This work presents a combination of machine learning approaches to classify and assess feature relevance in remotely sensed hyperspectral and multispectral data.
The approach starts with the automatic determination of the number of classes, which are the Optical Water Types (OWTs). OWTs refer to various conditions of oceanic waters. Then unsupervised classification is used; hence no labels are required. Feature relevance algorithms are applied to the OWTs, which allow understanding differences in the input features associated with the OWTs. Feature relevance algorithms aim to detect important variables in the input space which explain the output parameter.
The method was studied in both training and predictive mode. Using the approach in predictive mode allows determining the OWTs and associated relevant features for new unlabeled data, which shows the possibility of operation processing and assessment of water types.
The methodology was evaluated and presented on multispectral Arctic in-situ and satellite data and on a hyperspectral synthetized data. Using the approach can reveal how water quality changes in time and space in the Arctic oceans.
The paper “Learning Relevant Features of Optical Water Types” by K. Blix, A. B. Ruescas, J. E. Johnson and G. Camps-Valls (IEEE Geoscience and Remote Sensing Letters) can be accessed here.