Table 4_Using non-vegetated soil mosaics for soil property modelling to compare Google Earth Engine and THEIA platforms across Sminja Plain, Tunisia.pdf

This study examines the potential of Sentinel-2 (S2) temporal mosaics (TM) of non-vegetated soils to enhance soil property mapping in the semi-arid Sminja Plain, Tunisia (480 km²). Using multi-season data from 2019 to 2023, TMs were generated with two processing platforms, Google Earth Engine (GEE) and THEIA, and their predictive performances were compared. Non-vegetated soils were isolated using thresholds of NDVI < 0.35 and NBR2 < 0.09 to optimise non-vegetated soil extraction. Key soil properties, including electrical conductivity (EC), pH, soil organic carbon (SOC), base saturation (BS), exchangeable bases (K, Ca, and Na), granulometric fractions, and soil moisture contents (at field capacity and permanent wilting point), were analysed from 215 georeferenced samples systematically distributed across the study area. Random Forest (RF) models were calibrated using K-fold cross-validation, and their predictive performances were evaluated through RMSE, RPD, and RPIQ metrics. Results indicate that both platforms effectively predicted most of the inherent soil properties (SOC, CaCO₃, Ca, BS, granulometric fractions, and soil moisture content) with RPIQ values over 1.7. Conversely, predictions for dynamic soil properties (pH, EC, K, Na, and P₂O₅) were less or non-reliable, with RPIQ below 0.8. Seasonal mosaics had a minor effect on model performance, while platform choice showed only slight differences, with GEE performing marginally better in some cases. This highlights the limitations of the TM and RF models for certain dynamic soil properties in such environments. These findings emphasise the importance of multi-season non-vegetated soil mosaics for monitoring semi-arid soils and provide insights to support sustainable land management and agricultural planning.