A novel maximum likelihood based method for mapping depth and water quality from hyperspectral remote-sensing data

This article presents a novel statistical method for mapping water column properties from hyperspectral remote-sensing data. Usual inversion methods are based on a pixel-by-pixel approach. Therefore, they do not consider the spatial correlation between neighboring pixels, even though such pixels are often affected by the same water column if the spatial resolution is high enough. oposed method uses such redundant information performing local maximum likelihood (ML) estimation of depth and water quality in large zones. It provides multi-resolution maps, in which resolution depends on local depth. In shallow water, the signal-to-noise ratio is high so an accurate mapping can be performed. Therefore, we propose to model local depth variations with a linear model while water quality is still supposed to be locally homogeneous. In deep water, the signal-to-noise ratio is lower so estimating only the local mean depth with standard ML estimation is more robust and reliable. The entire image is divided into appropriate meshes. In every mesh, water column properties are estimated using both linear and constant depth models. Final maps are obtained combining these estimates. The deeper the water, the higher the influence of standard ML estimation maps. Using local information provided by neighboring pixels makes this method robust to noise. Moreover, the hyperspectral image and estimated bathymetry have the same resolution in shallow water since depth is modeled for every pixel. ficiency of our method was assessed with simulated and real hyperspectral images. Results proved that depth modeling improves depth and water quality estimations, especially in shallow water. In deep water, assuming that the bottom is locally flat is reasonable since depth variations are small relatively to depth mean value. With the considered water quality, the estimated bathymetry was accurate for depths up to 14 m. Estimated concentration maps were consistent for the whole range of depths. The spatial resolution of estimated bathymetry was 50 cm in shallow water (for depths up to 10 m), and ranged from 2.5 m to 5 m in deep water (for depths between 10 m and 14 m). fluence of bio-optical modeling is also demonstrated. We show that reliable models are absolutely necessary to obtain good estimation results.