Characterizing Spatiotemporal Patterns of Disasters and Climates to Evaluate Hazards to Crop Production in Taiwan

Abstract

Climate change causes frequent and severe disasters. A comprehensive assessment of disaster hazards is thus crucial to understanding variations in disaster patterns and planning mitigation and adaptation strategies. This study obtained information from a crop disaster dataset of Taiwan covering the period from 2003 to 2022. Additionally, principal component analysis and K-means clustering were used to create climate clusters to evaluate the effect of climate patterns on disaster hazards. The results revealed that tropical storm frequency substantially decreased, whereas rain disasters exhibited an increasing trend. The four regions of Taiwan exhibited variations in terms of hazards of various disasters. The cold wave hazard showed a significant upward trend in the central region. An upward trend of rain disaster hazards was only detected in the southern region. However, a downward trend in tropical storm hazards was detected across all regions. A distinct climate pattern was identified over the study period. After 2012, high temperature and dry climate were the primary climate patterns. These patterns exhibited a high hazard value for cold waves, droughts, and rain disasters. Hence, the present study’s findings indicate that managing cold waves and rain disasters is crucial to protecting crop production in Taiwan.

1. Introduction

Approximately 770 million individuals worldwide experienced hunger in 2020 [1]. To address this problem, crop production must be increased. Crop production is determined by factors such as variety, cultivation practices, and environmental conditions. Agriculture, unlike other industries, is highly sensitive to environmental influences. Several studies have reported a negative effect of global warming on crop production [2,3,4]. However, crop production is affected not only by global warming but also by extreme weather events. Because of the effects of climate change on agriculture, an additional 170 million individuals are projected to experience hunger by 2080 if conditions do not change [5]. Furthermore, climate change will cause more frequent extreme weather events [6,7,8,9], which are likely to threaten global crop production.

Disasters such as drought, floods, storms, and extreme temperatures caused a substantial loss in global crop production from 2008 to 2018 [10]. A study demonstrated a decline resulting from drought and extreme heat of between 9% and 10% in national cereal production across 177 countries [11], and another study observed that warm temperatures during the reproductive stages of maize caused a decline in yields of between 80% and 90% [12]. Moreover, excess soil moisture and flooding caused by heavy rainfall were reported to have substantially reduced maize (Zea mays L.) yields in the United States [13]. Rice (Oryza sativa L.) production has also been severely affected by extreme weather events [14,15,16]. Therefore, developing strategies to mitigate and adapt to disasters is thus crucial to ensuring food security.

Assessments of disaster risk provide crucial information for developing strategies to mitigate the effects of climate change on crop yields. According to the Intergovernmental Panel on Climate Change (IPCC), the risks of climate change involve hazard, exposure, and vulnerability [17]. “Hazard” refers to the potential occurrence of an event that may cause damage or loss of life, property, ecosystems, and environmental resources. “Exposure” refers to the individuals, livelihoods, species, or ecosystems that may be affected by disasters. “Vulnerability” involves sensitivity or susceptibility to harm and an incapacity to cope and adapt, and it is a measure of how severely exposed entities are affected by disasters. The present study examined the effects of disaster hazard on crop production in Taiwan.

Hazard assessment is the first step in determining the risk of environmental disasters to crops [18,19]. Studies on this topic have primarily used meteorological indices, such as the standardized precipitation index, waterlogging index, and heat index, as indicators of hazards [20,21,22]. However, these indices cannot fully reflect the influence of climate disasters on crops [23]. Another crucial source of information is disaster databases, which enable investigations of disaster characteristics and trends to reduce disaster risk and facilitate adaptation to climate change [24]. However, historical disaster data are difficult to assess and are primarily obtained on regional or national scales, rendering them unsuitable for use in small areas. Additionally, disaster databases such as the Emergency Event Database (EM-DAT), the NatCatSERVICE Database (Munich Re), and SIGMA (Swiss Re) [25,26,27], are focused on the effects of climate disasters on humans and may underestimate their effects on agriculture. To more thoroughly assess disaster hazards to crop production by using disaster data, researchers have typically employed historical disaster survey data or crop loss data [28,29]. In Taiwan, data on crop loss due to disasters can be obtained from the Report on Crop Production Loss Disasters of Taiwan (hereinafter, the Report).

