SEEDS
Soil Erosion and land covEr Data for Sustainability
Mapping terraces and land cover from space to predict rainfall-induced soil erosion in Italy.
The project
SEEDS is a project funded by the International Foundation Big Data and Artificial Intelligence for Human Development (IFAB). Its objective is to improve the assessment and prediction of rainfall-induced soil erosion in Italy, using Big Data Analytics and Machine Learning (ML) applied to Earth Observation (EO) data, primarily leveraging Copernicus services.
The project addresses two specific challenges: developing methods for a high spatial and temporal resolution estimation of rainfall erosivity, overcoming the current limitations, and analysing the dynamic landscape characteristics that influence soil susceptibility to erosion (vegetation cover, soil properties, and seasonal, phenological, and anthropogenic variations).
SEEDS sits at the intersection of three of the UN Sustainable Development Goals (SDGs): SDG 13 (Climate Action), SDG 15 (Life on Land) and SDG 11 (Sustainable Cities and Communities). A reminder that what happens to a hillside in the Apennines connects directly to flood risk downstream and to the long-term viability of the landscapes around it.
MEEO’s role
MEEO's involvement focuses on the semantic segmentation of Sentinel-2 satellite images to map urban infrastructure and agricultural terraces in the Lattari Mountains (Campania Region, Southern Italy) and the Idice river basin (Tuscan-Emilian Apennines, central-northern Italy).
- Pre-processing: Sentinel-2 images (bands B2, B3, B4, B8) are sharpened from 10 m to 3.3 m through super-resolution and combined with a Digital Elevation Model to build a richer, multi-layer input (NRGB + DEM). To train the model, reference maps are created by combining open geographic datasets (QuickOSM and Corine Land Cover) with manual visual interpretation of the terrain for terrace layers. These maps assign a land cover label to each pixel and, once aligned with the satellite images, tell the model what it should learn to recognise.
- Modelling: two models are trained on this data: a classic U-Net and the IBM-NASA Prithvi foundation model, both adapted to work with multi-source satellite inputs. The result is a detailed map of the land surface, classified into 8 categories (background, forests, fields, buildings, terraces, roads, rivers, and water) allowing precise identification of landscape features at local scale.
These maps feed into high-resolution soil erosion models and support the analysis of land surface changes, with direct reference to the 2023 flood events.
- 3.3 m
- Super-resolved Sentinel-2 input, from 10 m
- 8
- Land cover classes in the output map