Spatio-temporal mapping of soil water storage in a semi-arid landscape of northern Ghana – A multi-tasked ensemble machine-learning approach

Soil water storage (SWS) illustrates the available water capacity of soil horizons and its water reservoir from which crops can draw upon during transient water deficit periods. Information on SWS quantity and stability at spatial and temporal scales represented in a 4D predictive map is important to support sustainable agricultural intensification in sub-Sahara Africa, especially as ∼80% cash constrained and risk averse smallholder farmers are involved in rainfed agriculture. Yet, such information is non-existent in all the predictive soil mapping initiatives. We developed the first modelling framework to map 4D SWS information at 100 m spatial and 12-days temporal resolutions in Ghana. Our model is based on an ensemble machine-learning framework implemented with random forest for spatial data and extreme gradient boosting algorithms, which produces a 4D SWS map at six soil depth intervals following the GlobalSoilMap specifications. Model input datasets were optimally subsetted from an a priori set of covariates that drive SWS variability following the ‘Scorpan’ concept of soil mapping. The modelling framework was validated with a time-series in situ SWS data (n ≈ 2000; in a single season in 2018) measured at six depths until 100 cm interval using a calibrated moisture probe from key benchmark soils of the Guinea savannah landscapes of northern Ghana. The model was optimized and evaluated via a 5-fold leave-location-time-out cross validation algorithm. Our model recorded an RMSE accuracy of 0.22 mm (0.84 ≤ CCC ≤ 0.86), which is consistent with models with similar objectives. Across the different soil types and for all 0–100 cm depth intervals, predicted SWS contents were comparable to those of the in situ measurements with overall means 33.5 mm and 34.2 mm respectively, suggesting high model accuracy. Our results moreover indicated that SWS in top soil layers (0–15 cm) were highly variable, unstable, and consistently dry indicating high temporal instability, whereas SWS in bottom soil layers (>15 cm) showed time-stable wet spatial cluster of locations due to clay content build-up. The main application of the output in this study is that SWS time-stable locations where crop water requirements can be met can be explicitly identified for improved crop production, especially at the onset of rains. Thus, questions of how much, how long, where, and when soil water is adequate for cultivating, for example in short maturity crops or deep-rooted crops, can now be answered using our 4D maps for the Guinea savannah region of Ghana. Our outcomes form a core support system necessary to guide the implementation of drought-adaptation measures and complement existing predictive soil mapping initiatives.