The agriculture type of Taiwan is the small-scale farming system. The average size of cropping fields is 0.72 ha. Most of the crop producers work as an individual farm. Rice is the crop with the largest cropping area in Taiwan, followed by fruits and vegetables. Taiwan is an island located on the major tracks of typhoons (tropical storms) in the northwest Pacific region. An average of four typhoons affect the island each year, causing substantial rainfall [30]. Additionally, the mei-yu or “Plum Rains” front that typically occurs from May to June and diurnal rainfall cause extreme precipitation [31]. Furthermore, because of its location and geography, Taiwan’s climate ranges from tropical to subtropical, with a temperate climate in the mountains. Due to this diversity of climatic characteristics, Taiwan has experienced multiple climatic disasters, with an average crop production loss of USD 318 million annually from 2012 to 2017 [32]. To mitigate the economic effects of climatic disasters, Taiwan’s Agricultural Natural Disaster Relief Act, passed in 1991, provides disaster relief payments to compensate agricultural workers for their losses. Crop damage caused by disasters is assessed by agricultural technicians of district offices and aggregated by the Ministry of Agriculture to generate a yearly Report.

Although the agriculture of Taiwan experiences serious disaster damage every year, the analysis of assessment of disaster risk, especially for the agricultural sector, is rare. The assessment studies in Taiwan usually focused on a few specific disasters, such as flood and drought, and also quantified the hazard by meteorological indices [33,34,35,36]. In addition, the study region of these pieces of research was mostly restricted to a small area [37,38]. The impact of disaster to crop production in these studies was indirectly evaluated by mapping the hazard on the land use map [39,40]. Therefore, a comprehensive study that directly evaluates the hazard of various disasters to crop production is needed in Taiwan.

The present study extracted disaster records and crop production loss information from the Reports for the years 2003 to 2022 to evaluate the hazards of various disasters in two steps. First, the spatiotemporal variations in disasters were evaluated to examine disaster trends and distribution. Second, meteorological data were combined with disaster data and analyzed to elucidate the association between disaster hazards and climate. The results of this analysis may improve the understanding of agricultural disasters and provide information that can guide decision-makers in planning mitigation and adaptation strategies.

2. Materials and Methods

To characterize the spatiotemporal patterns of disasters and climates for the evaluation of disaster hazards, this study used two datasets and various statistical analysis techniques. The years of two datasets range from 2003 to 2022. The following sections contain the description of data and analysis techniques. A diagram of the workflow of each dataset and statistical analysis outputs are provided to illustrate the relation of datasets and statistical analysis techniques (Figure 1).

2.1. Study Area

This study was conducted in Taiwan, which is an island located in the northwestern Pacific Ocean. According to the similarity of climate between counties, the island can be briefly divided into four geographical regions (central, eastern, northern, and southern). The administrative divisions of Taiwan island are comprised of six cities (Kaohsiung, New Taipei, Taichung, Tainan, Taipei, and Taoyuan) and ten counties (Changhua, Chiayi, Hsinchu, Hualien, Miaoli, Nantou, Pingtung, Taitung, Yilan, and Yunlin). In this study, cities are described as counties and New Taipei city and Taipei city have been combined to form Taipei county.

2.2. Study Data

2.2.1. Meteorological Data

Daily weather data were collected from 535 weather stations and 251 automated rain gauge stations (https://codis.cwa.gov.tw/; accessed on 15 February 2024). The altitude of these stations was lower than 1200 m, which ensured that the recorded weather conditions corresponded with most agricultural environments. The rain gauge stations collected total daily precipitation data only. The weather variables examined in this study included daily maximum temperature (°C), daily minimum temperature (°C), daily total precipitation (mm), and daily average wind speed (m/s). Daily weather information for each county was calculated by averaging the data for each weather variable from the stations in the county. Missing values were imputed using the mean value of the corresponding date in the county. The county weather variables were subsequently used to calculate the mean daily maximum temperature (maxT), mean daily minimum temperature (minT), total precipitation (PREC), mean precipitation intensity (A_PREC, mm/day), number of dry days (DDS), and mean daily average wind speed (meanWS) for each month in each year. The mean precipitation intensity was calculated by dividing the total precipitation by the total number of wet days (precipitation > 0.1 mm). A dry day was defined as a day with total precipitation ≤ 0.1 mm. Additionally, this study assessed extreme weather conditions for each county by calculating the number of extreme weather days each month. Extreme weather conditions comprised extreme high temperature (daily maximum temperature > 35 °C, EHTDS), extreme low temperature (daily minimum temperature < 10 °C, ELTDS), heavy rains (daily total precipitation ≥ 80 mm, HRDS), and strong winds (daily average wind speed ≥ 4 m/s, SWDS). The ELTDS and HRDS were defined according to the threshold that the central weather administration of Taiwan placed on low temperature and heavy rain alerts. For EHTDS, the threshold was defined by the common failure point temperature of various crops [41]. The SWDS was defined by averaging the daily average wind speed during typhoons. These monthly weather variables were employed to characterize spatiotemporal variations in the climate.

