Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{203859,
        author = {Mihir Mohite and Akshat Yadav and Varsha Dange},
        title = {An Integrated Deep Learning Framework Using U-Net++ and Hybrid Loss for Class-Imbalanced Semantic Segmentation in Satellite Imagery},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {323-331},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=203859},
        abstract = {Accurate semantic segmentation of satellite imagery is essential for environmental monitoring, agricultural planning, water-resource assessment, and climate-driven land-use analysis. However, a major challenge persists in Deep Learning–based Earth Observation: class imbalance, where small but environmentally critical classes (e.g., water bodies, wetlands, deforested patches) appear far less frequently than dominant classes such as agricultural or barren land. Traditional U-Net models trained with Cross-Entropy loss tend to ignore such minority classes, resulting in low recall, imprecise boundaries, and unreliable predictions for rare features.
This research proposes a comprehensive deep-learning framework integrating U-Net++ architecture, class-aware sampling, targeted augmentation, and a three-part hybrid loss function combining Area-Weighted Binary Cross Entropy (AWBCE), Tversky Loss, and Boundary Loss. Training stability is further enhanced using the Ranger optimizer, which combines RAdam and Lookahead to better navigate complex loss landscapes. Experiments conducted on LISS-IV satellite imagery demonstrate substantial improvement in overall accuracy, mean IoU, and—most importantly—minority-class IoU. Water-body IoU improves by more than 23%, and the proposed method outperforms standard U-Net by 7.3% mean IoU.
The results validate that addressing class imbalance requires multi-level redesigns across data, architecture, optimization, and loss formulation.},
        keywords = {Earth Observation, semantic segmentation, U-Net++, hybrid loss, minority-class detection, remote sensing, deep learning.},
        month = {June},
        }