Landslide detection using deep learning on remotely sensed images

Natural hazards such as landslides pose significant geological threats that can severely endanger the safety and property of residents in affected areas. Therefore, the prompt detection and accurate localisation of landslides are crucial. With the advancement of remote sensing technology and computational methods, artificial intelligence (AI)-based landslide detection techniques have emerged as effective solutions. Compared to traditional methods, these AI-driven approaches offer enhanced efficiency, accuracy and reliability, improving the speed and precision of landslide detection. They also provide valuable data for disaster prevention, mitigation and the assessment of landslide susceptibility and hazard levels. This study focuses on the western Sichuan region and constructs a historical landslide dataset using Google Earth imagery, which includes 4280 landslide samples (3424 for training and 856 for validation). To augment the dataset, 11 data augmentation techniques were applied, including copy–paste, random horizontal flipping, mosaic, random rotation, random hue, saturation and value transformation, affine transformation, random Gaussian noise, random scaling, random brightness and contrast adjustment, mixup and random cropping. These methods improve the diversity of landslide data, helping deep learning models capture more comprehensive global and local information during optimisation. This research utilises the YOLOv10-n object detection framework, enhanced with RepBlock from EfficientRep, FusedMBConv and MBConv techniques derived from EfficientNetV2, CSCGhostblockv2 from GhostNetv2, CReToNeXt from Damo-YOLO and CSCFocalNeXt. These innovations explore the impact of different backbone architectures on model performance. Additionally, the model incorporates four distinct attention mechanisms—convolutional block attention module (CBAM), global attention mechanism(GAM), sim attention module(SimAM) and selective kernel(SK) attention—to assess their influence on detection accuracy. The detection heads are optimised by substituting with three alternatives—DynamicHead, adaptive spatial feature fusion and real-time detection transformer—to enhance feature integration and investigate their effect on model performance. The results indicate that combining EfficientNetV2 with CBAM and v10Detect yields the highest performance. When applied to the historical landslide dataset from the western Sichuan region, the YOLO-EfficientNetV2 model achieves an average precision of 0.861 and an F1 score of 0.82, with a model size of 5.54 M. This model demonstrates superior capability in accurately identifying landslide locations, addressing the common challenge of balancing detection precision and speed in traditional object detection models, while also reducing parameter size and increasing detection speed.