Freshwater marsh classification in the Lower Parana River floodplain: an object-based approach on multitemporal X-band COSMO-SkyMed data

A multitemporal approach to discriminate freshwater macrophyte vegetation types in the Lower Parana River floodplain is addressed. During a low-intensity flood pulse, seven X-band HH COSMO-SkyMed Hlmage images are acquired, covering a nine-month period. Scenes are segmented with a mean-shift segmentation algorithm. Objects are classified with an expectation maximization algorithm into clusters with different temporal signatures and are assigned to six information classes: water, bulrush marshes, short broad-leaf marshes, tall broad-leaf marshes, short grasslands and grass marshes, and tall grasslands and grass marshes. Class interpretation is based on backscatter dynamics, with focus on their correlation with hydrometric water level measured in the Parana River and/or with the floodplain area covered by water as estimated with a normalized difference vegetation index threshold criterion. The obtained product has a global accuracy of 75.4% and a kappa index of 67.2%. We point out the usefulness of X-band for flood monitoring and macrophyte vegetation type discrimination. However, we find limitations for the discrimination between high-biomass vegetation targets, such as tall broad-leaf marshes and tall grasslands. In a mosaic of herbaceous wetlands, the knowledge on the relation between vegetation and floods is essential for interpreting and predicting how back-scattering coefficients and other synthetic aperture radar-derivated parameters vary with flooding. (C) 2019 Society of Photo-Optical Instrumentation Engineers (SPIE)