Machine Learning Classification, Feature Ranking and Regression for Water Quality Parameters Retrieval in Various Optical Water Types from Hyper-Spectral Observations

This work presents an approach how machine learning techniques can be utilized for upcoming hyper-spectral missions to improve water quality monitoring globally by using a combination of classification and regression methods, and at the same time understanding the spectral relevance for various Water Quality Parameters (WQPs) in different Optical Water Types (OWTs). Machine learning methods are studied to classify OWTs and assign relevance to spectral bands for the WQPs from hyper-spectral observations. The model takes hyper-spectral Level-2 Rrs to classify the data into OWTs by using a non-linear kernel Support Vector Machine (SVM). Then, for selected WQPs feature ranking is performed to assign relevance to the spectral bands in the OWTs. In this study, two WQPs were selected, Chlorophyll-a (Chl-a) content and absorption from Colored Dissolved Organic Matter (aCDOM). The results are presented in spectral relevance maps, showing how spectral relevance varies with increasing optical complexity. These spectral relevance maps are used to select the relevant hyper-spectral bands for a given WQP, which then can be used for regression. Two regression models are evaluated, the kernel Support Vector Regression and Neural Nets.