Quantification of chlorophyll-a in inland waters by remote sensing algorithm based on modified equivalent spectra of Sentinel-2

Abstract

The development of remote sensing algorithms has traditionally relied on satellite spectra or simulated equivalents derived from in-situ spectra to monitor inland water quality. However, such equivalent spectra often result in significant errors when retrieving chlorophyll-a (Chl-a) concentrations due to discrepancies between in-situ and satellite-derived spectra. In this research, the authors innovatively adjusted the red-light component of in-situ spectra for application in two inland waters, Dongzhang Reservoir and Jie Zhukou Reservoir. Sentinel-2 multispectral images (MSI), standard equivalent spectra (ES), and modified equivalent spectra (MES) were utilized as input data to assess models' effectiveness in terms of accuracy, robustness, and generalizability. The research applied Chl-a retrieval models including deep neural networks (DNN), extreme gradient boosting (XGB), and conventional statistical approaches with various spectral indices, such as the red-NIR method, the three-band method, and the normalized difference chlorophyll index (NDCI). The results revealed that the MES-based model achieved best results in Chl-a retrieval (RMSE = 2.04 mg/m³) comparable to MSI-based model (RMSE = 2.07 mg/m³) and ES-based model (RMSE = 7.71 mg/m³). Moreover, MES-based model behaved robustness and precision within selected water bodies and temporal periods. Notably, the integration of the red-NIR method with DNN was particularly effective in retrieving Chl-a with higher accuracy, robustness, and generalizability. Enhancement method to the equivalent spectra methodology provided by the research have reduced retrieval errors in retrieving Chl-a, and providing a valuable reference for future model development in this domain.