Integrating Remote Sensing and AI modelling in Mediterranean Agroforestry and Croplands systems: A Methodological Perspective for spatial SOC monitoring in the MRV4SOC project, Spain

This study presents a robust framework for spatially explicit monitoring of soil properties and Above Ground Biomass (AGB) estimation in Mediterranean agroforestry and cropland systems by integrating remote sensing (RS) and artificial intelligence (AI). These variables are critical for assimilation into process-based models for Soil Organic Carbon (SOC) dynamics monitoring within a Monitoring, Reporting, and Verification (MRV) system. The framework was developed as part of the MRV4SOC project in Spain, aimed at designing a comprehensive, robust, and cost-effective Tier-3 approach. The primary goal is to produce high-quality geospatial layers of topsoil properties and AGB estima tion, which serve as key inputs for SOC dynamics modeling.

The methodology was tested at two long-term demonstration sites in Spain: Quercus ilex Dehesas in Extremadura (SW Spain) and rainfed cereal crops at La Canaleja experimental farm in central Spain. These agroecosystems provide diverse testing grounds for scalable and transferable SOC assessment methodologies within an MRV framework. The approach integrates multi-temporal remote sensing data (2018–2022) from Sentinel-2 and Landsat satellites with machine learning models to predict essential soil properties (SOC, Sand, Silt, Clay, pH, and Total N) and AGB. Ground truth data for AGB estimation were sourced from the Spanish National Forest Inventory (SNFI), while soil property predictions utilized the LUCAS 2018 topsoil libraries due to limited site-specific datasets for model training. A bare soil reflectance composite (2018–2022) derived from Sentinel-2 bands (B02–B12) at 20-meter resolution was employed for geospatial soil property mapping.

Given the limited availability of ground truth data, simpler models like Quantile Regression Forests (QRF) and XGBoost were selected. QRF achieved better accuracy for soil texture properties, with R² = 0.62 for clay and outperforming XGBoost for SOC (R² = 0.63) and pH (R² = 0.76) in the agroforestry site. However, XGBoost performed better for SOC (R² = 0.54) and total nitrogen in croplands, as well as for sand, silt, clay, and total nitrogen in the agroforestry site (R² = 0.61 for clay). For AGB estimation in the Dehesas area, a machine learning approach was implemented using SNFI data and remote sensing-derived transformation features. A gradient boosting algorithm (LightGBM) resulted in an R² value of 0.8. In La Canaleja, a bare soil reflectance composite was similarly employed for soil property mapping. Further analysis will be carried out to develop a bottom-up approach for monitoring SOC using these products and process-based models

Uncertainty analysis using Prediction Interval Ratio (PIR) assessment was conducted separately for landscape (L) and sub-landscape (SL) levels. While most properties showed medium to low uncertainty, sand and silt exhibited higher variability in croplands, and SOC displayed the highest uncertainty in the agroforestry site across L and SL levels.

This methodology contributes significantly to improving MRV systems by delivering high-quality geospatial layers for SOC dynamics monitoring in complex environments. Increasing ground truth data availability is essential for enhancing model accuracy and minimizing prediction uncertainties further.