Data Fusion of Regional Reanalysis- and Sentinel (Earth Observation)-based Products with Machine Learning Tools for Monitoring Evapotranspiration and Drought

Accounting of the hydrologic process of evapotranspiration (ET) or consumptive use of water is important for water resources allocation, irrigation management, drought early warning, climate change impact assessment as well as in agro-water-climate nexus modeling. In fact, monitoring the United Nations' sustainable development goals (SDGs) that emphasize on improved food security, access to clean water, promotion of sustainable habitats and mitigation of natural disasters (droughts) hinge upon access to better quality data of ET. Though numerous studies have targeted accurate estimation of potential evapotranspiration (PET) using earth observation (EO) data; hydrologists are yet to reach consensus on the best set of predictor variables that can be used irrespective of spatio-temporal scale. This can be attributed to the nonlinear and complex nature of the process of ET. When it comes to the estimation of actual ET (AET), studies employing Eddy Covariance (EC) towers have been successful in different regions of the world. However, the developing countries of the world lack access to EC observations, requiring viable economical methods for accurate ET measurement, even using reliable estimates of PET. The proposed study explored fusion of regional climate reanalysis data, EO data, and machine learning techniques for high-resolution PET estimation. In this analysis, owing to the documented success of data-driven models in hydrological studies, performance of two machine learning models- tree based Random Forest (RF) and regressor Multivariate Adaptive Regression Splines (MARS), are evaluated for estimating monthly PET. A suite of input predictors are chosen to describe three model categories: meteorological-, EO- and hybrid-based predictor models. There are about 10 input combinations that can be generated for the PET model development, particularly for an agriculture-dominated study region - Dhenkanal district, located in Odisha in eastern part of India. In this study, reanalysis-based (meteorological) inputs at a grid resolution of 0.12° and Sentinel 2A (EO) products at spatial resolution of 20 m have been used. Results of the analysis indicate that solar radiation is the most important meteorological variable that controls PET estimation. Among the vegetation indices obtained from remote sensing data, we find that the Normalized Difference Water Index (NDWI) that represents availability of water in plants and soil, is particularly useful. The best PET estimation model that uses only solar radiation and few vegetation indices (NDVI, NDWI) gave coefficient of determination (R²) 0.88 and root mean square error (RMSE) of 0.14 during validation stage, whereas the use of hybrid predictor model that utilize temperature and vegetation indices information further reduced the error and increased the prediction accuracy (6.86%). When the meteorological inputs: precipitation and wind speed are only used, model did not perform well. Mapping the ET using the proposed models can facilitate reporting of progress in SDG with regard to water use, crop water stress, adaptation to agricultural droughts and food security. In this context, the Evaporative Demand Drought Index (EDDI) is computed across the study region to understand the drought patterns in the region.

Keywords: Potential Evapotranspiration, Agricultural Drought, Food Security, EO Data, Random Forest, Machine Learning, Vegetation Indices