Mapping Soil Organic Carbon from Space and SkyFederated learning and hyperspectral drone imagery deliver sharper, privacy-friendly insights into soil carbon, from regional overviews down to within-field detail.
The Challenge
Soil organic carbon is one of the essential indicators of soil health. It improves soil structure, nutrient supply, and water retention, and it determines how much carbon our soils can lock away. With European initiatives such as the EU Soil Monitoring Law placing soil condition firmly on the agenda, the demand for dependable and detailed SOC maps keeps growing.
Creating such maps is anything but simple. Satellites like Sentinel-2 cover huge areas, but the models that turn their images into SOC values must be trained on laboratory results, and those are spread over many organizations. Since every soil sample is linked to a precise field location, data owners are frequently unwilling, or not legally allowed, to hand them over. Valuable measurements therefore stay locked away, and models built on a single dataset often struggle when applied elsewhere.
At field level, drones with hyperspectral cameras can record soil reflectance with a pixel size of roughly one centimeter. That level of detail creates a challenge of its own: a laboratory result typically comes from several soil cores mixed from an area of a few meters across, while one drone pixel only sees a tiny spot. Connecting the two correctly and testing the model in a way that does not overestimate its quality, is crucial for predictions that can be trusted.
Key Results
Federated learning: sharing knowledge, not data
Federated learning (FL) reverses the usual workflow. Instead of collecting all soil data on one central server, the model is sent to where the data is stored. Every partner trains it on its own samples and only sends back the updated model parameters, which a central server merges into an improved common model. The soil measurements and their coordinates stay with the organization that owns them at all times.
The Soil R&I Lab, with the support from the project technical provider INDRA, developed a federated learning pipeline starting from a European model: a convolutional neural network (CNN) trained on cropland samples from the LUCAS 2018 topsoil survey, linked to Sentinel-2 bare-soil reflectance composites for 2018–2020 and complemented with soil texture and bulk density information. This model was then further trained through FL on two private regional datasets: 211 samples from Flanders held by ILVO and more than 2,200 samples from Central Macedonia held by AUTH. About half of the Greek samples could be used, since many parcels are olive groves or orchards whose soil is rarely exposed.
The added value of working together is obvious when looking at the European model on its own, which showed a prediction error (RMSE) of more than 17 g kg-1 in Flanders (Table 1). The federated model brought this down to 4.7 g kg-1, close to a model built only on Flemish samples (4.2 g kg-1, Table 1). In Central Macedonia, the federated model even did slightly better than the purely regional one (3.3 versus 3.4 g kg-1, Table 1). Because the merging step weighs each partner by the number of samples it contributes, the much larger Greek dataset had more influence, which explains the small drop in accuracy for Flanders and highlights the value of more balanced contributions. The resulting model was applied to bare cropland in both regions, revealing broad regional patterns as well as differences within individual parcels (Figure 1 and Figure 2).
Table 1 Model performance on the independent test set for predicting topsoil SOC (g kg−1), comparing federated learning with baseline models. Figure 2. Close-up of the federated learning SOC map (g kg⁻¹) in Flanders, showing contrasts between and within individual cropland parcels. Hyperspectral drones: matching pixels to soil
samples
To zoom in further, the lab used a lightweight,
tunable hyperspectral camera from the project technology provider VTT mounted
on a drone. The wavelengths it records can be programmed before each flight;
six bands between 670 and 1100 nm were chosen for this work. Flying 80–100 m
above the ground, the system produced images with pixels slightly larger than
one centimeter. The flights took place over eight bare arable fields in
Merelbeke-Melle, Flanders (85 soil samples, autumn 2025) and five fields in Central
Macedonia (42 samples, winter 2026). The imagery was thoroughly checked before any
model was trained. Panels with known reflectance were laid out in every scene
for calibration, the Greek images were additionally compared with measurements
from a handheld field spectroradiometer, and pixels with missing or implausible
values were discarded. The main innovation lies in the way pixels are
associated with the soil samples. Rather than using only the pixel at the
sampling point, the team looked at all pixels within a circle of 0.25, 0.5 or 1
m around it and compared two strategies: - Aggregation condenses
the pixels in each circle into a single record per sample, using the average
reflectance of each band and how much it varies
- Augmentation treats every pixel in the circle as its own training example, labelled with the laboratory value of its sample, which greatly expands the training data.
Four machine learning methods were tested: Random Forest, XGBoost, a 1D CNN and TabPFN, a pretrained foundation model for tabular data (Table 2). To ensure statistically sound accuracy assessment, all pixels from the same sample were always kept together when splitting the data for cross-validation. The Flemish results were convincing. Based on the central pixel alone, the best models explained about half of the variation in SOC. Taking the surrounding pixels into account lowered the error in 23 of the 24 tested combinations, and 20 of those gains were statistically significant. The top performer was TabPFN with aggregation over a 1 m radius (R² = 0.78, RMSE = 0.16%, Table 2), reducing the error by more than a third. Augmentation improved every single model and was used to generate detailed predicted SOC maps at fields scale (Table 2, Figure 3). Table 2. SOC prediction accuracy (grouped 5-fold cross-validation) from VTT hyperspectral drone imagery in Flanders, comparing single-pixel extraction with the best neighbourhood configuration per model (radius in brackets). Figure 3. SOC map (%) of field RvK1A–B in Flanders predicted from VTT hyperspectral drone imagery with the Random Forest model using augmentation strategy, together with the laboratory-measured SOC values at the sampling points. In Central Macedonia, the winter flights coincided with moist soils, growing crops and permanent orchard canopies, while the measured SOC values varied only within a narrow range. Under these circumstances, no reliable SOC model could be built. The same workflow did, however, give promising results for other soil properties such as electrical conductivity, calcium carbonate and pH (R² between 0.61 and 0.69), which underlines how much SOC mapping depends on good bare-soil conditions at the time of flying.
Impact and future outlook
Both studies demonstrate that better soil carbon maps are not just a matter of choosing the right algorithm. Federated learning proves that organisations can pool their strengths to build more robust models while keeping sensitive field data under their own control, an important step towards trust and compliance with data protection rules. The drone work shows that aligning the image area with the way soil is actually sampled, backed by careful calibration and fair validation, can substantially sharpen SOC predictions at field level.
By uniting privacy-preserving collaboration with state-of-the-art sensing, the Soil R&I Lab is laying the groundwork for more detailed, reliable and scalable soil monitoring in support of sustainable land management across Europe.
|
