Optimizing methane flux prediction and key feature identification based on a novel hybrid machine learning model

Natural landscape methane (CH₄) emissions account for half of the global total, yet their quantification remains challenging due to high measurement costs and ecosystem complexity. In this study, the performance of three machine learning models, XGBoost (XGB), random forest (RF) and support vector machine (SVM) were assessed, to predict CH₄ fluxes, using data from 36 FLUXNET-CH₄ sites. Furthermore, the GWOPSO algorithm was applied to hybridize the XGB model to optimize input features and hyperparameters simultaneously. The results demonstrated that ensemble models (XGB and RF) significantly outperformed SVM. The hybrid model further improved prediction accuracy. Across most wetland types, it required only five key features to surpass the all-feature XGB model, reducing RMSE by 1.1%-10.8%. Soil temperature was identified as the most important predictive variable across most site types. Overall, this study recommends a reliable CH₄ prediction method, which provides a reference for CH₄ prediction of different site types.