Abstract. Soil moisture is highly variable in space, and its deficits (i.e. droughts) plays an important role in modulating crop yields and its variability across landscapes. Limited hydroclimate and yield data, however, hampers drought impact monitoring and assessment at the farmer field-scale. This study demonstrates the potential of field-scale soil moisture simulations to advance high-resolution agricultural yield prediction and drought monitoring at the smallholder farm field-scale. We present a multi-scale modeling approach that combines HydroBlocks, a physically-based hyper-resolution Land Surface Model (LSM), and machine learning. We applied HydroBlocks to simulate root zone soil moisture and soil temperature in Zambia at 3-hourly 30-m resolution. These simulations along with remotely sensed vegetation indices, meteorological conditions, and data describing the physical properties of the landscape (topography, land cover, soil properties) were combined with district-level maize data to train a random forest model (RF) to predict maize yields at the district- and field-scale (250-m) levels. Our model predicted yields with a coefficient of variation (R2) of 0.61, Mean Absolute Error (MAE) of 349 kg ha−1, and mean normalized error of 22 %. We captured maize losses due to the 2015/2016 El Niño drought at similar levels to losses reported by the Food and Agriculture Organization (FAO). Our results revealed that soil moisture is the strongest and most reliable predictor of maize yield, driving its spatial and temporal variability. Consequently, soil moisture was also the most effective indicator of drought impacts in crops when compared with precipitation, soil and air temperatures, and remotely-sensed NDVI-based drought indices. By combining field-scale root zone soil moisture estimates with observed maize yield data, this research demonstrates how field-scale modeling can help bridge the spatial scale discontinuity gap between drought monitoring and agricultural impacts.
Multilayer Soil Moisture Mapping at a Regional Scale from Multisource Data via a Machine Learning Method
