir water level spanning the period of 2009, 2011, 2012, and 2015. (**e**) Cumulative displacement at monitoring points ZG296 and ZG297, daily rainfall intensity, and reservoir water level spanning the period of June 2016 to October 2017. (**f**) Annual displacement at monitoring point ZG291, ZG294, ZG288, and ZG289, and reservoir water level spanning the period of 2007 to 2017.

(1) When the reservoir water level first rose from 135 to 156 m at the end of 2006, a significant annual displacement of 330 mm occurred at monitoring point ZG291 in 2007. Similarly, an annual displacement of 260 mm occurred at monitoring point ZG291 in 2009 when the reservoir water level rose from 156 to 172 m at the end of 2008. After 2009, the annual displacements shows a decreasing trend (Figure 6f). The results of the above analysis suggest that landslide deformation occurred at the preliminary operation phase, and more significant movement is likely to occur when the reservoir water level reaches a new higher level.

(2) A large deformation occurred when the reservoir water level slightly dropped from 175 m to 170 m in November to February (I in Figure 6). During 2009 to 2015, the monthly deformation rate during this drawdown period was greater than 20 mm per month (Figure 6a–d). For example, when the reservoir water level dropped from 174 m to 172.21 m in January 2012 to February 2012, the displacements measured at monitoring points ZG295, ZG296, ZG297, and ZG298 were 38.98, 26.39, 35.12, and 41.78 mm, respectively (Figure 6c). However, when the reservoir water level significantly dropped from 170 m to 145 m in February to June (II in Figure 6), the monthly deformation rate decreased to less than 20 mm per month.

(3) When the reservoir level remained at 145 m in July to September (III in Figure 6) and the landslide area suffered a heavy rainfall event, the landslide deformation was likely to be suspended except for 2012. The maximum monthly rainfall intensity during those suspended activities was 158 mm. In July 2012, the landslide area suffered from a heavy rainfall event with a monthly rainfall intensity of 208 mm, and the monitoring point deformed at a high rate. From those comparative analyses, we can speculate than the minimum triggering threshold consists of episodes lasting one month with cumulative rainfall exceeding 158 mm.

(4) When the reservoir rose from 145 m to 175 m in September to November (IV in Figure 6), monthly deformation rate decreased to a small positive (less than 10 mm per month) or even negative value.

(5) The near real-time monitoring data also showed the abovementioned trends: when the reservoir water level rose from 147.25 to 174.42 m on August 31 2016 to October 27 2016, the monthly deformation rates for ZG296 and ZG297 were 2.6 and 3.3 mm/month, respectively. When the reservoir dropped from 174.45 to 171.18 m on November 9 2016 to January 16 2017, deformations of 24.07 and 26.46 mm occurred at ZG296 and ZG297, respectively. The corresponding monthly deformation rates were 12.8 and 13.2 mm/month, respectively.

From the above analysis we can conclude that landslide movement was especially pronounced under prolonged periods of dropping reservoir levels, especially during periods of slight dropdown at the highest reservoir level, and the minimum triggering threshold consisted of episodes lasting one month, with cumulative rainfall exceeding 158 mm.

#### *3.4. QRNNs-KDE-Based Method for Ensemble Prediction*

#### 3.4.1. Data Splitting and Normalization

The available data (Figure 5) indicate that for the two active blocks, the largest displacements were observed at monitoring points ZG289 and ZG291, respectively. Therefore, monitoring points ZG289 and ZG291 were selected to establish a prediction model for the Fanjiaping landslide.

