Water Quality Inversion Framework for Taihu Lake Based on Multilayer Denoising Autoencoder and Ensemble Learning

In river and lake ecosystem management, comprehensive water quality monitoring is crucial. Traditional in situ water quality monitoring is costly, and it is challenging to cover entire water bodies. Remote sensing imagery offers the possibility of efficient monitoring of water quality over large areas. However, remote sensing data typically contain a large amount of noise and redundant information, making it difficult for models to capture the effective spectral information and the relationships in the water quality in the remote sensing data. Consequently, this hinders the achievement of high-precision water quality inversion performance. Therefore, this study proposes a comprehensive water quality inversion framework based on a multilayer denoising autoencoder that automatically extracts effective spectral features, utilizing a multilayer denoising autoencoder to extract effective features from Sentinel-2 remote sensing data, thereby reducing noise in the subsequent model input data and mitigating the overfitting problem in subsequent models. A bagging ensemble learning model was established to invert the total phosphorus concentration in Taihu Lake. This model reduces the prediction bias generated by a single machine learning model and was compared with decision tree, random forest, and linear regression models. The research results indicate that compared to a single model, the bagging ensemble learning model achieved better water quality retrieval results, with a coefficient of determination of 0.9 and an MAE of 0.014, while the linear regression model performed the worst, with a coefficient of determination of 0.42. Additionally, models trained using spectral effective information extracted by multilayer denoising autoencoders showed improved water quality retrieval accuracy compared to those trained with raw data, with the coefficient of determination for the bagging model increasing from 0.62 to 0.9. This study provides a rapid and accurate method for large-scale watershed water quality monitoring using remote sensing data, offering technical support for applying remote sensing data to watershed environmental management and water resource protection.