Intelligent Detection Of Landslides With Multi-source Data

Landslides are a frequent and widespread natural disaster that causes significant destruction.They block roads,disrupt traffic,and hinder emergency relief efforts.It is crucial to identify specific landslide-prone areas as soon as possible after severe geological events.Additionally,cataloging landslides is essential for analyzing the reasons,assessing susceptibility,monitoring and early warning.Therefore,landslides detection on a large scale is therefore of great significance for research related to landslides,risk prevention and management,and emergency rescue operations.Currently,artificial intelligence methods have made significant progress in detecting disastersinduced landslides.However,identifying old landslides and unstable deforming landslides still heavily relies on expert visual interpretation.To address this issue,this paper utilizes synthetic aperture radar(SAR),digital elevation models(DEM),and high-resolution optical images as the primary data sources.The study focuses on constructing a multi-source landslide identification dataset and automatically identifying landslides in Bijie City in Guizhou,the Sichuan-Tibet Railway,and Yan’an City in Shanxi.Finally,a hierarchical landslide identification method was developed.The main work and contributions of this paper are as follows:Firstly,this paper proposes a mask region?based convolutional neural network(R-CNN)approach with simulated hard samples to address the challenge that landslides are easily be confused with farmland,bare ground,and river beach land.Experiments show that our method reduces false detections of landslides by 30% and increases the F1 score by 18% compared with Mask R-CNN.Secondly,this paper proposes a transfer learning approach for detecting old landslides by leveraging the parameters and characteristics of the model trained by new landslides,to address the challenges of indistinct image features and limited samples.The results demonstrate that compared to the baseline model,Mask R-CNN,the precision,recall,and F1 score have all improved by approximately 10%.Thirdly,this paper proposes an automatic detection method for unstable slopes that are deforming using the YOLO and RXD-UTD algorithms.This approach integrates deformation and geological information to detect loess landslide hazards across an area of 60,000 square kilometers,automatically identifying 4,932 slope hazard areas.Furthermore,we construct two landslide identification datasets.One is the optical image dataset along the Sichuan-Tibet Railway,the other dataset in the Yan’an area is integrating optical,topographic,and In SAR data,which significantly reduces the time and effort to collect and label samples and address the shortage of public landslide identification datasets available for deep learning.