2.2.2. Disaster Data

The crop disaster dataset was used to evaluate disaster hazards [42]. The dataset contains county-scale information on the production loss of various crops caused by each disaster in a given year. The damaged field area in hectares (ha), damage level (%), actual damaged area (ha), estimated production loss (in tons), and estimated financial loss (in New Taiwan dollars, TWD) are recorded for each crop affected by each disaster. The damaged field area represents the total area of fields affected by disaster. The damage level is the average percentage of crop failure in the fields. The actual damaged area is calculated by multiplying the damaged field area by the damage level. In our analysis, disasters caused by high temperatures were excluded due to their low frequency in the crop disaster data. Crop production loss records from outlying islands were also excluded, as were records with damaged field areas of less than 5 ha, to prevent the data from representing the production losses of only a few fields. Finally, the disasters in the crop disaster data were categorized as cold waves, droughts, rains, tropical storms, or winds. Each classification of disaster includes one or more weather conditions. For example, rain category contains the disasters caused by heavy rain, extremely heavy rain, continuous rain, and so on. The detailed disaster data processing can be found in a previous study [42].

2.3. Hazard Assessment

The IPCC [43] defines “hazard” as the probability of the occurrence of a disaster. However, some studies have defined hazard as disaster severity or the product of the probability of occurrence and severity [18,44]. To investigate the yearly variation in disaster hazards, the hazard index was calculated by summing the severity of each disaster for each year. A county-scale hazard index was derived using the following equation:

$$Hazard index=\sum _{i=1}^5\sum _{j=1}^{n_i}{Severity}_{ij}$$

(1)

where Severity_ij represents the normalized severity value of the n_ith disaster of the ith disaster type (i = 1, …, 5), j indicates the number of ith disaster types in the county, and n_i represents the number of disasters of the ith type. The severity of each disaster is the logarithm of the actual damaged area. To eliminate variations in cropland area between counties, the severity values were normalized by county. The normalized severity value was calculated using the formula

$${Severity}_{ij}={\displaystyle \frac{\ln A_{ij}-\min (\ln A)}{\max (\ln A)-\min (\ln A)}}$$

(2)

where A_ij represents the total actual damaged area of the n_ith disaster of the ith disaster type. A is the value of total actual damaged area for a county.

2.4. Statistical Analysis

2.4.1. Principal Component Analysis

In total, 98 monthly weather variables, including maxT, minT, PREC, A_PREC, DDS, meanWS, and the number of extreme weather days were selected to represent the spatiotemporal climate patterns of 300 variable combinations for 15 counties. In assessing the variables, the present study only considered EHTDS from May to September. Additionally, ELTDS was only assessed for the months of January, February, March, November, and December. Furthermore, only data from the months from April to October were employed to calculate HRDS, and SWDS was calculated using data from January, February, March, July, August, September, October, November, and December. Principal component analysis (PCA) was employed to make the dimensionality reduction of monthly weather variables and characterize climate patterns. First, the monthly weather variables were standardized, and PCA was conducted using a covariance matrix. Monthly weather variables with larger loading absolute values were considered to contribute more to principal components (PCs). The first n PCs that explained 80% of the total variance were identified as crucial. The monthly weather variables for each year and county combination were multiplied by loadings to calculate PC scores. The combined score of each PC was considered to represent the climate characteristics. Combinations with comparable scores across multiple components were considered to indicate similar climate conditions.

2.4.2. Cluster Analysis

The PC scores were used to classify the climate patterns by using combinations of years and counties and the K-means procedure. The pseudo-F statistic (pseudo-F) and cubic clustering criterion (CCC) were employed to determine the optimal cluster number [45,46]. Pseudo-F and CCC were calculated for potential cluster numbers from 2 to 50. Subsequently, the potential cluster numbers were plotted against the pseudo-F and CCC values. The candidates for the optimal cluster number corresponded to the peaks in pseudo-F and CCC values.

2.4.3. Trend Analysis

To detect trends in the numbers of disasters, climate patterns, and disaster hazard variations, Kendall’s tau-b correlation coefficient (tau-b) was calculated [47]. Tau-b is a nonparametric measure of association based on the number of concordances and discordances in paired observations, and it has been applied to analyze trends in other studies [48,49]. The value of tau-b ranges from −1 to 1. The significance of tau-b (its distance from 0) was used to determine trends; positive and negative values of tau-b indicated upward and downward trends, respectively.

2.4.4. Linear Model

A linear model was used to investigate disaster hazards across climate clusters. The model contained only one factor (climate cluster) that was considered a fixed effect. When the assumption of homogeneity of variance was violated, log(y + c) transformation was applied [50]. In this formula, y indicates the observed values and c is a constant equal to 1/6. A one-way analysis of variance (ANOVA) was subsequently employed to assess the significance of the cluster effect.

3. Results

3.1. Disaster Distribution and Hazard Variation

3.1.1. Temporal Patterns of Disaster Frequency

To illustrate trends in the temporal variation of disaster frequency, the moving average of disaster frequencies over 3 years was calculated. A substantial upward trend in the yearly frequency of all disasters was observed, with an average of 12 disasters occurring per year (Figure 2a and

Table 1. Significant annual trends of disaster frequency and hazard index values for 2003 to 2022.

Disaster Classification	Entire Taiwan				Northern		Central		Southern		Eastern
	tau-b †	p Value †	tau-b	p Value	tau-b	p Value	tau-b	p Value	tau-b	p Value	tau-b	p Value
All disasters	0.598	<0.001	0.221	0.173	0.063	0.697	0.253	0.119	0.295	0.069	−0.316	0.052
Cold wave	0.512	<0.01	0.31	0.064	0.118	0.51	0.346	<0.05	0.385	0.055	0.105	0.586
Drought	0.281	0.238	0.256	0.222	−0.111	0.655	−0.2	0.624	0.4	0.327	−0.222	0.404
Rain	0.567	<0.001	0.411	<0.05	0.015	0.934	0.255	0.14	0.509	<0.01	0.2	0.299
Tropical storm	−0.489	<0.01	−0.484	<0.01	−0.346	<0.05	−0.415	<0.05	−0.505	<0.01	−0.505	<0.01
Wind	0.218	0.433	0.2	0.421					0.167	0.532

^† trends of disaster frequency.

). From 2003 to 2022, the frequencies of cold waves and rain disasters increased substantially (Figure 2b). After 2006, a substantial decreasing trend in tropical storm frequency was observed, with an average of 3.5 tropical storms per year. Drought and wind disasters were rare and exhibited no observable trend during this time frame. After 2011, rain disasters replaced tropical storms as the primary crop disaster. Furthermore, after 2011, the frequencies of cold waves approached those of tropical storms, with cold waves occurring more frequently than tropical storms from 2018 to 2022.

Although a downward trend in tropical storm disaster frequency was observed, the frequency of all disasters increased substantially. The increasing frequency of all disasters was attributable to the increasing frequency of cold waves and rain disasters. Additionally, the tau-b values for droughts and wind disasters exhibited an upward trend, although these values were nonsignificant.

3.1.2. Spatial Patterns of Disaster Frequency

The present study segmented Taiwan into northern, central, southern, and eastern regions for analysis. These four regions exhibited variations in total disaster frequency by county (Figure 3). The lowest disaster frequencies were observed in the northern region, whereas the southern region experienced disasters most frequently. Tropical storms accounted for the majority of disasters in the eastern, northern, and southern regions. However, rain disasters accounted for an average of 31.46% of disasters across all regions. In some northern and relatively high-altitude areas of Taiwan, the proportion of cold wave disasters exceeded 20%. Because tropical storms typically develop in the eastern sea region, the eastern region experienced the highest proportion of tropical storms, which were the predominant disaster type in this region. By contrast, the central region experienced few tropical storms because of the barrier of Taiwan’s mountainous terrain.

