Landslide Extraction Methods Based On Background Enhancement And Landslide-inducing Factors

Landslide occurs frequently worldwide.An accurate and quick detection of landslide can offer landslide’s location and the scope of its reach,then assist in releasing evacuation information,formulating emergency rescue plans,and assessing disaster loss.Landslide extraction techniques based on deep learning methods have alleviated problems like heavy artificial work and time consumption caused by other traditional approaches.However,in application,the quality of samples often fails the requirements of deep learning training,bringing adverse impact on landslide extraction accuracy.What’s more,on condition that landslide’s characters already show great diversity,some background objects also share high similarity in colors,textures,shapes with landslides.These features can be very confusing to landslide extraction models,causing a large amount of false and missed extractions.To solve the above problems,a background enhancement method was proposed to reduce false extractions by enriching the background characters of training samples.Considering that the environments of disaster areas play dominant role in the formation of landslides,landslide-inducing factors were used as supplements,providing landslide susceptibility information as a reference for landslide extraction models to further improve the accuracy of extraction results.Then,comparative experiments were conducted based on Mask R-CNN model and Yunnan Ludian landslide data,to validate the applicability and effectiveness of the proposed methods.The main contributions of this study are as follows:(1)Building landslide samples using post-disaster optical satellite images.Conducting landslide extraction experiment using Mask R-CNN model.The extraction result has large amount of false extractions on background objects and missed extractions on small landslides.By analyzing the samples,we concluded that samples directly splitted from satellite images often with simple background,leading to ineffective feature learning.(2)Developing background enhancement methods based on image splicing and region replacing.While increasing the amount of training data,the backgrounds of the samples are more complex,assisting models to distinguish landslide and background objects.With the help background enhanced samples,the F1 score was significantly improved by 16.60% compared with the traditional method.(3)Take Ludian as an example,selecting the factors having strongest relations with landslide occurrence through statistical analysis.Combining these landslide-inducing factors(DEM,slope,distance from river)with optical satellite images,providing landslide susceptibility information as a reference for landslide extractions.Comparative experimental results proved the effectiveness of using landslide-inducing information.Then,the experiment using both background enhancement method and landslideinducing information was conducted and achieved the best performance in this study.The accuracy,recall,F1 Score and m Io U of landslide extraction were 88.68%,89.49%,89.08% and 89.00% respectively,validating the applicability and effectiveness of the proposed methods.