A deep learning based framework for remote sensing image ground object segmentation

doi:10.1016/j.asoc.2022.109695

Return

A deep learning based framework for remote sensing image ground object segmentation

2022-11-01

PRE

DOI：10.1016/j.asoc.2022.109695

Original

Original request for help

Save

Abstract

En 中文

Semantic segmentation of very-high-resolution (VHR) remote sensing images is of great significance, in which remote sensing can be applied to numerous fields. However, VHR remote sensing images are taken at different seasons and regions, causing the large intra-class and low inter-class variations of pixels. Thus, the state-of-the-art semantic segmentation network have considerable misclassifications and blurring of object boundaries. To solve the above problems, a deep learning-based semantic segmentation framework (DLSS) of VHR remote sensing images is proposed in this study, which comprising three stages. At the pre-processing stage, a novel data pre-processing method named Image Block Segmentation (IBS) is proposed to coarse segmentation of VHR remote sensing images at the image block scale. At the image segmentation stage, the different strategies are adopted to segment image blocks of different categories for fine segmentation. Through these two stages, this study implements a coarse-to-fine segmentation strategy, which reduces the phenomenon of misclassification by using different network models for low inter class variations of pixels. At the post-processing stage, a novel post-processing method termed Superpixel Cluster (SPC) is proposed to modify the segmentation results. SPC can capture fine details of objects and aggregating continuous pixels with similar characteristics into a set of superpixels, so as to ensure the boundary accuracy and internal consistency of ground object. Extensive experiments, including a comprehensive ablation study, confirm that IBS is capable of effectively reducing the misclassification, and SPC can correct the segmentation results significantly. The experimental results on the Gaofen Image Dataset (GID) suggest that the overall accuracy (OA) of the commonly utilized models combined with DLSS framework can increase, and the average is 2.05%. The code of DLSS is available at https://github.com/dxj620/Deep-learning-semantic-segementation.(c) 2022 Elsevier B.V. All rights reserved.

Keywords:

Convolutional Neural Networks

Semantic segmentation

Remote sensing images

Superpixel segmentation

Landcover classification