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

Semantic segmentation, the process of assigning a semantic label to each pixel in an image, is a critical computer vision task for extracting detailed information from remote sensing data. However, its application to complex geographical contexts, such as coastal wetlands, is often constrained by the need for highly specialized implementations, class imbalance, and limited accessibility for non-specialists. This paper introduces a novel, modular, and adaptive open-source framework for semantic segmentation tailored to satellite imagery. Designed for maximum flexibility, the framework supports both binary and multi-class segmentation tasks and incorporates specific training strategies to handle severe class imbalances inherent in ecological detection, such as salt marsh mapping. The implementation provides a fully configurable pipeline that bridges the gap between Geographic Information Systems (GIS) and Deep Learning (DL). It integrates QGIS for intuitive spatial preprocessing and grid generation with a Python-based training and prediction workflow, thereby democratizing access to advanced segmentation techniques. The framework is architecture-agnostic, allowing the seamless deployment and benchmarking of various state-of-the-art encoder–decoder models, which are effective at combining multi-scale contextual information with high spatial resolution. A key contribution is the integration of a multifaceted training methodology that includes hybrid loss functions with dynamic class weighting and spectral-consistent data augmentation to ensure robust model generalization from limited and imbalanced datasets. We demonstrate the framework’s efficacy and scalability through two distinct case studies: a multi-class land cover classification on the Crozon Peninsula using Pléiades and a binary salt marsh detection in the Mont-Saint-Michel Bay Sentinel-2 imagery. The results show that accurate segmentation can be achieved with modest computational resources, promoting more sustainable and ethical AI applications in environmental monitoring. This work provides a critical tool for researchers and practitioners aiming to apply advanced DL segmentation to domain specific remote sensing challenges beyond conventional benchmarks. • Open-source QGIS-Deep Learning bridge for accessible segmentation. • Simple models outperform complex ones on small satellite datasets. • Hybrid loss and spectral augmentation mitigate class imbalance. • Accurate monitoring uses modest, sustainable computing resources.