Modeling evapotranspiration from remote sensing data using machine learning algorithm

Abstract

An accurate evapotranspiration (ET) estimation is crucial for hydrological, meteorological, and agricultural research and applications, such as weather forecasting, irrigation scheduling, drought monitoring, and energy budgeting. ET can be directly measured using lysimeters, but this procedure is expensive. Various ET estimation models / algorithms are also available to estimate ET locally to Globally. However, they require explicit characterization of numerous physical variables (e.g., precipitation, soil moisture, soil in-filtration capacity, soil texture, etc), which are sometimes difficult to obtain in data-scarce regions. The advent of satellite technology has inspired researchers worldwide to use remote sensing (RS) data to calculate ET from them. However, many existing ET estimation techniques using RS data still depend on ancillary ground observations. This study has employed machine learning algorithms like Support Vector Machine based regression (SVR) and Random Forest (RF) regression algorithms to estimate actual evapotranspiration for agricultural areas based on remote sensing data only.

This study first explored the relationships between ET and a number of satellite data-derived radiation and ecosystem variables, such as the Enhanced vegetation index (EVI), Global Vegetation Moisture Index (GVMI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Vegetation Index (NDVI), Land Surface Water Index (LSWI), albedo, land surface temperature (LST), etc. and found significant correlations between them. RS variables were obtained / derived from the moderate resolution imaging spectroradiometer (MODIS) images, and the observed ET was obtained from four eddy covariance towers located in the agricultural sites of Bangladesh, the United States, and South Korea. Different combinations of these RS variables have been used as input in the machine learning (ML) based model. Actual evapotranspiration is also computed by the well-known Sur-face Energy Balance Algorithm (SEBAL), which uses both satellite and weather data as input. The widely used Penman-Monteith (PM) equation was also used to compute ET in this study. The performances of all models were assessed against the observed data.

xii

The ML models showed promising potential in modeling ET from RS data. A number of SVR models were developed using single / multiple input variables in different combinations. The coefficient of determination (R2) and root mean squared error (RMSE) values for the best SVR-based model were 0.88 and 0.6 mm/day, respectively. But the general SVR model (based on the data of all sites belonging to different climatic zone and growing different crops) failed to capture the high ET values; therefore, individual station-wise models were then developed and tested to check if they could predict high ET values. But the performances of the station-wise models were poorer. Later, another ML algorithm, called random forest (RF) was employed to overcome this issue. The R2 and RMSE of RF-based models were 0.94 and 0.44 mm/day, respectively. Although the RF model has performed better than the SVR model, this model also missed some very high peaks. ML models are highly dependent on the quality and quantity of the data. There were only 2 to 3% data with very high values. If there were more data with high values, then these models would get many instances of high ET values to learn their fea-tures better and could have predicted them better. This can be one of the causes behind the failure to predict very high peaks of ET during the summer season. This study also compared the performance of the ML models with that of SEBAL and PM. The SEBAL model showed poor performance than the ML models. For the SEBAL model, R2 varied from 0.63 to 0.84, and RMSE varied from 1.19 to 1.64 mm/day for different stations. On the other hand, R2 and RMSE for the Penman-Monteith method varied from 0.21 to 0.53 and 0.92 to 2.01 mm/day, respectively. It is certain that among all the methods, machine learning-based models have performed better than the other existing methods. There-fore, this study demonstrates a promising possibility of estimating ET from solely remote sensing data in data-scarce regions like Bangladesh and may contribute to efficient water resources management. Further study is needed to validate the models developed in this study using data from newer stations and upcoming years when they become available.