**1. Introduction**

Mountains and hills make up more than 60% of the total area of Hunan province in China, half of which have slopes greater than 25◦ [1]. This area has high rainfall, so landslide disasters are frequent. According to statistics, 2449 various geological disasters occurred in Hunan Province in 2020, causing economic losses of 262.49 million RMB, of which 2116 were landslide disasters, accounting for 86.4% [2]. Deploying multiple types of sensors on landslides to gather information on deformation, rainfall, stress, and other physical parameters, and providing timely warning, are low-cost and reliable prevention methods that can effectively reduce casualties [3–5]. With the development of sensor technology and Internet of Things technology, landslide monitoring is gradually developing towards the direction of automation and intelligence [6–9]. It is of great significance to fully mine extensive monitoring data and extract and identify warning precursors for studying the mechanisms of landslide disasters and improving the accuracy of warning.

Early and accurate identification of landslide precursors is a prerequisite for early warning. The traditional precursors that can be used for early warning are mainly macroscopic phenomena such as surface cracks, slope toe uplift and other macro phenomena [10–12]. With the development of monitoring technology, landslide precursors can be mined from abundant monitoring data, of which the most widely used type of data is surface deformation. The accelerated deformation process of landslides is the most intuitive and reliable precursor, so it is widely used in the study of landslide early warning. Xu et al. [5,13] proposed to use the normalized tangent angle as an indicator for early warning of landslides.

**Citation:** Xu, J.; Bai, D.; He, H.; Luo, J.; Lu, G. Disaster Precursor Identification and Early Warning of the Lishanyuan Landslide Based on Association Rule Mining. *Appl. Sci.* **2022**, *12*, 12836. https://doi.org/ 10.3390/app122412836

Academic Editors: Jinrong Jiang, Yangang Wang and Yuzhu Wang

Received: 9 November 2022 Accepted: 12 December 2022 Published: 14 December 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

Jeng et al. [14] proposed to use displacement-velocity ratio as an indicator for landslide warning. Valletta et al. [15] proposed a multicriteria approach to identify accelerated deformation processes in landslides. Bai et al. [16] proposed a hybrid warning algorithm that could identify the landslide acceleration process quickly, automatically, and accurately in an online monitoring and warning system, and achieved the balance of warning immediacy, accuracy, and computational resources through different strategies.

Although displacement, as a precursor of landslide disaster, can give early warning quickly and accurately, it also has many shortcomings. First, the current sensors for displacement monitoring are highly susceptible to environmental influences and often generate false alarms during the warning process [16–18]. Second, displacement is the result of a combination of multiple factors, both internal and external to the landslide. The acceleration of displacement foreshadows the initiation of the landslide process, and the warning window is very short [19–21]. Finally, the use of a single displacement characteristic for early warning does not take into account the impact of external trigger factors such as rainfall, earthquakes, and construction on the disaster, and is therefore necessarily incomplete.

The development of data mining technology in recent years has provided new research ideas for landslide precursor identification. Data mining technology can filter and analyze useful information and important events from massive data to reveal the internal relationships and hidden rules of data, which have been widely used in the commercial [22,23], industrial [24,25], engineering [26,27], medical [28–30] and educational [31,32] fields with remarkable effect. The application of data mining techniques in the field of landslides is mainly focused on susceptibility assessment [33–35], aiming to analyze landslide instability risk at the regional scale, while there are very few studies on application in specific landslide monitoring. Ma et al. [36,37] first used modern data mining techniques integrating two-step clustering, association rule mining, and decision trees to analyze data from the Majiagou landslide and the Zhujiadian landslide in the Three Gorges reservoir area. These studies not only identified landslide disaster factors but also realized the prediction of displacement evolution, which was the earliest research to carry out data mining for single landslide monitoring. Miao et al. [38] and Guo et al. [39] adopted the same data mining technology to analyze the trigger factors of the Baishuihe landslide and the Shuping landslide in the Three Gorges Reservoir area, and determined the warning threshold. All these studies have fully and comprehensively considered the correlation between multi-source monitoring data and provided causal relationships between different monitoring variables, which are very helpful for the analysis of landslide damage mechanisms and instability patterns. Most of these studies focused on reservoir landslides in the Three Gorges region of China, with monitoring data collected over several years and on a monthly scale. Therefore, these studies were more focused on the long-term deformation patterns of landslides. However, the daily-scale or even hourly-scale short-term deformation patterns of landslides are equally important in landslide early warning studies. Such short-term deformation patterns contain more reliable precursors of landslide disasters than deformation features, which are important for early warning decisions. In addition, these studies all adopted a two-step clustering algorithm, which is a kind of hierarchical clustering and divides clusters through the process of splitting or clustering, so there is no need to determine the number of clusters. However, for the clustering of daily or even hourly monitoring data, we prefer to flexibly adjust the number of clusters. This kind of data is very complex, and human subjective judgment is still needed. At this time, partition clustering represented by k-means is more appropriate.

The purpose of this paper was to mine the short-term deformation patterns of landslides, identify the precursors of landslides, and obtain more reliable early warnings. In this study, the Lishanyuan Landslide in Hunan Province was taken as the case study. First, the sliding window method was used to extract features from the original monitoring data, then the k-means algorithm optimized by particle swarm optimization (PSO) was used to cluster the features and construct the item set, and the Apriori algorithm was finally

used to mine the association rules between different features and determine the short-term deformation pattern of landslides according to the given confidence levels to analyze the precursors of landslide disasters and provide early warnings.
