Improved Joint Estimation of Daily Actual Evapotranspiration and Gross Primary Production by Integrating Landsat, Sentinel-2, Sentinel-1, and Climate Data With Machine Learning

Accurate fine spatial resolution evapotranspiration (ET) and gross primary production (GPP) estimates will help us understand water-carbon interactions and optimize water resource management for enhancing ecological and agricultural applications. However, previous studies usually estimated ET or GPP separately at relatively coarse spatial or temporal resolution that is often insufficient for agricultural management. Besides, although Landsat provides fine-scale data, a single Landsat optical satellite can obtain limited observations. To address these issues, this study attempts to jointly estimate daily actual ET and GPP at 30-m resolution by integrating Landsat, Sentinel-2, Sentinel-1, and climate data with machine learning. Multisource optical and radar data were first integrated into unified Sentinel-2 vegetation indices (VIs) using linear and random forest (RF) models. Improved daily VIs were generated using climate data and temporally adjacent multisource satellite VIs with a bidirectional iteration approach. Finally, daily actual ET and GPP were jointly estimated using improved daily VIs and climate data. The results showed that the bidirectional iteration model improved daily VIs (R² > 0.882; RMSE < 0.077). Using the improved daily VIs, the total number of daily actual ET and GPP estimates were greatly improved (N = 23,669 vs. 3,657 and 8,771) compared to Sentinel-2 and integrated multisource satellite VIs. The RF model performed better than the other five evaluated machine learning algorithms for estimating ET (R² = 0.815; RMSE = 0.716 mm/d) and GPP (R² = 0.753; RMSE = 2.011 gC/m²/d). The study proposed a feasible methodological framework to jointly estimate daily actual ET and GPP at 30-m resolution, presenting great potential to monitor small-scale water stress and plant growth.