Reference evapotranspiration of Brazil modeled with machine learning techniques and remote sensing

Area of Science:

• Hydrology and Remote Sensing
• Agricultural Meteorology
• Environmental Science

Background:

• Reference evapotranspiration (ETo) is crucial for hydrology and irrigation, typically estimated using the Penman-Monteith method.
• Limited weather station data in developing regions necessitates alternative ETo estimation methods.
• Remote sensing products, like MOD16's potential evapotranspiration (PET), offer a viable solution for ETo estimation.

Purpose of the Study:

• To estimate monthly reference evapotranspiration (ETo) using the MOD16 potential evapotranspiration (PET) product.
• To evaluate the effectiveness of various machine learning algorithms for ETo estimation.
• To identify the optimal machine learning model for accurate ETo prediction using MOD16 data.

Main Methods:

• Collected data from 265 Brazilian weather stations (2000-2014) for Penman-Monteith ETo as the standard.
• Acquired MOD16 PET data for Brazil and WorldClim covariates.
• Trained and tested eight machine learning regression algorithms (including Cubist, Random Forest) to model ETo from MOD16 PET.

Main Results:

• MOD16 PET values were generally higher than Penman-Monteith ETo across Brazil.
• The MOD16 product demonstrated a good correlation with ETo, confirming its utility for estimation.
• All machine learning models improved ETo prediction accuracy; the Cubist model yielded the best performance (R²=0.91, NSE=0.90, nRMSE=8.54%).

Conclusions:

• Machine learning, particularly the Cubist algorithm, effectively enhances ETo estimation using the MOD16 product.
• The MOD16 product is a reliable source for predicting monthly ETo, especially in data-scarce regions.
• This approach facilitates broader applications of ETo data in hydrological and agricultural management.