Modelling Underwater Hyperspectral Images
Abstract
The underwater hyperspectral imager is a new type of imager able to acquire hyperspectral images under water. Common consumer cameras have poor spectralresolution and only captures light in three bands, in the red, green and blue partsof the light spectrum. Hyperspectral cameras, on the other hand, might capture hun-dreds of spectral bands in both the ultraviolet, visible and infrared regime. Thisproperty makes such cameras ideal for automatic mapping and object classification, which is important tools in fields such as environmental mapping.
In this work, we have taken the first steps towards a physical model for imagesacquired with an underwater hyperspectral imager. Light is heavily attenuated inwater, and corrections need to be made if measurements from different distancesand water bodies are to be compared. The proposed model consists of three sub-models: a model describing a lamp, a model for spectrally dependent attenuationin water and a model for reflectance. The lamp was modelled as a point sourcepositioned behind the real lamp, emitting light in a Gaussian angular distribution.Attenuation in water was modelled using radiative transfer theory, and reflectanceusing the Oran-Nayar reflectance model. With the proposed model we were ableto describe the collimation of the light well, and the assumed distribution as a reasonable first approximation. The model was also partly able to describe the changeof shape of the light spectra with attenuation.
We have tested classification in underwater hyperspectral images to see the ef-fect of uneven illumination and attenuation in water on the classification accuracy.It was also used as a tool for evaluating the proposed model. Uneven illumination and attenuation had a significant effect on classification accuracy, and henceneed to be accounted for in real world applications. Two approaches were testedfor removing these effects: normalisation and estimating reflectances. Normalisation gave a near perfect classification, while estimating reflectances using theproposed model performed worse than classification on the raw spectra. The es-timated reflectances had however similar shapes, and it is thus mainly a problemwith estimating the correct magnitudes for the illumination. The preprocessingmethod which seems to have the most potential when classifying different waterbodies is to first estimate reflectance, and then normalise. We then circumvent theproblem with illumination, and are able to classify between water bodies as longas the attenuation is known.
The proposed model performs within the expectations of a first attempt at modelling images acquired with an underwater hyperspectral imager. The developedsimulation- and analysis tools will be useful in further research and field measure-ments.