Estimating reference evapotranspiration using hybrid models optimized by bio-inspired algorithms combined with key meteorological factors

As a fundamental parameter for assessing farmland hydrological environment and formulating irrigation plans, the accurate estimation of reference evapotranspiration (ETo) is essential for efficient agricultural water resource management. To improve the precision of ETo estimation under limited conditions, this study used Gradient boosting decision tree (GBDT), Kendall' tau correlation coefficient (Kendall), Maximum information coefficient (MIC), and Partial correlation analysis (PCA) to analyze the influence of meteorological factors on ETo estimation and constructed different combinations based on the importance ranking of variables. A novel hybrid model (DBO-RF) was established by combining the Dung beetle optimizer (DBO) with the Random forest (RF) model, and the daily meteorological data from 30 stations in the North China Plain (NCP) from 1968 to 2019 were used to compare the DBO-RF with the RF model optimized based on the Whale optimization algorithm (WOA) and the Particle swarm optimization (PSO), as well as the original RF, Adaptive boosting (AdaBoost), Decision Tree, Knearest neighbor (KNN), and Multivariate adaptive regression splines (MARS) models. The results showed that by comparing the estimation accuracy of the ETo models under all combinations, the input combinations extracted by GBDT significantly improved the model's estimation accuracy. Among the five important combinations, the DBO-RF model demonstrated more stable and superior performance in ETo estimation across different regions of the NCP. Furthermore, even under external assessment, the DBO-RF model based on the M4 combination still exhibited better adaptability (R2 > 0.95). In conclusion, the study provides reliable technical support for efficient agricultural water resource management in the NCP.