SemanticSeg4EO: An open-source unified framework and QGIS plugin for semantic segmentation in Earth Observation

Semantic segmentation of satellite and aerial imagery is a fundamental task in Earth Observation (EO). However, exploiting these data with state-of-the-art deep learning methods remains out of reach for most Geographic Information Systems (GIS) practitioners due to programming barriers and infrastructure complexities. While tools such as Deepness, GeoAI, and QLearn have begun to bridge this gap, they either lack a full training pipeline or suffer from dependency conflicts when loading deep learning libraries directly into the QGIS environment. This paper presents SemanticSeg4EO, a unified, fully open-source framework that bridges this gap. It provides both a Python command-line toolbox for reproducible research workflows and a QGIS plugin that makes deep learning accessible via a graphical interface. A key technical contribution is the environment isolation strategy: the plugin delegates all computation to an external Python environment via subprocess calls, preserving QGIS stability while providing full access to GPU-accelerated deep learning. The framework supports an end-to-end pipeline from georeferenced patch extraction to large-scale prediction with Gaussian blending. With support for over 20 architectures and 15 loss functions, SemanticSeg4EO ensures transparency and reproducibility in environmental monitoring.