Modeling transient soil moisture dichotomies in landscapes with intermixed land covers

Large-scale in situ soil moisture monitoring networks are becoming increasingly valuable research tools, but existing networks feature almost exclusive deployment of stations in grassland vegetation. These grassland soil moisture observations are unlikely to adequately represent the real soil moisture patterns in landscapes with intermixed land cover types. Here we demonstrate the severity of the problem for one particular landscape and introduce a flexible new method for solving the problem. The specific objectives of this study were (i) to compare root-zone soil moisture dynamics of two dominant vegetation types across Oklahoma, grassland (observed) and winter wheat cropland (simulated); (ii) to relate the soil moisture dynamics of grassland and cropland vegetation using an artificial neural network (ANN) as an observation operator; and (iii) to use the resulting ANN to estimate the soil moisture spatial patterns for a landscape of intermixed grassland and wheat cropland. Root-zone soil moisture was represented by plant available water (PAW) in the top 0.8 m of the soil profile. PAW under grassland was calculated from 18 years of soil moisture observations at 83 stations of the Oklahoma Mesonet, whereas PAW under winter wheat was simulated for the same 83 locations using a calibrated and validated soil water balance model. Then, we trained an ANN to reproduce the simulated PAW under winter wheat using only six inputs: day of the year, latitude and longitude, measured PAW under grassland, and percent sand and clay. The resulting ANN was used, along with grassland soil moisture observations, to estimate the detailed soil moisture pattern for a 9 × 9 km2 Soil Moisture Active Passive (SMAP) grid cell. The seasonal dynamics of root-zone PAW for grassland and winter wheat were strongly asynchronous, so grassland soil moisture observations rarely reflect cropland soil moisture conditions in the region. The simple ANN approach facilitated efficient and accurate prediction of the simulated PAW under winter wheat, RMSD = 21 mm and normalized RMSD = 0.17, using observed PAW under grassland as an input. This new method for estimating soil moisture under adjacent, contrasting land covers at a relatively low computational cost could be employed for any region and land cover pairing with training data available, and it may significantly enhance the applications of existing large-scale soil moisture monitoring networks.