ABCNet: Attentive bilateral contextual network for efficient semantic segmentation of Fine-Resolution remotely sensed imagery

• Published in: ISPRS journal of photogrammetry and remote sensing Vol. 181; pp. 84 - 98
• Main Authors: Li, Rui, Zheng, Shunyi, Zhang, Ce, Duan, Chenxi, Wang, Libo, Atkinson, Peter M.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2021
• Subjects: Attention Mechanism; Bilateral Architecture; Convolutional Neural Network; data collection; Deep Learning; economic analysis; environmental protection; neural networks; photogrammetry; precision agriculture; remote sensing; satellites; Semantic Segmentation; Semantic Segmentation; Convolutional Neural Network; Bilateral Architecture; Deep Learning; Attention Mechanism
• ISSN: 0924-2716; 1872-8235
• DOI: 10.1016/j.isprsjprs.2021.09.005

Semantic segmentation of remotely sensed imagery plays a critical role in many real-world applications, such as environmental change monitoring, precision agriculture, environmental protection, and economic assessment. Following rapid developments in sensor technologies, vast numbers of fine-resolution satellite and airborne remote sensing images are now available, for which semantic segmentation is potentially a valuable method. However, because of the rich complexity and heterogeneity of information provided with an ever-increasing spatial resolution, state-of-the-art deep learning algorithms commonly adopt complex network structures for segmentation, which often result in significant computational demand. Particularly, the frequently-used fully convolutional network (FCN) relies heavily on fine-grained spatial detail (fine spatial resolution) and contextual information (large receptive fields), both imposing high computational costs. This impedes the practical utility of FCN for real-world applications, especially those requiring real-time data processing. In this paper, we propose a novel Attentive Bilateral Contextual Network (ABCNet), a lightweight convolutional neural network (CNN) with a spatial path and a contextual path. Extensive experiments, including a comprehensive ablation study, demonstrate that ABCNet has strong discrimination capability with competitive accuracy compared with state-of-the-art benchmark methods while achieving significantly increased computational efficiency. Specifically, the proposed ABCNet achieves a 91.3% overall accuracy (OA) on the Potsdam test dataset and outperforms all lightweight benchmark methods significantly. The code is freely available at https://github.com/lironui/ABCNet.