Landslide Extraction Using Mask R-CNN with Background-Enhancement Method

The application of deep learning methods has brought improvements to the accuracy and automation landslide extractions based on remote sensing images because techniques have independent feature powerful computing ability. However, in application, quality training samples often fails requirement for networks, causing insufficient learning. Furthermore, some background objects (e.g., river, bare land, building) share similar shapes, colors, textures with landslides. They can be confusing automatic tasks, contributing false missed extractions. To solve above problems, a background-enhancement method was proposed enrich complexity samples. Models learn differences between landslides more efficiently through background-enhanced samples, then reduce objects. Considering that environments disaster areas play dominant roles formation landslides, landslide-inducing attributes (DEM, slope, distance from river) were used as supplements, providing additional information extraction models further improve results. applied extract occurred Ludian county, Yunnan Province, August 2014. Comparative experiments conducted using mask R-CNN model. experiment both showed satisfying result an F1 score 89.08%. Compared only satellite input data, it significantly improved by 22.38%, underscoring applicability effectiveness our method.