Semantic Segmentation of Super Resoluted Sentinel-2 Images for Urban and Agricultural Surface Mapping in Soil Erosion Studies

Identifying anthropic infrastructure and soil erosion systems is critical for analyzing the P-factor of RUSLE (Revised Universal Soil Loss Equation) formula which plays a crucial role in assessing the effectiveness of conservation practices at reducing soil erosion. SEEDs is a CMCC (Mediterranean Centre for Climate Change) project for IFOA (Training organization) focused on the production of high-resolution assessments of soil erosion risk, combining Earth Observation data and advanced analytics to support sustainable land management in Italy. The current work explores semantic segmentation of Sentinel-2 imagery, utilizing both its native 10-meter resolution and super-resolved 3-meter images, to map urban infrastructure (buildings and roads) and agricultural terraces critical for soil erosion analysis, within the context of SEEDs project.

It is focused on two distinct areas in Italy: the Lattari Mountains (Campania, southern Italy) and the Idice river basin (Emilia-Romagna, northern Italy), which was impacted by significant flood events in May 2023 and September 2024. These regions differ in landscape characteristic and erosion control measures, providing a valuable comparison to evaluate influence of these factors on erosion dynamics.
The methodology incorporates deep learning U-Net architectures, fine-tuned using Sentinel-2 multispectral data and elevation (DEM) data. Mask preparation for training and validation involves data from OpenStreetMap, Corine Land Cover and visual interpretation using QGIS software, specifically for Liguria terraces mapping, which were used as a unique training dataset for identifying terraced landscapes in the study areas, due to the scarcity of available labelled datasets.  

The analysis mapped key infrastructures in the study areas using both original resolution and super-resolved imagery.  In addition, preliminary results indicated differences in infrastructure before and after the flood events of 2023, suggesting potential impacts on both agricultural lands and urban areas. This approach demonstrates potential for precise urban and agricultural mapping in erosion-prone landscapes. Ongoing work focuses on refining model performance and validating results across diverse terrains and regions, ultimately enhancing soil erosion risk assessments and supporting more effective land management strategies.