Machine learning-based retrieval of chlorophyll-a and total suspended matter from HY-3A CZI: Model development, validation, and application

Résumé fourni par la source

Water quality monitoring via optical remote sensing provides essential insights for large-scale and long-term assessment and management of aquatic ecosystems. However, the inherent complexity of biogeochemical optical processes presents challenges in developing robust and generalizable retrieval algorithms. Machine learning (ML) has emerged as an effective alternative to handle the nonlinear relationships between apparent optical properties (AOPs) and water quality parameters (WQPs). Here, we introduced ML models to retrieve two key WQPs, chlorophyll-a (Chla) and total suspended matter (TSM), from the second-generation Coastal Zone Imager (CZI) onboard the HY-3A satellite. First, we compiled a global in-situ dataset (N = 7,535) comprising co-located remote sensing reflectance ( R r s ), Chla (range: 0.01–360.02 mg m −3 ) and TSM (range: 0.1–2626.82 g m −3 ) measurements, spanning diverse aquatic environments including oceans, coasts, estuaries, and inland waters. Based on this comprehensive dataset, we developed and evaluated five ML models, Random Forest (RF), Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN) and Mixture Density Network (MDN), to retrieve Chla and TSM from HY-3A CZI. Model testing showed that ML models outperformed the traditional retrieval algorithms, with MDN achieving the best performance, with Median Absolute Percentage Error (MAPE) of 33.64% for Chla and 30.79% for TSM. To further assess model applicability and product reliability, the trained models were applied to atmospherically corrected CZI images for two purposes: (1) independent validation using satellite and in-situ matchups, and (2) spatial pattern mapping and analysis across various aquatic environments. In the matchup validation, the quality assessments of input features revealed band-/sensor-dependent MAPE values ranging from 32.08% to 37.26% for individual bands, while those for band ratios were below 12%. By inputting quality-assessed data into the calibrated models, we obtained reliable Chla and TSM estimations with MAPE values ranging from 9.87% to 23.28%. However, the limited spatial and optical coverage of the matchup dataset warrants further investigation. Spatial pattern maps indicated that all models effectively captured realistic water quality gradients and heterogeneity, though performance varied among models. RF and XGBoost produced similar retrievals but struggled under extreme conditions. SVR was effective in TSM estimation but tended to overestimate Chla in the Pearl River Estuary. DNN and MDN exhibited superior robustness across diverse optical conditions and enabled the simultaneous retrieval of Chla and TSM. Through comprehensive model development, evaluation, and application, this study highlights the potential of ML-based approaches for accurate and scalable water quality monitoring with HY-3A CZI.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.