A Scalable Open-Source System for Impervious Land Mapping Using GRASS and the Python Ecosystem

Abstract. This paper presents a scalable open-source system for national-level impervious-surface mapping and monitoring that combines GRASS, HDF5 and Python-based machine-learning libraries. Continuous Sentinel-2 multispectral monitoring is paired with targeted very high resolution (VHR) orthophoto segmentation to efficiently detect and delineate land transformation. Sentinel-2 imagery is retrieved from the Microsoft Planetary Computer, organized into a GRASS space-time raster dataset and classified using TorchGeo models fine-tuned on the EuroSAT dataset. Persistent transitions toward impervious-related classes identify disturbance candidates, and these areas trigger semantic segmentation of corresponding orthophoto tiles using GPU-accelerated Random Forests in RAPIDS cuML. The resulting outputs are vectorized, enriched with registry data and disseminated through spatial database services.

Applied to Slovenia for the 2020–2025 period, the system detected 99 km² of new impervious surfaces across a 20,271 km² study area, corresponding to 0.5 % land transformation (0.1 % annually). These results demonstrate that integrating continuous multispectral time-series analysis with event-driven VHR segmentation provides an efficient, reproducible and high-detail approach for operational land-change detection and environmental assessment.