Previous correlation analysis in [3] revealed that weak to very strong correlations exist between landslide displacement and triggering and state variables. Therefore, based on triggering factor analysis and previous work on correlation analysis in [3], seven variables including four trigger variables and three state variables were selected as the inputs: rainfall intensity over the past month (*x*1(*t*)), rainfall intensity over the past two months (*x*2(*t*)), average reservoir water level in the current month (*x*3(*t*)), variation in the reservoir water level in the current month (*x*4(*t*)), displacement over the past one month (*x*5(*t*)), displacement over the past two months (*x*6(*t*)), and displacement over the past three months (*x*7(*t*)). In addition, the displacement in the current month (*Y*(*t*)) was selected as the output. A data set (*xi*(*t*),*Y*(*t*)), *i* = 1, 2, · · · , 7 was generated based on the inputs and corresponding outputs. For the Tanjiahe landslide, the data from October 2006 to January 2015 with a size of 100 were treated as the training set, and the data from February 2015 to June 2015 with a size of 5 were used as the testing set. For the Muyubao landslide, the data from October 2006 to January 2015 with a size of 133 were treated as the training set, and the data from November 2017 to October 2018 with a size of 12 were used as the testing set.

#### 3.4.2. QRNN Modelling

Two nonlinear models with a sigmoidal transfer function and linear transfer function for τ = 0.01, 0.02, . . . , 0.98, 0.99 with an interval of 0.01 were trained for monitoring points ZG289 and ZG291 to generate multiple base learners. The number of hidden nodes in the QRNNs model was set to 5. The penalty for weight delay regularization was set to 1 to prevent overfitting in QRNNs model construction. For each monitoring period, a total of 99 base learners were obtained at conditional quantities ranging from 0.01 to 0.99 based on the QRNNs. The main parameters applied in the modelling of QRNNs are shown in Table 1.


**Table 1.** The parameters utilized in the QRNNs modeling for the Fanjiaping landslide.

#### 3.4.3. PDF Estimation by KDE

The multiple base learners from QRNNs were employed as inputs of Epanechnikov KDE to estimate the PDF. The optimal bandwidths for PDF estimation were calculated based on Silverman's rule of thumb. The optimal bandwidths for PDF estimation of testing data at ZG289 were set to 7.98, 8.05, 5.66, 7.05, and 9.26. The optimal bandwidths for PDF estimation of testing data at ZG291 were set to 5.28, 8.61, 7.77, 5.62, and 5.30.

#### 3.4.4. Final Ensemble Prediction

Final ensemble predictions for the Fanjiaping landslide were generated through a probability combination scheme. PIs were constructed from the obtained PDF to estimate the predictive uncertainty. For the purpose of aiding decision-making, it is preferable to have prediction information with high confidence levels to reduce risks. Therefore, PIs at a high PINC value of 90% were obtained and analyzed in the study.

#### **4. Results**

PDFs: The PDFs of predictive displacement at ZG289 and ZG291 constructed by the proposed QRNNs-KDE approach are shown in Figures 7 and 8. The fast movement is the main concern in landslide displacement prediction. Here, only a portion of the prediction describing the fast landslide is selected and shown. Figures 7 and 8 show that rather than a single estimate, the range and complete PDF of the predictive displacement are provided by the proposed approach. All landslide displacement observations are distributed in the middle of the PDFs with high probability in addition to the observations of May and June at ZG289, which appear at the tail of the probability density curve. The small fraction falling into the right tail follows the increase in the prediction period; here, there are more uncertainties associated with longer-term landslide predictions.

**Figure 7.** Probability density functions (PDFs) for the Fanjiaping landslide at ZG289 from February 2015 to June 2015.

**Figure 8.** PDFs for the Fanjiaping landslide at ZG291 from December 2017 to May 2018.

Final ensemble prediction: Figure 9 shows the final ensemble predictions. As shown in Figure 9, the ensemble predictions obtained via the probability combination scheme showed a high degree of consistency in the landslide displacement observations, with coefficient of determination values of 0.999932 and 0.999944. To further evaluate the prediction performances of ensemble prediction based on the QRNNs-KDE, the evaluation metrics of the BP, RBF, ELM, and SVM approaches are shown in Table 2. As shown in Table 2, the final ensemble predictions using the QRNNs-KDE approach outperformed the persistence methods with the smallest MSE, RMSE, NRMSE, and MAPE and the largest R<sup>2</sup> . Moreover, compared with predictions at monitoring point ZG289 using the Copula-KSVMQR approach in [3], the QRNNs-KDE approach provided more accurate prediction with smaller MAPE and RMSE.