3.1.3. Spatiotemporal Variations in Disaster Hazards

A high disaster hazard index value can result from a high disaster frequency, a large area of damaged crops, or both. For the entirety of Taiwan, the disaster hazard index did not exhibit a significant upward trend during the study period (Figure 4e and Table 1). Additionally, contrasting hazard trends were observed between rain disasters (which trended upward) and tropical storm disasters (which trended downward). Moreover, the cold wave disaster hazard index value exhibited a slightly significant upward trend (tau-b = 0.31, p value = 0.064). A significant downward trend in the tropical storm disaster hazard index values was detected in all four regions (Figure 4a–d and Table 1). The remaining disasters exhibited various region-specific trends. Specifically, the hazard index value for cold wave disasters increased significantly in the central region, whereas the southern region experienced a significant increase in rain disaster hazards. Drought and wind disasters occurred infrequently in the four regions and did not demonstrate a significant trend. Although the results were only slightly significant, upward (tau-b = 0.295, p value = 0.069) and downward (tau-b = −0.316, p value = 0.052) trends for all disaster hazards were observed in the southern and eastern regions, respectively.

3.2. PCs of Weather Variables

The first 15 PCs explained 80.12% of the spatiotemporal climate pattern variations. PC1, PC2, and PC3 explained 20.96%, 13.94%, and 9.88% of the total variations, respectively. The contributions of monthly weather variables, including maxT, minT, PREC, A_PREC, meanWS, EHTDS, ELTDS, HRDS, and SWDS, to the first two PCs are illustrated in the loading plots in Figure 5. The magnitudes of the loadings represent the contributions of the weather variables to the PCs, whereas the direction of the loadings indicates a positive or negative association between the weather variables and PCs.

The maxT and minT values of each month were positively correlated with PC1, and most months exhibited magnitudes greater than 0.5 (Figure 5a,b). This positive correlation indicates that warmer climates exhibited a higher PC1 score. Additionally, although the loadings of maxT and minT were not high for PC2, the loading directions for PC2 indicated that summer exhibited a different temperature tendency from those of other seasons (Figure 5a,b). Specifically, summer temperatures that were above average resulted in a high PC2, as did below-average temperatures in other seasons. By contrast, above-average temperatures during non-summer months and below-average values during summer months resulted in a low PC2 value. The loadings of EHTDS and ELTDS indicated contrasting directions for PC1 (Figure 5g). However, these loadings had low magnitudes for both PC1 and PC2. PREC and A_PREC, exhibited varying trends between summer and other seasons for both PC1 and PC2 (Figure 5c,d), as did RHDS (Figure 5h). Moreover, DDS contributed substantially to the PC2 value in July and August, whereas the DDS values for February, March, and April were correlated with PC1 (Figure 5e). Furthermore, the meanWS values for all months had loadings with magnitudes higher than 0.5 that were correlated with PC2 (Figure 5f), as did SWDS in January, February, March, September, October, November, and December (Figure 5i). These results indicate that PC2 primarily represents meanWS and SWDS variations. The values of the loadings for the important PCs are presented in Supplementary Table S1.

3.3. Cluster of Climate Patterns

3.3.1. Determination of Cluster Number

Multiple peaks were evaluated to determine the optimal number of clusters (Figure 6). The CCC and pseudo-F statistics exhibited global peaks at 47 and 16, respectively. However, the optimal cluster number was expected to be lower than 15 for results with explanatory power. The CCC and pseudo-F statistics both suggested an optimal number between 6 and 11 clusters. Although the 6-cluster peak was slightly higher than the 11-cluster peak for pseudo-F, 11 was selected as the optimal cluster number to enable inclusion of more climate patterns.

