An accurate estimation of spatially and temporally continuous latent heat flux (LE) is essential in the assessment of surface water and energy balance. Various satellite-derived LE products have been generated to enhance the simulation of terrestrial LE, yet each individual LE product shows large discrepancies and uncertainties. Our study used Extremely Randomized Trees (ETR) to fuse five satellite-derived terrestrial LE products to reduce uncertainties from the individual products and improve terrestrial LE estimations over Europe. The validation results demonstrated that the estimation using the ETR fusion method increased the R² of five individual LE products (ranging from 0.53 to 0.61) to 0.97 and decreased the RMSE (ranging from 26.37 to 33.17 W/m²) to 5.85 W/m². Compared with three other machine learning fusion models, Gradient Boosting Regression Tree (GBRT), Random Forest (RF), and Gaussian Process Regression (GPR), ETR exhibited the best performance in terms of both training and validation accuracy. We also applied the ETR fusion method to implement the mapping of average annual terrestrial LE over Europe at a resolution of 0.05 ◦ in the period from 2002 to 2005. When compared with global LE products such as the Global Land Surface Satellite (GLASS) and the Moderate Resolution Imaging Spectroradiometer (MODIS), the fusion LE using ETR exhibited a relatively small gap, which confirmed that it is reasonable and reliable for the estimation of the terrestrial LE over Europe.

Ocean latent heat flux (LHF) is an essential variable for air–sea interactions, which establishes the link between energy balance, water and carbon cycle. The low-latitude ocean is the main heat source of the global ocean and has a great influence on global climate change and energy transmission. Thus, an accuracy estimation of high-resolution ocean LHF over low-latitude area is vital to the understanding of energy and water cycle, and it remains a challenge. To reduce the uncertainties of individual LHF products over low-latitude areas, four machine learning (ML) methods (Artificial Neutral Network (ANN), Random forest (RF), Bayesian Ridge regression and Random Sample Consensus (RANSAC) regression) were applied to estimate low-latitude monthly ocean LHF by using two satellite products (JOFURO-3 and GSSTF-3) and two reanalysis products (MERRA-2 and ERA-I). We validated the estimated ocean LHF using 115 widely distributed buoy sites from three buoy site arrays (TAO, PIRATA and RAMA). The validation results demonstrate that the performance of LHF estimations derived from the ML methods (including ANN, RF, BR and RANSAC) were significantly better than individual LHF products, indicated by R² increasing by 3.7–46.4%. Among them, the LHF estimation using the ANN method increased the R² of the four-individual ocean LHF products (ranging from 0.56 to 0.79) to 0.88 and decreased the RMSE (ranging from 19.1 to 37.5) to 11 W m⁻². Compared to three other ML methods (RF, BR and RANSAC), ANN method exhibited the best performance according to the validation results. The results of relative uncertainty analysis using the triangle cornered hat (TCH) method show that the ensemble LHF product using ML methods has lower relative uncertainty than individual LHF product in most area. The ANN was employed to implement the mapping of annual average ocean LHF over low-latitude at a spatial resolution of 0.25° during 2003–2007. The ocean LHF fusion products estimated from ANN methods were 10–30 W m⁻² lower than those of the four original ocean products (MERRA-2, JOFURO-3, ERA-I and GSSTF-3) and were more similar to observations.

Accurate estimation of satellite-derived ocean latent heat flux (LHF) at high spatial resolution remains a major challenge. Here, we estimate monthly ocean LHF at 4 km spatial resolution over 5 years using bulk algorithm COARE 3.0, driven by satellite data and meteorological variables from reanalysis. We validated the estimated ocean LHF by multiyear observations and by comparison with seven ocean LHF products. Validation results from monthly observations at 96 widely distributed buoy sites from three buoy site arrays (TAO, PIRATA, and RAMA) indicated a bias of less than 7 W/m² with R² of more than 0.80 (p<0.01) and with a King–Gupta efficiency (KGE) of over 0.84. Our estimated ocean LHF also performs well in simulating annual variability and predicting between-site variability, as indicated by a bias of lower than 6 W/m² and an R² of more than 0.84 (p<0.01). Overall, the average KGE for estimated ocean LHF increased by 18%–23% compared to other LHF products, indicating robust LHF estimation performance. Importantly, our estimated annual ocean LHF has similar global spatial distribution compared to other LHF products, although there are general differences in LHF values due to the difference in the models and the spatial resolution.

As a major agricultural hazard, drought frequently occurs due to a reduction in precipitation resulting in a continuously propagating soil moisture (SM) deficit. Assessment of the high spatial-resolution SM-derived drought index is crucial for monitoring agricultural drought. In this study, we generated a downscaled random forest SM dataset (RF-SM) and calculated the soil water deficit index (RF-SM-SWDI) at 30 m for agricultural drought monitoring. The results showed that the RF-SM dataset exhibited better consistency with in situ SM observations in the detection of extremes than did the SM products, including SMAP, SMOS, NCA-LDAS, and ESA CCI, for different land cover types in the U.S. and yielded a satisfactory performance, with the lowest root mean square error (RMSE, below 0.055 m³/m³) and the highest coefficient of determination (R², above 0.8) for most observation networks, based on the number of sites. A vegetation health index (VHI), derived from a Landsat 8 optical remote sensing dataset, was also generated for comparison. The results illustrated that the RF-SM-SWDI and VHI exhibited high correlations (R ≥ 0.5) at approximately 70% of the stations. Furthermore, we mapped spatiotemporal drought monitoring indices in California. The RF-SM-SWDI provided drought conditions with more detailed spatial information than did the short-term drought blend (STDB) released by the U.S. Drought Monitor, which demonstrated the expected response of seasonal drought trends, while differences from the VHI were observed mainly in forest areas. Therefore, downscaled SM and SWDI, with a spatial resolution of 30 m, are promising for monitoring agricultural field drought within different contexts, and additional reliable factors could be incorporated to better guide agricultural management practices.