A hyperspectral image contains hundreds of narrow and contiguous spectral bands. Because of this high spectral resolution, hyperspectral images provide valuable information from the earth surface materials and objects. Therefore, target detection (TD) is a key issue in processing such data. In fact, the aim of TD algorithms is to find specific targets with known spectral signatures. In the other point of view, the enormous amount of information provided by hyperspectral images increases the computational burden as well as the correlation among spectral bands. Besides, even the best TD algorithms exhibit a large number of false alarms due to spectral similarity between the target and background especially at subpixel level in which the size of target of interest is smaller than the ground pixel size of the image. Thus, dimensionality reduction is often conducted as one of the most important steps before target detection to both maximize the detection performance and minimize the computational burden.

This paper presents a method to improve the efficiency of subpixel TD based on selection of appropriate bands using genetic algorithm (GA). To use GA for band selection, two similar fitness function are proposed in this study. The first fitness function is introduced for cases in which the position of target is known. Regarding this, maximizing the output values of TD algorithm on the target pixels is used as the evaluation function. This maximization is roughly equivalent to minimizing the false alarm rate. The main problem in the use of the first fitness function is its need to know the correct position of target pixels in the image. Hence, the second function is proposed to solve this problem. In this function, the output value of TD algorithm is maximized on the simulated targets.

In this study, the adaptive coherence estimator (ACE) as the well-known subpixel TD algorithm is used in its local form for the evaluations. Moreover, the target detection blind test data set is employed for the experiments. The data sets includes HyMap reflectance image of Cook City in Montana, USA. The ground resolution of imagery data is approximately 3 m. In the HyMap image, 12 targets, at full and subpixel sizes, were located in an open grass region, which included six fabric panels for the self-test and six for the blind-test.

In this study, the local ACE algorithm is implemented using inner and outer detection windows with sizes of 3×3 and 5×5 pixels, respectively. Also, GA is performed with the population number of 100, the probability of mutation of 0.2, the probability of crossover of 0.8, and the maximum generation number of 100. Experimental results obtained for detecting the 10 subpixel targets considered show that, the number of false alarms produced when using dimension reduction by GA is completely low in comparison to that of obtained using all bands. Based on the results, the use of GA with first and second fitness functions reduce the false alarm rate by 95% and 75%, respectively, in comparison to using all bands. For fair comparison with the proposed method, the GA-contrast method is also performed on the same data set. The results show that, compared to the GA-contrast method, GA with the first and second fitness functions reduce the false alarm rate by 94% and 70%, respectively.