Improved mapping of dynamic soil properties by comparing date-specific vs. temporal mosaicking strategies on bare agricultural soils using Sentinel-2

Abstract

Mapping dynamic soil properties (DSPs), such as pH and nutrient levels, that fluctuate under management practices (fertiliser application, liming, irrigation, etc.), seasonal cycles, and environmental factors, is essential for precision agriculture, yet reliably quantifying them from satellite imagery remains a challenge. In this study, conducted in a semi-arid agricultural region of Tunisia covering 480 km², we challenge the standard composite-first paradigm by systematically evaluating the relationship between specific satellite acquisition dates and the predictability of DSPs. Using a dense time series of Sentinel-2 imagery (2019-2023) and 215 soil sampling points, we modelled key DSPs (pH, K, P₂O₅, EC, and Na) and soil moisture suctions (pF2.8 and pF4.2) using both temporal mosaic and date-specific approaches, with the latter being applied only on dates with at least 50 valid samples after cloud and vegetation filtering. Our results reveal a crucial limitation of the temporal mosaic approach, which yielded poor predictive performance, validated by low RPIQ values, for all properties. In contrast, date-specific analysis showed that certain DSPs, notably pH, K, and P₂O₅, could be predicted with high accuracy on specific optimal dates. At the same time, EC and Na remained poorly predicted, likely due to a low proportion of saline points in the dataset. We conclude that in such semi-arid agricultural environment, the temporal context of image acquisition is a decisive factor for successfully mapping specific DSPs, mandating a strategic shift from universal composites toward date-specific modelling for operational soil mapping.