Benggang segmentation via deep exchanging of digital orthophoto map and digital surface model features

Abstract

Benggang is a typical fragmented erosional landscape in southern and southeastern China, posing significant risk to the local residents and economic development. Therefore, an efficient and accurate fine-grained segmentation method is crucial for monitoring the Benggang areas. In this paper, we propose a deep learning-based automatic segmentation method for Benggang by integrating high-resolution Digital Orthophoto Map (DOM) and Digital Surface Model (DSM) data. The DSM data is used to extract slope maps, aiming to capture primary morphological features. The proposed method consists of a dual-stream convolutional encoder-decoder network in which multiple cascaded convolutional layers and a skip connection scheme are used to extract morphological and visual features of the Benggang areas. The rich discriminative information in the DOM and slope data is fused by a channel exchanging mechanism that dynamically exchanges the most discriminative features from either the DOM or DSM stream according to their importance at the channel level. Evaluation experiments were conducted on a challenging dataset collected from Guangdong Province, China, and the results show that the proposed channel exchanging network based deep fusion method achieves 84.62% IoU in Benggang segmentation, outperforming several existing unimodal or multimodal baselines. The proposed multimodal segmentation method greatly improves the efficiency of large-scale discovery of Benggang, and thus is important for the management and restoration of Benggang in southern and southeastern China, as well as the monitoring of other similar erosional landscapes.