Applying hyperspectral remote sensing methods to ship detection based on airborne and ground experiments

Abstract

Recent advances in hyperspectral remote sensing using hundreds of spectral channels make it possible to derive oceanic variables and to understand the spatial distribution of diverse oceanic phenomena. In this study, hyperspectral measurements were obtained using airborne and ground experiments and hyperspectral data processing techniques were applied for the detection of vessels. The observed hyperspectral data were analysed by using the spectral characteristics of endmembers and a representative spectral mixture analysis technique. The pixels in the image were classified into endmembers by using spectral matching algorithms and the marine library spectra from preliminary experiments. The estimated lengths and widths of the detected ships based on a series of hyperspectral data processing methods show root mean square error (RMSE) of 2.28% and 8.72% with respect to actual ship sizes, respectively. The investigation of the errors revealed that undulating surface waves, surface water penetration of the deck, and the submerged portion of the ship below the water surface potentially contributes to unclear ship boundaries. In this study, an optimal method for the objective determination of the threshold of abundance fractions based on pixels with multiple endmember contributions is proposed. The validation results based on the proposed hyperspectral method are in good agreement with the actual sizes of the ships.