Answers are grounded in this paper and cite their sources.
This study assessed the effect of training record length on daily pan evaporation estimation using Multiple Linear Regression (MLR), Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Extreme Gradient Boosting (XGBoost) models in a semi-arid region of Mexico. Two historical data scenarios comprising 10 and 20 years of daily observations were evaluated using different combinations of temperature, relative humidity, wind speed, and solar radiation variables. Both scenarios followed chronological training and validation periods and used the same independent testing period (2023–2024). The 10-year scenario used 2015–2020 for training and 2021–2022 for validation, whereas the 20-year scenario used 2005–2018 and 2019–2022, respectively. Among the selected configurations, MLR - M5 showed the most favorable overall performance for the 10-year scenario, with an RMSE of 1.98 mm/day, while LSTM - M9 was selected as the reference configuration for the 20-year scenario, achieving 1.97 mm/day. Extending the training record reduced LSTM RMSE by 0.12 mm/day (5.74%), whereas MLR RMSE increased by 0.05 mm/day (2.53%). Notably, LSTM using 20 years achieved an RMSE only 0.01 mm/day (0.51%) lower than MLR using 10 years. These results indicate that increasing the training record does not guarantee uniform improvements and that its effect depends on the learning paradigm and estimation configuration.
No takes yet. Share an insight, caveat, or question.
Castillo-Martínez et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: