Water resources optimization and irrigation advisory based on farm-level remote sensing and crop water demand

Abstract:

Agricultural water management decisions are usually made based on in-situ measurements (e.g. precipitation, soil moisture). Such approaches are costly, hardly scalable, and do not consider available weather forecasts. Recent advances in high-resolution remote sensing and modeling can offer farm-level information with reasonable confidence, and with potential for better-informed water resources management. The question underlying this project is: can a farm manager be advised how much water is needed to be supplied to the crops for the following week to optimize the end of season crop yield? For this part of the study, we propose a framework to monitor and forecast (with a 3-7-day lead-time) subfield crop water stress. To do this, we use a fine-resolution (70 m) dataset from six ~190 ha study farms across the provinces of Saskatchewan and Manitoba in Canada. At each site, we have assembled data such as precipitation (MSWEP), topography (SRTM), surface soil moisture (SMAP SPL3SMP), root zone soil moisture (SPL4MAU), LAI (Landsat), ET/ET0 (LSM or ECOSTRESS), and crop yield data (surveyed). We downscaled the value of each covariate in every 70 m grid location for the study period. By using the crop’s phenological development schedule, we link the relevant depth of soil moisture with crop water demand and soil water supply. To separate the spatial and temporal components of crop water stress and to reduce the dimensionality of the data we used the Empirical Orthogonal Function (EOF) approach. For the prediction of crop water stress, we use techniques such as Multilinear regression and Support Vector Machine. Our prototype is developed and tested on farms in the Canadian Prairies, but the framework can be transferred to other parts of the world.