"Rainfall-induced Clustered Landslide Dataset"

"Intense rainfall often triggers co-occurring clustered landslides, which are characterized by their large quantity, high spatial density, and extensive impact areas, posing severe threats to human lives and infrastructure. Rapid and accurate identification of such clustered landslides is crucial for effective disaster response and risk assessment. In recent years, deep learning techniques have demonstrated significant potential in the automatic detection of landslides from remote sensing imagery. However, the lack of high-quality annotated datasets specifically targeting clustered landslide events has become a major limitation, hindering the performance and generalization ability of deep learning models in this field. To address this gap, we developed a Rainfall-induced Clustered Landslide Dataset (RCLD) based on high-resolution remote sensing imagery. The dataset encompasses three clustered landslide events triggered by extreme rainfall in China in 2024, comprising 3,297 images (512 \u00d7 512 pixels), with an average of 19.27 landslides per image. To evaluate the effectiveness of the dataset, three widely used deep learning models\u2014FCN, U-Net, and DeepLabV3+\u2014were trained and tested on RCLD. Furthermore, the generalization capability of models trained on RCLD was assessed by comparing their performance with models trained on several publicly available landslide annotated datasets. The results indicate that models trained on RCLD achieved superior performance, with an Intersection over Union (IoU) of 68.10% and an F1-score of 81.02%, significantly outperforming those trained on existing datasets. This study highlights the substantial value of the RCLD dataset in enhancing the automatic detection of clustered landslides induced by extreme rainfall events, which is particularly meaningful in the context of increasingly frequent climate-related geohazards."