A Lightweight Multilevel Feature Fusion approach for Accurate Semantic Segmentation of Remote Sensing Data

Abstract

Remote sensing has become essential for land cover classification, leveraging deep learning-based segmentation to improve accuracy in environmental monitoring, agriculture, and urban planning. This study introduces Disc Net, a streamlined DeepLabv3+ adaptation optimized for semantic segmentation on the OpenEarthMap dataset. The model employs a structured pipeline, integrating InceptionV3 with squeeze-excitation (SE) attention in the encoder to enhance feature extraction, while multilayer feature fusion combines information from both shallow a deep layers for richer representations. The decoder, built on DeepLabv3+ with depthwise separable convolutions, reduces computational complexity without compromising segmentation quality. Preprocessing with stratified sampling ensures balanced class representation, further refining model performance. DiscNet outperforms existing baselines, achieving an mIoU of 79.6%, significantly surpassing U-Net (52.7%), simpleCNN (51.3%), and ResUNet (42.5%). This strong performance, validated through qualitative and quantitative assessments, demonstrates DiscNet’s ability to deliver high-precision segmentation while maintaining efficiency, making it a promising solution for large scale land cover mapping.