Towards a predictive and explainable mapping of soil organic carbon stocks with TabPFN and SHAP: A case study from agro-sylvo-pastoral ecosystems in Senegal

Source Agritrop Cirad (agritrop.cirad.fr) International audience Soil organic carbon (SOC) plays an essential role in the mitigation of climate change, agricultural productivity, and environmental sustainability. This study aims to map the SOC stocks in three village territories within the Senegalese groundnut basin located in Niakhar. This Basin, the agricultural heartland of Senegal, dominated by smallscale family farming and characterized by declining organic inputs to the soil and critically low stocks of SOC. In this study, we apply a novel approach that leverages contextual learning through the TabPFN network, a foundation model specifically designed for tabular data. The performance of this model is compared to traditional ML algorithms (Random Forest and XGBoost) in the prediction of SOC stocks using a combination of environmental covariates and in situ SOC measurements. Furthermore, model interpretability is assessed using Explainable AI (XAI) techniques based on SHAP values, which allow for interpreting the influence of each covariate on the variability of the predicted SOC stocks. This study use 1813 georeferenced ground samples collected across three terroirs, along with environmental covariates including topographic, climatic, and remote sensing variables extracted from the Google Earth Engine (GEE) platform to predict SOC stocks in the 0-30 cm soil layer. Our findings reveal that environmental covariates alone explain 52% of the variability in SOC stock using the TabPFN model, outperforming XGBoost (47%) and Random Forest (46%). To improve predictive accuracy, we compared the performance of these models using another dataset in which environmental covariates were supplemented with soil carbon content. The models achieved significantly higher predictive performance (R2 = 0.95, 0.96, 0.80) compared to those trained with environmental covariates alone (R2 = 0.52, 0.47, 0.46), suggesting that environmental variables alone do not fully capture the variability influencing the spatial distribution of SOC stocks. The SHAP analysis identified sentinel-2 spectral data, gross primary productivity (GPP), precipitation, and elevation as the most important predictors. This study highlights the potential of the TabPFN network for SOC stock mapping and the use of SHAP values to improve the interpretability of the model, allowing the identification of the most influential environmental covariates and their effects on the variability of SOC stocks.