Landslide Detection Based on Improved YOLOv5 and Satellite Images

Since the 21st century, due to the repeated deterioration of the natural climate and the increasing impact of human production activities on the ecological environment, landslides have become a common high-hazard natural phenomenon. Traditional landslide detection is mainly done through field detection, synthetic aperture radar, and other technologies. With the increase in the accuracy of satellite imagery and the rapid development of deep learning, the use of deep learning to realize landslide detection has gradually become a trend. Our work mainly includes two parts: landslide data set production and model performance improvement. We have produced a complete general landslide dataset based on the high-resolution remote sensing images obtained from open satellites and the annotations of professional researchers. We will publish our dataset later. Based on the research of previous researchers and based on the basic framework of YOLOv5, we improved the feature splicing method of YOLOv5 by adding Adaptively Spatial Feature Fusion (ASFF), and fused feature information of different scales to improve the model. To better mine shallow feature information, we introduced the Convolutional Block Attention Module (CBAM) module to improve the performance of the model. Experiments have proved that our model has a 1.64% performance improvement.