A coupled LSTM–XGBoost framework for daily estimation of Net ecosystem productivity using remote-sensing and environmental data

Net ecosystem productivity (NEP) of terrestrial ecosystems is a key indicator for measuring carbon sink capacity, and its high-precision simulation is of great significance for achieving the ‘dual carbon’ strategic goals. In view of the limitations of traditional methods in characterizing the temporal nonlinearity of NEP, this paper proposes a prediction framework that integrates deep temporal modelling with ensemble learning. First, a long short-term memory network (LSTM) is employed to extract key time-dependent features from multi-source driving variables; these features are then introduced as auxiliary variables into an XGBoost model to improve prediction accuracy and stability. Based on daily observations from nine eddy-covariance flux sites in China (2003–2010), we constructed a multi-source input feature set and compared the proposed framework with mainstream models such as Random Forest (RF), Support Vector Machine (SVM) and ensemble learning. The results show that extracting key temporal features with LSTM and feeding them into XGBoost significantly enhances the accuracy and stability of carbon-flux prediction: the mean R2 increases by approximately 2%, the error is markedly reduced, and the model’s temporal sensitivity and ecological adaptability are improved. In addition, Shapley Additive Explanations (SHAP) were employed to interpret model predictions and quantify the contributions of key environmental drivers to daily NEP dynamics. This study integrates deep time-series features with ensemble-learning methods for NEP estimation, achieving a dual improvement in predictive accuracy and ecological interpretability and providing a useful technical tool for regional carbon-flux monitoring and carbon-neutrality planning.