Multi-Scale Dense Feature Fusion Based Loess Landslide Recognition

Abstract

Loess landslide geological disasters are widely distributed in Northwest China, but there are few relevant attention and researches. Landslide recognition can provide information help for landslide disaster management and risk management. Previous works of landslide recognition of remote sensing images based on deep learning, due to the lack of high resolution multi-source datasets, the boundary of landslide recognition is missing and not obvious and the identification accuracy is not ideal. In this work, a multi-scale dense feature fusion loess landslide recognition network (MDFF) was proposed and an open dataset of loess landslide samples (MSLLD) based on GF-2 images and DEM was constructed, which has spectral and topographic information. The MDFF network retains different levels of features by means of dense connection mechanism to make up for the loss of detailed features, the dense connected dilated convolution layer is introduced into the network to capture the different scale features of landslide images, expand the receptive field and avoid convolution degradation. When testing different networks on MSLLD, the proposed network achieves the most advanced performance, mIoU and F1-score were 82.31 % and 84.59% respectively, indicating that the proposed network can effectively recognize landslides, which is of great value for the investigation and analysis of loess landslide disasters.