Landslide detection based on efficient residual channel attention mechanism network and faster R-CNN

Abstract

Accurate landslide detection plays an important role in land planning, disaster prediction and disaster relief. At present, field investigation and exploration based on professional personnel is the most widely used landslide mapping and detection technology, but this method consumes a lot of manpower and material resources and is inefficient. With the development of artificial intelligence, landslide identification and target detection based on deep learning have attracted more and more attention due to their remarkable advantages over traditional technologies. It is a technical problem to identify landslides from satellite remote sensing images. Although there are some methods at present, there is still room for improvement in the target detection algorithm of landslides against the background of the diversity and complexity of landslides. In this paper, target detection algorithm models such as Faster R-CNN apply to landslide recognition and detection tasks, and various commonly used recognition and detection algorithm network structures are used as the basic models for landslide recognition. Efficient residual channel soft thresholding attention mechanism algorithm (ERCA) is proposed, which intends to reduce the background noise of images in complex environments by means of deep learning adaptive soft thresholding to improve the feature learning capability of deep learning target detection algorithms. ERCA is added to the backbone network of the target detection algorithm for basic feature extraction to enhance the feature extraction and expression capability of the network. During the experiment ERCA combined with ResNet50, ResNet101 and other backbone networks, the objective indicators of detection results such as AP50 (Average Precision at IOU=0.50), AP75 (Average Precision at IOU=0.75) and AP (Average Precision) were improved, and the AP values were all improved to about 4%, and the final detection results using ResNet101 combined with ERCA as the backbone network reached 76.4% AP value. ERCA and other advanced channel attention networks such as ECA (Efficient Channel Attention for Deep Convolutional Neural Networks) and SENet (Squeeze-and-Excitation Networks) are fused into the backbone network of the target detection algorithm and experimented on the landslide identification detection task, and the detection results are that the objective detection indexes AP50, AP75, AP, etc. are higher for ERCA compared with other channel attention, and the subjective detection image detection effect and feature map visualization display are also better. We released our code at: https://github.com/fluoritess/Efficient-residual-channel-attention-mechanism-network-and-Faster-R-CNN.