Classification is an important and widely used technique for remotely sensed hyperspectral data interpretation. Although most techniques developed for hyperspectral image classification assume that the spectral signatures provided by an imaging spectrometer can be interpreted as a unique and continuous signal, in practice, this signal may be obtained after the combination of several individual responses obtained from different spectrometers. In this work, we propose a new spectral partitioning strategy prior to classification which takes into account the physical design of the imaging spectrometer system for partitioning the spectral bands collected by each spectrometer, and resampling them into different groups or partitions. The final classification result is obtained as a combination of the results obtained from each individual partition by means of a multiple classifier system (MCS).
The proposed strategy not only incorporates the design of the imaging spectrometer into the classification process but also circumvents problems such as the curse of dimensionality given by the unbalance between the high number of spectral bands and the generally limited number of training samples available for classification purposes. This concept is illustrated in this work using two different imagingspectrometers: the airborne visible infra-red imaging spectrometer, operated by NASA, and the digitalairborne imaging system (DAIS), operated by the German Aerospace Center. Our experiments indicate that the proposed spectral partitioning strategy can lead to classification improvements on the order of 5% overall accuracy when using state-of-the-art spatial–spectral classifiers with very limited training samples.