3.3.2. Climate Clusters

The temperature of Cluster 1 was near the average temperature for all clusters throughout the year, with the exception of February and July (Figure 7a,b). Additionally, because the value for DDS was similar to the average in Cluster 1, that cluster exhibited relatively low PREC and A_PREC during summer (Figure 7c–e). Moreover, Clusters 2, 4, and 7 had higher than average temperatures in all months (Figure 7a,b). Specifically, Cluster 2 exhibited high DDS, low PREC, and low A_PREC from January to May and October to December (Figure 7c–e). Cluster 4 experienced high PREC and A_PREC in July and high HRDS and SWDS (Figure 7c,d,f,h). Cluster 7 exhibited notable EHTDS from late spring to early fall (Figure 7g). Furthermore, Clusters 3, 6, and 10 exhibited lower-than-average temperatures in most months (Figure 7a,b). Finally, Cluster 3′s high A_PREC in June, July, and August was due to high DDS and PREC in these months (Figure 7c–e).

Except during July and August, Cluster 6 exhibited lower than average DDS (Figure 7e). Additionally, Cluster 10 recorded the highest PREC, A_PREC, and HRDS of the clusters in October (Figure 7c,d,h). Moreover, Cluster 8 exhibited a consistently low temperature and wind speed throughout the year (Figure 7a,b,f). Cluster 5 was characterized by high meanWS and high SWDS in the fall and winter (Figure 7f,i), and Cluster 9 experienced relatively cold temperatures from January to March (Figure 7a,b). Finally, Cluster 11 exhibited the highest A_PREC and HRDS in April and August (Figure 7d,h).

3.3.3. Cluster Frequency

Clusters 1 and 2 exhibited the highest frequencies of all clusters (

Table 2. Cluster characteristics and frequencies of climate patterns between 2003 and 2022.

Cluster	Frequency	Defining Characteristics
1	61	Low total precipitation and precipitation intensity in summer. Cold in February and hot in July.
2	72	High temperature throughout the year. Dry in winter and spring.
3	14	Cold in winter and early spring. Relatively wet in February, March, May, June, July, and August.
4	22	Warm and dry in spring, fall, and winter. Extremely wet with high frequency of strong winds in July.
5	24	Dry in late spring and summer. Wet in fall. High wind speed in fall and winter.
6	32	Cool and wet in spring, fall, and winter.
7	6	High temperature throughout the year. Extreme high temperatures from May to September. Dry in spring, fall, and winter.
8	27	Low temperature and wind speed throughout the year. High frequency of heavy rain in September.
9	21	Cold in January and February.
10	8	Wet with low temperatures in spring, fall, and winter. Extremely wet in October.
11	13	High frequency of heavy rain in April and August. Dry in fall.

). Cluster 1 was observed to have a high frequency after 2012 (Figure 8e), and the frequency of Cluster 2 steadily increased after 2006. Cluster 1 was notably frequent in the northern region, whereas Cluster 2 was predominant in the southern region (Figure 8a,c). In the central region, Cluster 2′s frequency increased after 2009 (Figure 8b). Clusters 3 and 4 were primarily observed between 2003 and 2011. By contrast, the frequencies of Clusters 5 and 6 varied during the study period, particularly in the eastern region (Figure 8d). Furthermore, Clusters 7 and 10 exhibited frequencies below 10. Cluster 7 was only observed in the central and southern regions, whereas Cluster 10 occurred only in the northern and eastern regions. Cluster 8 occurred throughout the study period and was primarily observed in the central region. Cluster 9 was primarily detected between 2010 and 2018. Finally, Cluster 11 exhibited a low frequency and was primarily observed in the northern and central regions.

3.4. Effects of Climate on Disaster Hazards

To investigate the effects of climate on disaster hazards, the average hazard index values of each disaster in the climate clusters were calculated and compared. Further comparisons were conducted between some clusters because cold waves and droughts were not observed in all clusters. Additionally, wind disasters were excluded from the comparison due to their low frequency.

Variations in disaster hazards were observed between clusters (Figure 9e). Additionally, the results of an ANOVA revealed a significant effect of climate clusters on disaster hazards (Table S2). Cluster 3 exhibited the highest disaster hazard index values, whereas Clusters 1, 2, 4, and 8 exhibited moderate disaster hazard values. Cluster 1 exhibited the highest cold wave hazard values (Figure 9a), whereas Clusters 4 and 11 exhibited low cold wave hazard values. However, climate clusters did not significantly affect cold wave hazard values (p value = 0.0631). Only Clusters 1, 2, 5, 6, 8, and 11 experienced drought disaster hazards (Figure 9b), but no significant differences in drought disaster hazard values were observed between them. Additionally, the effects of climate on rain and tropical storm disaster hazards were significant. Clusters 1, 2, 3, and 10 exhibited high rain disaster hazard values (Figure 9c), whereas high tropical storm hazards were observed only in Clusters 3, 4, and 8 (Figure 9d).

