Evaluating convolutional neural networks and machine learning algorithms for accurate landslide detection from satellite data

Landslides threaten human life, infrastructure, and environmental stability, necessitating rapid response systems enabled by accurate and timely detection. This study carried out a comparative evaluation of the deep learning methods against the traditional machine learning algorithms in the detection of landslides using satellite remote sensing imagery. The proposed CNN-based architecture, comprising VGG and ResNet models, is compared with conventional ML algorithms, namely, Random Forest, Support Vector Machine, Decision Tree, K-Nearest Neighbors, and Logistic Regression. Experiments were carried out on two benchmark datasets Recent Landslide Database and Relict Landslide Database using False Positive Rate (FPR) and Matthews Correlation Coefficient as key metrics. Results reveal that the proposed ResNet gives the minimum FPR (0.0683% and 0.1296% for RLD and LLD, respectively) and maximum MCC values (0.7012 for RLD and 0.724 for LLD), thus besting all traditional models. VGG gives competitively very high MCC scores with stable accuracy. RF offers good accuracy but suffers from more false positives, while SVM and KNN offer the worst of classification results, especially on LLD. Additionally, an Average Relative Predictor Importance (ARPI) study is performed that distinguishes slope gradient and curvature as the most important features for landslide prediction.