This is published under the terms of CC BY-NC.

Downloaded: 307 times

Abstract

This work proposes an Adaptive Multi-feature Balanced network (AMBNet) for semantic segmentation in complex urban remote sensing scenarios. To fully exploit optical images and Digital Surface Models (DSM) data obtained from remote sensing sensors, a Depth Feature Extraction and Balancer (DFEB) module is devised to estimate and balance the depth information of all pixels by capturing detailed structural compositions of the ground surface. After that, a Parallel Multi-Stage Segmentator (PMSS) comprised of a dual-branch Encoder and Decoder with skip connections is constructed to perform effective segmentation by exploiting the balanced DSM (BDSM) and optical information. As a result, the proposed AMBNet can make effective use of optical images to complete depth information, so as to achieve multimodal informationassisted semantic segmentation for complex remote sensing scenes. Comprehensive experiments performed on the ISPRS Vaihingen and Potsdam remote sensing datasets confirm the segmentation performance of the proposed method.