A small attentional YOLO model for landslide detection from satellite remote sensing images

The use of high-spatial-resolution remote sensing image technology on mobile and embedded equipment is an important and effective way for emergency rescue and evaluation decision-makers to quickly and accurately detect landslide areas. Deep learning-based landslide detection models include one-stage and two-stage models. The two-stage landslide detection models are slower. The one-stage landslide detection models are faster but less accurate. Both types of detection models have many parameters. This research aims to improve the speed, accuracy, and parameters of landslide detection models. A you only look once-small attention (YOLO-SA) landslide detection model is proposed. YOLO-SA is an improved version of the one-stage detection model YOLOv4. First, the group convolution (Gconv) and ghost bottleneck (G-bneck) residual modules are used to replace the convolution components and residual module consisting of standard convolution. The purpose is to reduce the parameters of the model. Then, on this basis, an attention mechanism is added to improve the detection accuracy of the model. Finally, the position of the attention mechanism is adjusted to determine the framework of YOLO-SA. Qiaojia and Ludian counties in Yunnan Province, China, are used as the study area to acquire three-channel (red, green, blue) historical landslide optical remote sensing images from Google Earth, with a total of 1818 images, for training the model. YOLO-SA is compared with 11 advanced models, including Faster-RCNN, 3 types of EfficientDet, 2 types of Centernet, SSD-efficient, and 4 types of YOLOv4 models. The results show that the number of YOLO-SA parameters is reduced to 1.472 mb compared to EfficientDet-D0; the accuracy is improved to 94.08% compared to Centernet-hourglass; and the speed is up to 42 f/s. In addition, the effectiveness of the YOLO-SA model for potential landslide detection is verified, with an F1 score of 90.65%.