Use of surface soil moisture to estimate profile water storage by polynomial regression and artificial neural networks

Abstract

Water storage in the soil profile is an important agronomic variable but its measuring is rather difficult for farmers in production fields. We tested the possibility of using samples from the upper soil layers, which are usually taken for soil fertility evaluation, for whole profile water storage estimation. A data set of 712 water profiles from the subhumid-semiarid portion of the Pampas in Argentina was used, generated under a wide range of soil types, crops, tillage systems, soil cover, and rainfall scenarios. To calculate stored water, soil was sampled up to 140 cm in layers of 20 cm, water content was gravimetrically determined and bulk density also assessed. Polynomial regression and artificial neural networks were used for modeling, randomly partitioning the data set into 75% for model fit and 25% for independent testing. It was possible to estimate with good fit soil profile water storage using as independent variables in regression, or inputs in neural networks, water content in the upper three soil layers (0-20, 20-40, and 40-60 cm) and depth to petrocalcic layer in soils which have this type of horizon. Similar performance was attained with both modeling methods (R2> 0.93, RMSE = 11% of mean water content). Other soil and environmental properties had only a minor impact on estimations and were dropped from models. Because of its simplicity, regression is the recommend method for estimation of water content in the soil profile for agronomist.