A novel comprehensive interpolation to produce spatiotemporally continuous daily evapotranspiration

Evapotranspiration (ET) is a significant component of the global hydrologic cycle, playing a crucial role in transferring water and the Earth's energy from the surface to the atmosphere through latent heat. Due to its importance in the hydrologic cycle, the estimation and mapping of ET using satellite remote sensing data have been the subject of numerous research in agriculture and applied hydrology. However, a single satellite ET retrieval still has a couple of limitations, including a trade-off between spatial and temporal resolutions and missing information due to cloud contamination (e.g., cloud cover and its shadows) and defective satellite sensors (e.g., the Landsat 7 scan line corrector-off). As a solution to the current issues of a single satellite ET product, the present study developed a season and land cover specific interpolation (SLSI) technique using decision tree-based machine learning models and various remote sensing products. Using the SLSI, this study produced spatiotemporally continuous daily ET data (ETSLSI) on a statewide scale using the existing Landsat-derived ET product (ETLS) that are discontinuous in time and space. Based on a comparative analysis between three decision tree-based models, including Classification and Regression Trees (CART), random forest (RF), and eXtreme Gradient Boosting (XGB), the SLSI adopted RF for predicting ET over the missing areas due to its robust performance with high accuracy. The developed SLSI was tested in Illinois. The resulting ETSLSI was compared with the two eddy covariance flux site measurements. The results validated the reliable performance of the SLSI, showing good agreement with observations and keeping the accuracy of the source data (i.e., ETLS). In addition, the results of the experimental evaluation, where the non-cloud contaminated ETLS scene was masked with artificially created clouds with four different cloud cover rates (20, 40, 60, 80%) and then reconstructed by the SLSI, proved that the SLSI could successfully reproduce the missing spatial information for all cloud rates regardless of cloud shape and cloud cover rates.