March 3, 2026Open Access

Interpretable Temperature‐Based Deep Learning for Evapotranspiration: SHAP ‐Based Feature Analysis in CNN ‐ GPU

Puntos clave

Evapotranspiration was accurately estimated using temperature-based deep learning models, achieving a high Nash-Sutcliffe efficiency of 0.98.
The CNN-GPU model displayed superior performance, reducing training time by 20.4% compared to the CNN-RNN model while maintaining low RMSE of 0.23 mm/day.
Shapley Additive Explanations analysis highlighted significant influences of solar radiation and wind speed on evapotranspiration predictions despite limited input variables.
This modeling approach may improve irrigation scheduling and water management strategies under climate change conditions, but further validation is needed.

Resumen

ABSTRACT Reference evapotranspiration ( ET o ) is a critical parameter for assessing crop water requirement and formulating irrigation scheduling and water management practices under climate change conditions and water shortage. Classical approaches e.g., the FAO ‐Penman‐Monteith ( FPM ‐56) equation generally require several meteorological data inputs, which are often unavailable or limited. In the present study, CNN ‐ RNN and GPU ‐accelerated CNN ( CNN ‐ GPU ) models were utilized for temperature‐dependent ET o estimating. ‘ SHapley Additive exPlanations ’ ( SHAP ) analysis revealed that solar radiation and wind speed exerted high degrees of influence, even after their exclusion from the input matrix, which clarified these implicit nonlinear relationships captured by the model. CNN ‐ GPU model outperformed CNN ‐ RNN in both accuracy ( RMSE = 0.23 mm/day, NS = 0.98) and computational efficiency with a faster training time by 20.4%. Despite training with limited input variables (temperature records), the proposed DL ‐based models successfully captured complex temporal and spatial meteorological patterns in the study region.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo