Data scale over model complexity: evaluating multi-architecture neural networks for daily reference evapotranspiration across the CONUS-AgMet dataset

Chatzithomas, Constantinos; Malamos, Nikolaos

doi:10.22098/mmws.2026.20570.1875

Data scale over model complexity: evaluating multi-architecture neural networks for daily reference evapotranspiration across the CONUS-AgMet dataset

مقالات آماده انتشار

نوع مقاله : پژوهشی

نویسندگان

Academic Staff, Department of Natural Resources and Agricultural Engineering, School of Environment and Agricultural Engineering, Agricultural University of Athens; Athens, Attica

10.22098/mmws.2026.20570.1875

چکیده

In this study, we developed 96 neural network (NN) models, representing all combinations of 8 input‑variable sets, 4 network architectures, and 3 activation functions. All models estimated daily ETo using a subset of the CONUS-AgWeather dataset, which was split into training (40%) and testing (60%) periods. The split was chronological, and the test period exhibited higher average Tx, Tn, and Tav values, reflecting an underlying warming trend. From the full model ensemble, the eight best‑performing models, one for each unique input‑variable combination, were selected based on RMSE. Each selected model produced ETo estimates for the entire spatial domain of the dataset. The top model, which used all available input variables, achieved an RMSE of 0.19 mm/d. All high‑performing models included Rs as an input, with RMSE values ranging from 0.60 to 1.01 mm/d. The model using only Tav yielded an RMSE of 1.30 mm/d; and a MBE of -0.03 mm/d. Given the large spatial extent and climatic diversity of the dataset, this performance is considered satisfactory. Feature importance analysis revealed Rs as the most important variable except for models where Tav is combined with RH and u2, which reflect the aerodynamic part of ETo. Overall, the analysis showed that neither the choice of activation function nor the network architecture had a substantial impact on model performance. Instead, the quantity and quality of the dataset were key factors enabling the strong performance of the NN models as tested with the CONUS-AgWeather dataset and the specific architectures and activation functions.

کلیدواژه‌ها