[Prediction of Nitrogen Concentration in Agricultural Drainage Water by Fusing Multispectral Images from Unmanned...

Rapid and accurate monitoring of nitrogen concentration in farmland drainage ditch water can help to analyze the migration pattern of agricultural surface source pollution and enhance watershed water quality management and pollution prevention and control. In order to solve the problems of large nitrogen migration flux and difficult spatio-temporal dynamic monitoring of farmland drainage ditches in Ningxia Yellow River Irrigation District， taking the fourth drainage ditch of farmland in Ningxia Yellow River Irrigation District as the research object， and based on the dynamic water quality monitoring data and UAV multispectral images of the ploughing period in 2024， we chose six machine learning models including BP， SVM， CNN， RF， XGBoost， and CatBoost and conducted an analysis modeling and predicting total nitrogen （TN）， ammonium nitrogen （NH₄⁺-N）， and nitrate nitrogen （NO₃^--N）. On this basis， the RF， XGBoost， and CatBoost models with better performance were selected for stacked model development， and the distribution of the prediction results was plotted. The results of the study showed that： ① The average TN content of each monitoring point in the fourth drainage ditch was higher than the standard limit of Class V water （2 mg·L^-1） during the 2024 cropping period， and the NO₃^--N accounted for 26.16%-84.57% of TN， which was the main form of nitrogen pollution in the water body of the fourth drainage ditch. ② The normalized summation value （k） was able to combine the advantages of the Pearson correlation analysis and the random forest feature importance analysis and fully resolved the complex dependence between spectral features and nitrogen concentration， which is more reliable as the basis for feature screening. ③ Using the XGBoost model for TN， NH₄⁺-N， and NO₃^--N， the R² of the test set was 0.81， 0.55， and 0.79； the MAE was 0.67， 0.12， and 0.68 mg·L^-1； and the RMSE was 0.90， 0.16， and 0.90 mg·L^-1， respectively， which were the best machine learning models in this study. ④ The RF-XGBoost models for TN， NH₄⁺-N， and NO₃^--N had test set R² of 0.89， 0.66， and 0.86； MAE of 0.56， 0.10， and 0.61 mg·L^-1； RMSE of 0.68， 0.14， and 0.73 mg·L^-1； MSE 0.46， 0.02， and 0.53 mg·L^-1； and CCC of 0.94， 0.74， and 0.92， respectively. The model had high prediction accuracy， small error， high consistency of prediction results， and strong extrapolation ability， and it was the best stacking model at the field scale in the present study. ⑤ The RF-XGBoost model tended to under-predict the extreme values， over-predict the extreme minima， and predict the intermediate values more accurately， and the consistency between the predicted and measured plots of TN and NO₃^--N was stronger， while the prediction results of NH₄⁺-N were poorer. In summary， the RF-XGBoost stacking model can achieve accurate prediction of TN and NO₃^--N in the fourth drainage ditch of farmland in Ningxia Yellow River Irrigation District by relying on spectral features only， and it has strong extrapolation ability， which can be used in the follow-up promotion. However， it is difficult to fully reflect the changing law of NH₄⁺-N by relying on spectral features only， so it is necessary to explore the modeling feasibility of other correlation indexes in the follow-up study. Therefore， it is necessary to explore the feasibility of modeling other related indexes in the subsequent research， and this result can provide new ideas and solutions for the research in this direction.