Determining noise in hyperspectral imagery for the application of oversampling to supervised classification

A hyperspectral sensor imager has many (∼200) contiguous spectral bands. It is this number and bandwidth of spectral bands that defines the spectral resolution of a sensor. In general, having a greater number of spectral bands has advantages, but this number is limited by implementation considerations. One of the most important is the limitation imposed on the sensor because of noise. In some cases, it is useful to apply signal-processing techniques to increase the resolution of the measured signal instead of or in addition to making physical changes to the sensor.; The resolution of the signal is defined by the noise level, and is measured in terms of signal to noise ratio (SNR). In practice, the number of binary digits that represent the amplitude of the signal is typically used.; Previous research has shown that spectral oversampling is typical in hyperspectral images of many physical objects and can be exploited with signal processing algorithms. These algorithms can be used to decrease the noise level of a signal, thus increasing the percentage of correct classification in supervised and unsupervised algorithms.; The research presented here investigates how oversampling techniques can be applied in a useful manner to hyperspectral images in supervised systems. An additive noise model is proposed for the noise signal, and used to design the oversampling algorithm. Also, a statistical characterization of the noise is obtained by modeling it as a stochastic process. Resolution enhancement algorithms are proposed in different classes: spectral filters, spectral-spatial filters and eigen-filters.; Results have shown an improvement in the correct classification accuracy percentage with the proposed spectral filter algorithm. Other techniques present diverse results.