Actual evapotranspiration by machine learning and remote sensing without the thermal spectrum

Area of Science:

• Remote Sensing
• Agricultural Science
• Machine Learning

Background:

• Accurate monitoring of actual evapotranspiration (ETr) is crucial for efficient irrigation management in agriculture.
• Sentinel-2 satellite data offers a valuable, high-frequency source for deriving vegetation parameters.
• Estimating ETr often relies on thermal bands, limiting temporal resolution.

Purpose of the Study:

• To develop and evaluate machine learning models for estimating evapotranspiration fraction (ETf) using Sentinel-2 spectral data.
• To assess the models' performance in estimating actual evapotranspiration (ETr) for center-pivot irrigated crops.
• To analyze the spatial and temporal applicability of these models for agricultural water management.

Main Methods:

• Two scenarios for ETf estimation using Sentinel-2 spectral bands, vegetation indices, and normalized ratio procedures (NRPB).
• Six regression algorithms were employed, with the Cubist model selected as the best performer.
• Actual evapotranspiration (ETr) was calculated using ETr-Brazil and Hargreaves-Samani methods, validated against the Simple Algorithm For Evapotranspiration Retrieving (SAFER).

Main Results:

• The Cubist model demonstrated superior performance in estimating ETf across both input scenarios.
• The ETr-Brazil approach provided more accurate ETr estimates compared to Hargreaves-Samani, which overestimated values for sugarcane.
• Spatially and temporally accurate ETr monitoring is achievable using Sentinel-2 data without thermal bands, enhancing temporal frequency.

Conclusions:

• Machine learning models, particularly Cubist, effectively estimate ETf using Sentinel-2 spectral data.
• ETr can be reliably monitored in agricultural areas without thermal infrared data, improving temporal resolution.
• The findings support enhanced irrigation management through frequent, accurate ETr assessments.