Modular and adaptive implementation of Semantic Segmentation Models for Satellite Images and Open Source tools suitable for complex geographical contexts

A modular, open-source framework for semantic segmentation of satellite images demonstrates that simple deep learning models can outperform complex ones when applied to small, imbalanced datasets like those found in coastal wetlands. The framework, which bridges Geographic Information Systems (GIS) and Deep Learning (DL), integrates QGIS for spatial preprocessing with Python-based training pipelines, making advanced segmentation accessible to non-specialists. Key innovations include hybrid loss functions with dynamic class weighting and spectral-consistent data augmentation, which mitigate class imbalance—a persistent challenge in ecological detection. Case studies on the Crozon Peninsula and Mont-Saint-Michel Bay highlight its efficacy in multi-class land cover classification and binary salt marsh detection using modest computational resources, promoting sustainable AI applications in environmental monitoring.

Le Guillou, A., & Niculescu, S. (2026). Modular and adaptive implementation of Semantic Segmentation Models for Satellite Images and Open Source tools suitable for complex geographical contexts. International Journal of Applied Earth Observation and Geoinformation, 146, 105069. https://doi.org/10.1016/j.jag.2025.105069