**Figure 9.** Comparisons of the final ensemble predictions and observations for the Fanjiaping landslide at ZG289 and ZG291.


**Table 2.** Comparisons of predictions obtained from QRNNs-KDE, BP, RBF, ELM, and SVM for the Fanjiaping landslide.

*Note*: The most accurate prediction results are shown in bold italics.

Uncertainty quantification: Based on the PDFs shown in Figures 7 and 8, PIs at a high confidence level (90%) were constructed for ZG289 and ZG291 (Figure 10a,c, respectively). To evaluate the prediction performances based on the QRNNs-KDE approach, 90% PIs were constructed based on the bootstrap-ELM-ANN approach (Figure 10b,d). The corresponding evaluation metrics are shown in Table 3. As shown in Figure 10 and Table 3, the constructed PIs based on the QRNNs-KDE approach perfectly covered the observations with a high percentage, and the QRNNs-KDE approach outperformed the bootstrap-ELM-ANN approach with smaller NMPIW and CWC. For example, the performance indices NPIW and CWC of 90% PIs at ZG289 were 0.0215 and 0.1661, respectively, which were lower than those obtained using the bootstrap-ELM-ANN approach. The normalized mean PI width using the QRNNs-KDE approach was approximately 90% narrower than that for the bootstrap-ELM-ANN approach.

The experimental results show that the final ensemble predictions based on the QRNNs-KDE approach outperformed the traditional BP, RBF, ELM, SVM, and Copula-KSVMQR algorithms with regard to deterministic point prediction. The QRNNs-KDE approach was more informative than traditional algorithms because it provided the likely range of landslide displacement. The landslide observations were distributed in the middle of the prediction range with high probability. Moreover, regarding the aspect of uncertainty quantification, the QRNNs-KDE provided more satisfactory PIs than the bootstrap-ELM-ANN approach. Therefore, we believe that the final ensemble predictions based on the QRNNs-KDE approach have the advantages of accurate prediction and uncertainty quantification of landslide displacement.



*Note*: The prediction results with a narrower PI range are shown in bold italics.

**Figure 10.** *Cont*.

**Figure 10.** Comparisons of the observations and the constructed PIs at a 90% confidence level for the Fanjiaping landslide at ZG289 and ZG291 using QRNNs-KDE and bootstrap-ELM-ANN. (**a**) 90% PIs at ZG291 using QRNNs-KDE; (**b**) 90% PIs at ZG291 using bootstrap-ELM-ANN; (**c**) 90% PIs at ZG289 using QRNNs-KDE, (**d**) 90% PIs at ZG289 using bootstrap-ELM-ANN.

#### **5. Discussion**

In this study, with regard to point prediction, the probability-scheme combination ensemble prediction, which employs QRNNs-KDE, provided the best prediction. The fundamental reasons behind this can be explained from statistical, computational, and representational perspectives [47]. From a statistical perspective, the available training data set may not be able to provide sufficient information for training the true model (*h* \* in Figure 11). Constructing an ensemble model (*h* ' in Figure 11) might not be better than the single best prediction model *h* \* , but it does reduce the risk of choosing a bad learner with poor generalizability (schematic in Figure 11a). From a computational perspective, in a single model the training algorithms might get stuck in lock optima by only performing a local search. Constructing an ensemble model by searching from different starting positions might be a better alternative (schematic in Figure 11b). From a representational perspective, it is possible that the searched hypothesis space might not contain the true model *h* \* . Constructing an ensemble model might expand the representable space (schematic in Figure 11c).

**Figure 11.** Schematic that shows the fundamental benefits of the ensemble prediction model from statistical (**a**), computational (**b**), and representational (**c**) perspectives. *h* \* is the true prediction model; *h*1 , *h*<sup>2</sup> , and *h*<sup>3</sup> are single prediction models; and *h* ' is the ensemble prediction model obtained by combining the single prediction models *h*<sup>1</sup> , *h*<sup>2</sup> , and *h*<sup>3</sup> . The outer black curve is the hypothesis space of all possible models. The inner blue curve denotes the subset of hypotheses that give reasonable accuracy with the available training data (modified from [47]).

