Semantic Labeling in Remote Sensing Corpora Using Feature Fusion-Based Enhanced Global Convolutional Network with High-Resolution Representations and Depthwise Atrous Convolution

Abstract

This paper addresses improving semantic segmentation in remote sensing for aerial and satellite images, which is crucial for agriculture, map updates, route optimization, and navigation. We propose enhancements to the state-of-the-art Enhanced Global Convolutional Network (GCN152-TL-A) by introducing a High-Resolution Representation (HR) backbone for better feature extraction, Feature Fusion (FF) to capture low-level details, and Depthwise Atrous Convolution (DA) for refined multi-resolution features. Experiments on Landsat-8 and ISPRS Vaihingen datasets demonstrate our model’s superior performance, achieving over 90% accuracy in F1 scores and outperforming baseline models.