4. Discussion

4.1. Threat of Disasters under Climate Change

This study quantified disaster hazards by assessing the frequency and severity of disasters. Under climate change, the frequency of disasters caused by extreme weather events is projected to increase. Globally, the frequency of floods, storms, droughts, and heat waves increased from 1970 to 2010 [10,51,52]. The increase in disaster frequency under climate change is primarily attributable to variables such as temperature and precipitation, which are influenced by higher atmospheric greenhouse gas concentrations [43,53]. Additionally, both disaster frequency and intensity are affected [17,52]. In climate change projections, high sea surface temperatures are the primary factor responsible for increasing tropical storm intensity [54]. Moreover, an increase in intense rain events was observed from 1977 to 2016 [55]. Furthermore, in China, the intensities of drought and heat events during the maize growth period increased between 1990 and 2020 [56]. Because of the increasing frequency and intensity of disasters, several studies have investigated their spatiotemporal variations [29,57,58]. Reported variations in cold waves, droughts, rains, and tropical storms are discussed in the following subsections.

4.1.1. Cold Waves

Under global warming, average temperatures have increased yearly. 2023 was the warmest year since global records began in 1850 [59]. Because of the warming climate, cold weather events are expected to become less frequent. Research revealed a nonsignificant effect of extreme cold on cereal production that could be explained by cold weather occurring outside the growing season in most agricultural regions [11]. In the present study, warm climates (Cluster 2) exhibited an upward trend in the northern, central, and southern regions from 2003 to 2022. However, cold waves exhibited an upward trend in frequency and hazard. Similarly, a significant increase in crop areas affected by frost disasters from 1990 to 2011 was observed in a study conducted in China [60]. Another study conducted in China detected a change in the frequency and scale of devastation of extreme low temperatures (shift from a decreasing trend toward stability after the mid-2000s) and a slight increase in the hazard index from 1990 that was attributable to the increased duration and intensity of low temperatures [61]. Additionally, an observational study ascribed winter continental cooling to Arctic amplification caused by global warming [62]. Severe winters in East Asia are associated with rising temperatures in the Barents–Kara Sea region [63]. A significant increase in stratospheric polar vortex stretching associated with warming of the Barents–Kara Sea region was observed from 1980 to 2020 [64]. Consequently, cooler winters have increased in frequency in East Asia. In the present study, Cluster 1 dominated the climate patterns of the northern, central, and eastern regions of Taiwan after 2012, with cold temperatures and a high frequency of extreme low temperatures observed in February. Additionally, Clusters 8 and 9, characterized by cold climates, occurred frequently after 2012. Notably, the warming environment during other seasons may induce premature crop development, leading to higher vulnerability during frosts [65,66].

4.1.2. Droughts

Clusters 1 and 2 exhibited dry seasonal conditions, indicating a high drought hazard. Because of the predominance of these climate clusters, the drought hazard in Taiwan is projected to increase. Trends toward an increasing incidence of drought have also been observed in studies conducted in other countries; long-term climate change projections indicate a rise in drought hazard from 2021 to 2050 and from 2041 to 2070, with varying hazards across three southeastern regions of Italy [57]. In the present study, neither drought frequency nor drought hazard exhibited an increasing trend from 2003 to 2022. However, a nonsignificant trend in drought frequency was expected because multiple drought disasters rarely occur within a single year in Taiwan. Most drought disasters occurred in the years after 2010 and occurred in consecutive years from 2017 to 2022. For comparison, in South Africa the frequency of dry years for rainfed agricultural systems increased dramatically after 2010, and larger areas of irrigated cropland have been affected by drought since 2012 [67]. Regarding spatial variations of drought hazards in Taiwan, the central and southern regions experienced such hazards only in the years following 2010. However, the incidence of droughts is likely to increase due to Taiwan’s warm climate. One study conducted in China observed that provinces with a warm temperate monsoon climate exhibited a greater increase in drought hazard compared with provinces with a continental monsoon climate [68]. Globally, semiarid areas, such as Central Asia, the western United States, Australia, and southwestern Africa, experience high drought hazards, whereas some humid areas, such as northwest France and southeast Brazil, experience moderate to severe drought hazards [69]. The heightened drought hazard in these humid areas is due to the increased frequency and intensity of droughts [70].

