Automatic detection of sinkhole collapses at finer resolutions using a multi-component remote sensing approach

Abstract

Sinkhole development is a typical geological disaster found in areas of carbonate bedrock. Compared with other geological disasters, sinkholes are considerably smaller and scattered according to scale and spatial distribution. Nevertheless, detecting and investigating sinkholes have become increasingly challenging. This study proposes a novel method by applying case-based reasoning (CBR) combined with object-based image analysis and genetic algorithms (GAs) to detect the sinkholes using high-resolution aerial images. This case study was performed in Paitan Town, Guangdong Province, China. The method comprises three major steps: (1) multi-image segmentation, (2) GA-based feature selection, and (3) application of CBR techniques. The detected sinkholes were categorized into three classes: buried, collapse type I, and collapse type II. The experiment demonstrated that the proposed method can obtain higher accuracy compared with the traditional supervised maximum likelihood classifier (MLC). The overall accuracy of CBR classification and MLC for the collapse area was 0.88 and 0.71, respectively. In addition, the kappa coefficient for CBR classification (0.81) was higher than that for MLC (0.5). A similar case library was also applied to another trial area for validation, the satisfactory results of which suggested that CBR is applicable for independently detecting sinkholes. The proposed method will be useful for preparing hazard maps that express the relative probability of a collapse in similar regions.