In the proposed QRNNs-KDE approach, the probability combination scheme is employed to combine 99 base learners into one final ensemble to improve the model performance. However, a concern about computational time may be associated with this ensemble strategy. The required computational time is highly related to the number of base learners. For the case of ZG 291, the required computation time is 191.85 s to train 99 base learners in RStudio Version 1.2.5042 on an Intel(R) Xeon(R) E-2176M @ 2.70 GHz CPU with 64 GB RAM. Thus, we believe that the proposed approach is computationally efficient.

Nevertheless, the probability-scheme combination ensemble prediction, which employ QRNNs and KDE, also holds inherent limitations associated with data-driven models, such as the lack of an explicit input-output relationship, and the requirement of large training data to maintain the model performance.

In practical applications, the main motivation for the construction of predictive range and complete PDF is to quantify the likely predictive uncertainty in the deterministic point predictions. Availability of range and complete PDF of the predictive displacement allows the researchers and practitioners to efficiently quantify the level of predictive uncertainty with the deterministic point predictions and to consider a multiple of solutions/scenarios for the best and worst conditions. Wide ranges are an indication of presence of a high level of uncertainty in the operation. This information can guide the researchers and practitioners to avoid the selection of risky actions under uncertain conditions. In contrast, narrow range means that decisions can be made more confidently with less chance of confronting an unexpected condition in the future, for example, if a sharp displacement increment with a wider range was predicted for the further. An alert should be carefully determined whether reaching tertiary creep stage by researchers and practitioners through comprehensive analysis. Under this circumstance, time-of-failure forecasting should be run in parallel, and a multiple of solutions/scenarios should be considered until either failure precursors are identified or the movements suspended.

The proposed QRNNs-KDE approach is suitable for medium-term to long-term horizon forecasting. Results from previous studies [2,48] have shown that the performance of data-driven models varies for landslides with different deformation behaviors. Usually, for landslides with drastic step-like deformation, the prediction accuracy is lower, and the corresponding prediction error is larger. Therefore, in practical applications of medium-term to long-term horizon forecasting, when predicting landslides with drastic deformation, the proposed QRNNs-KDE approach should be applied with caution. To achieve excellent performance, sufficient data are recommended and needed for model training.

#### **6. Conclusions**

In this study, a QRNNs-KDE approach was proposed to improve the prediction accuracy and uncertainty quantification of landslide displacement. The Fanjiaping landslide in the TGRA was selected as a case study to explore the performance of the QRNNs-KDE approach. The following conclusions from the study were obtained:

The movements of the Fanjiaping landslide was especially pronounced under prolonged periods of dropping reservoir levels, especially during periods of slight dropdown at the highest reservoir level, and the minimum triggering threshold consists of episodes lasting one month, with cumulative rainfall exceeding 158 mm.

The QRNNs-KDE approach achieves perfect performance and outperforms the traditional BP, RBF, ELM, SVM, bootstrap-ELM-ANN, and Copula-KSVMQR methods. Additionally, the proposed approach is more informative by providing the likely range and complete PDFs of landslide displacement. The landslide displacement observations are distributed in the middle of the prediction range with high probability.

In practical application, the proposed QRNNs-KDE approach is suitable for medium-term to long-term horizon forecasting. The range and complete PDF of the predictive displacement can supplement final point predictions for decision making.

**Author Contributions:** The work was carried out in collaboration between all the authors. J.M. and X.L. guided and supervised this research; X.N., Y.W., T.W., J.Z., and Z.Z. performed the field investigation; J.M. wrote the original draft; and J.M. and X.L. reviewed and edited the draft. All authors have contributed to, seen, and approved the manuscript. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the National Key R&D Program of China (Grant No. 2017YFC1501305), the National Natural Science Foundation of China (Grant Nos. 41702328 and 41572279), the Hubei Provincial Natural Science Foundation of China (Grant No. 2019CFB585), and the Huaneng Lancang River Hydropower Co., Ltd. (HNHJ18-H24).

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

International Journal of *Environmental Research and Public Health*
