Um novo método Wrapper multiobjetivo para seleção de bandas de Imagens Hiperspectrais

Abstract

Hyperspectral images (HIs) are characterized by higher spectral resolution than other kind of images, having applications in areas such as medicine, mining, and especially in agriculture. These images in conjunction with remote sensing have become a useful tool for precision farming, enabling identification and analysis of health conditions in agricultural areas. For this identification it is necessary to segment the images that can be obtained through classification. An intrinsic problem with HIs is the volume of data that can pose a challenge in terms of transmission, storage, processing and also the performance of classification algorithms (caused by the curse of dimensionality). Techniques that reduce dimensionality are promising for HIs, but many of them are designed to deal with a single objective and cannot assure a balance between conflicting objectives. Examples of conflicting objectives can be based on improving the pixel classification and reducing the number of HIs bands simultaneously, the latter being related to the dimensionality of these images. To try to deal with solutions for conflicting objectives, it can be applied multiobjective algorithms that are designed for this purpose. Band selection methods based on multiobjective algorithms have recently been proposed in the literature, but many strategies have not yet been explored or properly combined. Based on different approaches from the literature, in this research it was developed a multiobjective band selection method called Wrapper Multiobjective Evolutionary Band Selection (WMoEBS) composed of strategies that were experimentally tested. WMoEBS is based on the Wrapper strategy incorporating the Support Vector Machine (SVM) classifier, using spatial and spectral information as input, it makes an initial selection to narrow down correlated bands, being a multiobjective algorithm dealing with classification results and number of bands simultaneously and a decision maker to return a single final solution. WMoEBS has been compared with state-of-the-art methods in the classification improvement criteria for different metrics and bandwidth reduction capability. Experiments have shown that WMoEBS presents superior results to most cases tested for classification metrics, including when applying statistical tests, and it is also advantageous in reducing the number of bands.