A framework for estimating actual evapotranspiration at weather stations without flux observations by combining data from MODIS and flux towers through a machine learning approach

Abstract

The scarcity of site-scale actual evapotranspiration (ET) measurements poses a challenge for modelling and verifying regional ET. Taking the grassland ecosystem in arid and semi-arid regions of Northern China (ASNC) as an example, we proposed a framework for estimating ET at weather stations without flux observations through a machine learning approach by combining data from MODIS and fifteen flux towers distributed in alpine grassland (AG) and temperate grassland (TG) across ASNC. First, we analyzed the temporal characteristics of grassland ET at site scale. Second, we applied the machine learning approach Random Forest (RF) to develop a robust model for simulating site-scale grassland ET, and used random and spatial cross-validations (CVs) and three metrics to evaluate the RF models performance. Two strategies (using pooled data from all flux towers for a general model, and dividing data by grassland type to develop models specific to AG and TG) and four temporal resolutions (daily, 8-day, monthly and seasonal) were used to develop the RF models. Third, we investigated how the importance of predictor variables for estimating grassland ET changed in different CVs, strategies and temporal resolutions. The ET of AG and TG showed similar dynamic patterns but differed in magnitude. The RF models showed good performance in both strategies and all four temporal resolutions (R2 > 0.64, MAE < 0.53 mm d−1, RMSE < 0.72 mm d−1). However, seasonal ET simulations performed better than that of daily, 8-day and monthly using pooled data from all flux towers, especially in spatial CV. Meteorological variables (temperature, precipitation and radiation), vegetation (NDVI, LAI and FPAR) and soil (soil water content and soil temperature) were strong predictors of variation in grassland ET. Changes in the importance ranking and value of predictor variables partly explained the variations of model performance. The robust model performance in spatial CV proved that the framework developed in this study was reliable when applied to weather stations without flux observations, thereby overcoming the scarcity of site-scale actual ET measurements.