Soil assessment and restoration decision support
SOIL-AI predicts the condition of soil in burned areas from high-resolution aerial imagery and turns those predictions into site-specific restoration suggestions. Deep-learning models trained on multispectral aerial data and laboratory-analysed soil samples produce maps of pH, organic matter, carbon, nitrogen and other soil properties. A decision support system then combines these maps with fire severity, erosion risk and terrain data to flag where intervention is needed, and which agroforestry restoration practices fit each zone.
SOIL-AI was developed by indigo as Task 6.1 of TREEADS (A Holistic Fire Management Ecosystem for Prevention, Detection and Restoration of Environmental Disasters), an Innovation Action funded by the European Union under Horizon 2020 (Green Deal call, grant agreement 101036926, 2021–2025). It was built and validated on wildfire-affected areas in Central Greece in collaboration with the Lithuanian Research Centre for Agriculture and Forestry (LAMMC), the University of Girona (UdG) and the Mediterranean Agronomic Institute of Chania (MAICh), and is integrated as a service in the TREEADS platform restoration module.
For the people who restore land after a wildfire
Forest and land managers
Assess soil health across a burned area within days of a fire, without waiting for a full ground survey, and see exactly which zones need intervention.
Restoration stakeholders and forestry organisations
Receive case-specific agroforestry and erosion-mitigation suggestions for each intervention area, as map layers ready for planning.
Regional authorities and policy makers
Prioritise post-fire restoration budgets using replicable, data-driven vulnerability maps rather than uniform treatments.
Soil maps, vulnerability and restoration advice from one survey
Soil property prediction from aerial imagery
Deep neural networks and tree-based models trained on multispectral orthomosaics and laboratory soil analyses predict pH, organic matter, carbon, total nitrogen, CaCO₃, NO₃-N, texture and electrical conductivity at 1 m resolution.
Soil health vulnerability maps
Each predicted parameter is classified against expert- and literature-derived thresholds (e.g. pH 6.0–6.5, organic matter 2–4 %, carbon 2–6 %, total N 0.2–0.5 %) to show where soil has fallen outside the healthy range.
Fire severity and erosion risk assessment
Sentinel-2 dNBR fire severity and RUSLE-based soil erosion risk (rainfall, erodibility, slope, cover and practice factors) identify the zones most prone to degradation.
Agroforestry restoration suggestions
A decision support system evaluates whether an intervention is triggered and matches each zone with suitable practices: immediate actions (leaving burned debris, mulching, barriers), soil-specific amendments (organic matter, liming, nitrogen-fixing microorganisms) and medium- to long-term vegetation establishment.
Web-GIS and API access
Predictions and intervention areas are exported as Cloud-Optimised GeoTIFF and GeoJSON layers and exposed through the Soil Characteristics and Restoration API for integration into the TREEADS platform or any web-GIS.
From a survey flight to restoration suggestions
- 01
Acquire aerial imagery
A light manned aircraft flies the burned area with an RGB camera and a Sequoia multispectral sensor (green, red, red-edge, near-infrared). Images are processed into orthomosaics and a digital elevation model at 1 m resolution.
- 02
Collect ground truth
Topsoil samples (0–15 cm) are taken across the burned area and analysed in the laboratory for pH, electrical conductivity, texture, organic matter, CaCO₃, nitrogen, phosphorus, potassium and micronutrients.
- 03
Predict soil properties
Spectral bands and derived indices (NDVI, MSAVI2, brightness, salinity and bare-soil indices, among others) are stacked and used to train deep neural networks and tree-based models that map each soil parameter across the whole area.
- 04
Assess vulnerability
Predicted parameters are classified against health thresholds and combined with Sentinel-2 fire severity (dNBR), RUSLE erosion risk, slope and aspect, land cover and vegetation (NDVI) to produce vulnerability maps.
- 05
Recommend restoration
For each zone the DSS evaluates whether an intervention is triggered (the most common outcome is "no intervention") and, where it is, proposes practices from an expert-curated table, exporting intervention areas as GeoJSON through the API.
Aerial and open data in, map layers and an API out
- High-resolution aerial multispectral imagery (user input, GeoTIFF, 1 m)
- Laboratory soil sample analyses (CSV, GeoJSON, Shapefile)
- Copernicus Sentinel-2 imagery, via Google Earth Engine (10–20 m)
- EFFIS burned-area perimeters (GeoJSON)
- Copernicus CORINE land use / land cover (100 m)
- ESDAC–JRC soil datasets: rainfall erosivity (R), erodibility (K), practice factor (P)
- LUCAS topsoil survey data
- Digital elevation model, slope and aspect (from aerial data or public DEMs)
- Raster maps of predicted soil characteristics (Cloud-Optimised GeoTIFF)
- Classified vulnerability maps per soil parameter, fire severity and erosion
- Vector layers of intervention areas with proposed restoration practices (GeoJSON)
- Soil Characteristics and Restoration API (Python) for the TREEADS platform restoration module
- Deep learning (U-Net, FCN) and machine learning (Random Forest, CART) in TensorFlow/Keras
- Spectral indices for soil and vegetation
- dNBR fire severity from pre- and post-fire Sentinel-2
- RUSLE soil loss modelling
- Threshold-based decision rules from literature review and expert knowledge (LAMMC, UdG)
- QGIS, Global Mapper, Google Earth Engine, Agisoft Metashape, Docker, cloud storage