Spatiotemporal soil moisture estimation for agricultural drought risk management

In this study we proposed a new spatiotemporal soil moisture estimation model for nowcasting and forecasting of the agricultural drought. Temporal rain, irrigation coverage, normalized difference vegetation index (NDVI) and parcel based evapotranspiration data are used as inputs of a multiple-input time-delay neural network for spatiotemporal multi-depth soil moisture estimation. For this purpose, we combined the outcome of another study that provides conversion of multi-temporal satellite image originated NDVI data to spatiotemporal NDVI data. Agricultural drought and water stress risk management are complex processes since they are not only dependent on soil structure and climatic parameters but also they have relevance to plant cover status. Our soil moisture estimation model provides nowcasting ability for water-stress management. It also provides drought forecast information in order to utilize water resources in an optimal way for yield loss reduction. The main capability of this method depends on the applicability of both past time-window data and weather forecast information as the future data to the trained neural network. Agro-meteorological data from randomly placed 6 stations, chosen to be within different agricultural fields, are used in this study. Soil moisture estimation model was performed through a 4-layer time-delay neural network. Past and predicted meteorological parameters are used as a part of inputs to the neural network. We have shown that estimated soil moisture has approximately 3% root mean square error at 15cm and 45cm depths of reference verification points.