Prediction of Landslides Using Extracted Environmental and Topographic Features from Multispectral Satellite Images

Landslides are natural disasters that lead to the loss of lives and economic damage in a particular location. In India, particularly in hilly areas, landslides occur owing to factors such as rainfall, vegetation changes, changes in atmospheric conditions, and topographic features. The occurrence of landslides disrupts transportation, destroys infrastructure, and blocks roads for people near the location. Using satellite images, environmental and topographic features such as aspect, slope, curvature, contour, rainfall, soil moisture, atmospheric changes, and variations in vegetation index values can be extracted. These extracted features were used for landslide prediction, and the newly labeled dataset, comprising 1,738 landslides and 1,598 randomly generated non-landslide locations in India, was used for our experiment. Some previous studies considered either topographic or environmental features, whereas others used small datasets. In this work, both environmental and topographic features of all these locations were extracted from multispectral Landsat satellite images. A new dataset with features such as Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Built-up Index (NDBI), slope, aspect, elevation, and precipitation was prepared. Using this dataset, Artificial Intelligence (AI) models were built utilizing various Machine Learning (ML) algorithms, such as the Light Gradient Boosting Machine (LightGBM), Random Forest (RF), CatBoost, and Multilayer Perceptron (MLP) neural network models, to predict landslides. These models were assessed using confusion matrices, accuracy, Receiver Operating Characteristic (ROC) curves, and comparative metrics. RF, LightGBM, and CatBoost achieved the highest accuracy of 96%, and the MLP achieved an accuracy of 93%. The ROC–Area Under the Curve (ROC–AUC) score was 0.86 for the LightGBM model, which was the highest compared to the other three models. This work used multispectral satellite images and advanced ML models for reliable landslide predictions in India.