Reduction of the Dimensionality of Hyperspectral Data for the Classification of Agricultural Scenes

Recent advances in sensor technology opened new possibilities for remote sensing. For example, the appearance of sensor higher spatial and spectral resolution. In terms of spectral resolution, the number of available bands increased significantly, resulting in hyperspectral sensors. Hyperspectral remote sensing images are characterized by the division of the electromagnetic spectrum in a great number of narrow spectral bands, which enable greater detail of spectral variation of targets. High dimensionality demands special attention in the classification process. The main problem caused by the increase of the dimensionality is the reduction of the efficiency of the classifiers. This problem is known as the Hughes phenomenon. The occurrence of the Hughes phenomenon is caused by the exaggerated increase of the dimensions of the variance covariance matrix (increase of the dimensionality), compared to the limited number of available training samples. As a result, recent approaches focus reduction of the dimensionality. In this paper, a method of feature selection from hyperspectral images is presented. The proposed method, based in the use of the Genetic Algorithms, is evaluated with a AVIRIS data set and the results are compared to the results of other algorithms (Sequential Forward Selection and Sequential Backward Selection), recognized as techniques for reduction of dimensionality. A Genetic Algorithm can be described as a global search technique for optimization purposes inspired by the natural evolutionary process. The experiments show that Genetic Algorithms based reduction method can be used to reduce the dimensionality within image classification in remote sensing.