Abstract
The main goal of remote sensing image classification is to associate land cover classes to each pixel in the monitored area. In this sense, hyperspectral images play a key role by providing detailed spectral information per pixel. On the other hand, although the huge amount of spectral bands enables the creation of more accurate thematic maps, they can compromise the quality of results due to data redundancy, high-dimensionality problems and noisy bands. Many dimensionality reduction techniques have been proposed in order to better use the available information. An effective strategy is to perform a band selection, which aims at selecting the best bands for classification. This process decreases the dimensionality without degrading information, i.e., it keeps the physical properties acquired by the sensors. As a drawback, the dimensionality reduction process can take a lot of time to be performed. In this paper, we propose a new unsupervised band selection method based on the dissimilarity among neighboring bands by exploiting an intermediary representation called spectral rhythm. Our approach can take advantage of a pixel sampling strategy to improve its efficiency without significant reduction on the quality of selected bands. Experimental results reveal that our method can efficiently select suitable bands to represent the whole data by producing accuracy results as good as the baselines in the classification problem.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have