Intelligent identification of clustered landslides based on semantic segmentation method

Clustered landslides induced by earthquake and rainfall are characterized by large quantity, high frequency, uneven spatial distribution, and extensive impact areas. Traditional landslide identification relies primarily on visual interpretation of remote sensing imagery, which depends on the expertise of professionals and the analysis of pre- and post-event image textures. Although these methods can yield reasonable accuracy, they are labor-intensive and struggle to provide rapid and large-scale assessments, resulting in inefficiency and high operational costs. To date, with the rapid development of artificial intelligence, a semantic segmentation method based on data-driven and deep learning has been widely used in the automatic detection of clustered landslides. Taking Niangniangba Town, Qinzhou District, Tianshui City, Gansu Province, as a case study, this study established a comprehensive landslide inventory containing 3,026 rainfall-triggered landslides through systematic visual interpretation of Google Earth satellite imagery. The full-scene images were subsequently cropped into 3,026 patches (1,200×1,200 pixels). These patches were then stratified and randomly allocated into training (2,526), validation (250), and test sets (250). The PSPNet semantic segmentation model was employed to automatically identify rainfall-induced landslides in the study area. The model obtained at the 87,000th iteration was selected as the best model. On the validation set, the Precision, F1 score, Recall, and Intersection over Union (IoU) were 0.903, 0.821, 0.752, and 0.696, respectively. On the test set, these metrics were 0.920, 0.865, 0.816, and 0.761. Among the test samples, 204 landslides were correctly identified, 18 were falsely detected, and 28 were missed. The proposed intelligent recognition model based on semantic segmentation demonstrates high accuracy in identifying clustered landslides of varying scales and morphologies, with strong generalization capability. This approach provides an effective and efficient solution for the rapid and precise identification of clustered landslides.