191 / 2024-02-27 11:28:50

Machine learning based total nitrogen retrieval in small inland waterbodies

total nitrogen; machine learning; unmanned aerial vehicles; spatiotemporal distribution

Inland small waterbodies are important water sources for human, and are also the cornerstone of maintaining ecological environment stability. It is of great significance to carry out accurate, fast, and green efficient monitoring of inland water quality. However, there are almost no ground-based water quality monitoring points in small inland waterbodies, and unmanned aerial vehicles (UAV) low-altitude remote sensing provides opportunities and challenges for water quality retrieval in these regions. This study proposed a machine learning water quality retrieval framework that integrated spatial information, feature engineering and intelligent optimization algorithms to retrieve total nitrogen in inland small waterbodies. The results showed that the RMSE of the training, validation, and test sets was 0.036-0.21 mg/L, R² was 0.918-0.998, RPD was 3.483-20.052, indicating that the established framework has high accuracy. Furthermore, the established framework was applied to multispectral images obtained from 6 UAV flights, and it was found that the predicted quartiles and mean values of TN concentration in each study area were close to the measured, with an error of less than 20%. Therefore, it is considered that the established model framework has high transferability. Although there are still many uncertainties in the water quality inversion model constructed in this study, such as significant differences in the retrieval results of various models in part of the water bodies, the overall spatial distribution is consistent. This study has achieved a transformation of water environment monitoring from point to surface, which can provide reference for intelligent management of small waterbodies.