Abstract: [Purposes]Total phosphorus (TP)concentration is a key indicator for measuring eutrophication and ecological health in water bodies.Traditional manual sampling methods are costly and have limited spatial coverage.Although satellite remote sensing inversion technology is an effective alternative,it faces challenges in model fitting and generalization ability when the sample size is insufficient.[Methods]To overcome the accuracy limitations under small sample conditions，this study proposes a Stacking integrated remote sensing inversion model for small sample data.By constructing a multidimensional spectral feature set to mine effective information,combined with recursive feature elimination (RFE)and turbidity mediation model to optimize input feature structure and physical consistency,heterogeneous base learners such as XGBoost,LightGBM,random forest,CatBoost,and partial least squares regression (PLSR)are integrated,and high-order fusion is performed by meta learners.[Findings]In the Sentinel-2 image experiment,the model achieved an inversion accuracy of R² better than 0.92.[Conclusions]This method significantly improves the robustness and generalization ability of the model under small sample conditions,providing a reliable method for remote sensing monitoring of total phosphorus in water bodies．

Key words: total phosphorus concentration, remote sensing inversion, Stacking ensemble learning, small sample learning, Sentinel-2