An Improved Segmentation Method for Automatic Mapping of Cone Karst from Remote Sensing Data Based on DeepLab V3+ Model

Han Fu,Bihong Fu,Pilong Shi

doi:10.3390/rs13030441

Abstract

The South China Karst, a United Nations Educational, Scientific and Cultural Organization (UNESCO) natural heritage site, is one of the world’s most spectacular examples of humid tropical to subtropical karst landscapes. The Libo cone karst in the southern Guizhou Province is considered as the world reference site for these types of karst, forming a distinctive and beautiful landscape. Geomorphic information and spatial distribution of cone karst is essential for conservation and management for Libo heritage site. In this study, a deep learning (DL) method based on DeepLab V3+ network was proposed to document the cone karst landscape in Libo by multi-source data, including optical remote sensing images and digital elevation model (DEM) data. The training samples were generated by using Landsat remote sensing images and their combination with satellite derived DEM data. Each group of training dataset contains 898 samples. The input module of DeepLab V3+ network was improved to accept four-channel input data, i.e., combination of Landsat RGB images and DEM data. Our results suggest that the mean intersection over union (MIoU) using the four-channel data as training samples by a new DL-based pixel-level image segmentation approach is the highest, which can reach 95.5%. The proposed method can accomplish automatic extraction of cone karst landscape by self-learning of deep neural network, and therefore it can also provide a powerful and automatic tool for documenting other type of geological landscapes worldwide.

Highlights

Published: 27 January 2021Karst landscapes are general term for the surface and underground landforms formed by the dissolution of water on soluble rocks
The mean intersection over union (MIoU) is a general standard for evaluating the accuracy of deep neural network segmentation, which means the ratio of the intersection and union of the ground truth value set and the predicted value set
By replacing the green band of the Landsat data by the digital elevation model (DEM) data, the influence of atmosphere on remote sensing image is reduced while the elevation information is included, the segmentation accuracy is higher than the results using other three-channel training samples

Summary

Introduction

Karst landscapes are general term for the surface and underground landforms formed by the dissolution of water on soluble rocks. They are widely distributed in the world and have a total area of 1.25 million km in China [1]. Most of karst areas have complex topography and geomorphic conditions, fragile ecological environment and poor accessibility, which significantly restrict the development of regional land use, urban planning, mineral survey, geological disaster prevention and control, ecological environment protection and tourism resources management [3]. Extracting the geomorphic information and spatial distribution of karst landscapes effectively is essential and significant for land use planning, ecological environment protection, and management for these heritage sites

Methods

Results

Discussion

Conclusion