4.1.3. Rains

Rainfall is a crucial source of water for agricultural systems. However, excessive precipitation can cause waterlogging and flooding, threatening crop production [71,72]. Numerous studies have demonstrated an increase in extreme precipitation worldwide [73,74,75,76], although trends vary between regions. The seasonal melting dynamics in mountainous areas are assumed to affect regional precipitation and may be associated with a significant positive trend in heavy rain frequency in some climatic zones of Nebraska in the United States [77]. In Mexico and the southern United States, regions frequently affected by tropical storms and other tropical system activities exhibited a significant upward trend in extreme summer precipitation [78]. In Taiwan, an increase in extreme precipitation was observed in the northern, central, and southern regions, with a large and sudden increase in the southern region [79]. The present study uncovered spatial variation in rain disaster hazards corresponding to variations in extreme precipitation. Specifically, although rain disaster hazards exhibited an overall upward trend, a significant trend was observed only in the southern region of Taiwan. Additionally, varying trends in extreme precipitation frequency were observed between seasons [31]. The results of the present study reveal that the increase in rain disasters was associated with a high frequency of Climate Cluster 2, which exhibited high total precipitation, precipitation intensity, and heavy rain frequency in June and August. In the central region, Climate Clusters 1 and 2 dominated climate patterns after 2012, and although a significant upward trend in Cluster 2 was observed in this region, rain disaster hazards exhibited a nonsignificant upward trend due to the high frequency of Cluster 2, which is characterized by low precipitation.

4.1.4. Tropical Storms

Tropical storms considerably influence crop production [80,81,82] and cause substantial crop production losses annually in Taiwan [83]. Nevertheless, from 2003 to 2022, both the frequency and hazard index values of tropical storms exhibited a substantial downward trend. Additionally, the tropical storm hazard index value decreased consistently across Taiwan’s four regions. The high frequency of Clusters 1, 2, 9, and 10 after 2010, which are characterized by low total precipitation, precipitation intensity, and heavy rain frequency during Taiwan’s primary tropical storm season (July and August), contributed to the low frequency of tropical storms. A similar reduction in tropical storm frequency after 2010, attributable to the northward shift of prevailing tropical storm tracks caused by warming sea surface temperatures, was observed in southeastern China [84,85]. Additionally, several studies have indicated that tropical storm frequency is influenced by the El Niño Southern Oscillation [86,87,88]. In Taiwan, La Niña and its precursor years are associated with low tropical-storm-induced extreme precipitation frequency [31]. Since 2006, La Niña events have been frequent, which has caused decreasing tropical storm frequencies and hazard index values.

5. Conclusions

This study utilized a 20-year dataset to assess crop disaster hazards in Taiwan. The annual disaster hazard index values for each county were quantified on the basis of the number and severity of disasters. Additionally, daily weather variables for each county were analyzed using multivariate analysis to reduce data dimensions and generate climate clusters. Spatiotemporal variations in climate clusters were also examined to investigate changes in climate patterns from 2003 to 2022. Finally, the association between disaster hazards and climate patterns was assessed, and the following conclusions were reached:

• Although the frequency of tropical storm disasters decreased, the frequency of other disasters exhibited a significant upward trend due to the increasing frequency of cold wave and rain disasters.
• Overall, the disaster hazard index values did not increase from 2003 to 2022.
• The increasing hazard index values for cold waves and rain disasters indicate an urgent requirement for mitigation and adaptation strategies to manage them.
• Given the variations in the disaster patterns and hazard trends across Taiwan’s four regions, region-specific strategies are required to address disaster hazards.
• The results of multivariate analysis revealed that temperature and other weather information, such as that regarding precipitation and wind speed for each month, characterize climate patterns.
• The climate in Taiwan, especially in summer, has shifted from wet to dry.
• Climate clusters significantly affect the hazard index values of all disasters.
• The primary climate clusters after 2012 exhibited a relatively high hazard index value for cold waves, droughts, and rain disasters.