Evaluating machine learning and deep learning for upscaling instantaneous latent heat flux to daily scale

Abstract The aim of this study is the evaluation of machine learning (ML) and deep learning (DL) models as an alternative to traditional methods for upscaling hourly latent heat flux (LE i ) to daily latent heat flux (LE d ). For these purposes, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Long Short-Term Memory-Convolutional Neural Network (LSTM-CNN) models were developed using the FLUXNET2015 dataset and compared the performance of these models with the evaporative fraction (EF) approach. Also, the most suitable combinations of time and features were determined to upscale LE i to LE d . Results indicated that the model performance was stable during midday hours, with the highest correlation at 11:00 (R²=0.75, RMSE = 24.25 W m⁻², and MAE = 17.58 W m − 2 ). To determine the best features, Pearson Correlation analysis was conducted, and results revealed that hourly LE i , net radiation (Rn i ), reference evapotranspiration (ETo i ), incoming solar radiation (SW in ), air temperature (Ta i ) and daily net radiation (Rn d ) had the greatest influence on LE d . Among the generated models, MLP performed best (R²=0.85, RMSE = 15.96 W m⁻² and MAE = 11.43 W m⁻²) to estimate LE d . However, in terms of computational efficiency, XGBoost outperformed other models with significantly lower training times, whereas DL models required intensive computational effort. This study showed that ML and DL models are robust and more accurate for upscaling LE compared to traditional methods. Specifically, MLP is prominent as a suitable model for accuracy.