Initial Experiments with a Scalable Machine Learning Based Approach for Downscaling the MOD16A2 Evapotranspiration Product

Vatsal Jingar, Stitiprajna Sahoo, Siddharth S, Dharmisha Sharma, Shivani A. Mehta, Aaditeshwar Seth

Abstract: In countries like India which have historically been reliant on rainfed agriculture, the increasing need of water for irrigation to support greater cropping intensity and shifts towards horticulture, has largely been supported through groundwater based irrigation. Cheap electricity has enabled a rapid increase in borewells almost across the country, which to some extent has enabled more equitable access to water than other irrigation approaches like canals, but has also led to groundwater stress in many regions. One way to indirectly estimate groundwater abstraction is to estimate evapotranspiration from cropping areas as a proxy for crop water consumption. Remote sensing based methods have been used to estimate evapotranspiration but existing open data products largely have a low spatial resolution which is not adequate to support local decision making for water use. In this study, we build machine learning methods to develop downscaled data outputs of evapotranspiration at fine spatial scales. Our approach uses satellite data, meteorological variables, and land surface characteristics as input features to obtain field-scale fortnightly time-series of evapotranspiration. We validate the results results across multiple geographic locations and also study its correlation with in situ evapotranspiration measurements. We find that our method is not able to accurately match in situ data but is able to successfully provide relative differences in evapotranspiration. We make our trained models available on the Google Earth Engine platform for use by other researchers and practitioners to obtain evapotranspiration outputs for their areas of interest. Our research contributes a scalable and adaptable solution to address the growing demand for fine-resolution hydro-climatic information.