Research on the application of large-scale convolution kernels and multi-scale fusion networks in landslide remote sensing image

Owing to the defects of semantic segmentation of deep convolution network and the confusion and multi-scale problems of landslides in remote sensing images, large-scale spatial separation convolution kernel and multiscale fusion semantic segmentation network is proposed. By using a large spatially separable convolution and channel attention mechanism on the encoder, the landslide image is extracted with large-scale information, which ensures the accurate extraction of landslide edge information; A skip connection is adopted between the encoder and the decoder to recover the context loss caused by the down sampling of the encoder; At the same time, the atrous spatial pyramid pooling (ASPP) module is applied to extract and fuse multi-scale features, so as to further improve the performance. The experimental results show that the segmentation effect of the proposed network on landslide dataset is better than FCN, SegNet, U-Net, DeeplabV3+ and other semantic segmentation methods, and it also verifies that the network has good landslide recognition ability in medium and high vegetation coverage areas. Experimental results demonstrate that the proposed network significantly outperforms existing semantic segmentation methods such as FCN, SegNet, U-Net, and DeepLabV3+ on landslide datasets, and exhibits strong landslide recognition capabilities in areas with medium to high vegetation coverage.