Landslide Monitoring Using State of the Art Machine Learning Techniques with Image Processing of Space Borne Remote Sensing Imagery Data

Abstract

Landslides present significant worldwide difficulties, leading to profound repercussions such as property destruction and human death. As a result, the methods we currently have, including light detection and ranging (LiDAR) and on-site assessments by experts to look at things like topography, ground fractures, slope stability, and changes in vegetation, are not good enough to accurately predict and prevent these events. The rising frequency of landslides worldwide, intensified by alterations in weather patterns and human actions, underscores the necessity for sophisticated alert and surveillance tools. Landslides can result in devastating outcomes, such as fatalities, displacement of individuals, and significant economic repercussions. Technological advancements, namely in machine learning, remote sensing, and image processing, provide promising options to improve landslide prediction and mitigation efforts. This work aims to address the critical issue of evaluating and predicting landslides, with the objective of creating a dependable approach. We extract landslide causative factors from remote sensing imagery data through image processing, compile them into a single dataset, and then use this dataset to train machine learning algorithms. The landslide causative factors included in our dataset are the Topographic Wetness Index (TWI), Peak Ground Acceleration (PGA), Stream Power Index (SPI), Terrain Ruggedness Index (TRI), Normalized Difference Vegetation Index (NDVI), curvature, elevation, faults, rainfall, geology, solar radiation, aspect, landcover, slope, streams, and roads. We tested how well different machine learning algorithms could predict the risk of landslides. These included Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), AdaBoost, Long Short-Term Memory (LSTM) networks, Deep Neural Networks (DNN), NGboost, and 1D Convolutional Neural Networks 1D (CNN 1D). The results demonstrate significant accuracy scores for each model, with the LSTM network emerging as the most efficient, showcasing an accuracy of 0.91 and an amazing Area Under the Curve (AUC) score of 94%. The success of the LSTM can be due to its proficiency in managing sequential and time-dependent data, which is essential for comprehending and predicting the dynamic characteristics of landslides. The LSTM’s ability to utilize temporal patterns improves the effectiveness of early warning systems, representing a notable progress in landslide prediction approaches.