Hyperspectral Image Feature Selection Based On Evolutionary Optimization

Hyperspectral feature selection, which is based on the reservation of the image’s physical information, is a search process to identify the most representative band combination. Evolutionary algorithm is a kind of global optimization search methods, which simulates the evolutionary mechanisms of nature in view of population. In this thesis, we apply evolutionary algorithms to hyperspectral band selection, the main contents include:(1) First of all, in accordance with the problems in the combination of hyperspectral feature selection and evolutionary algorithm, we design appropriate coding schemes and target functions. Hyperspectral feature selection, which based on the evolution optimization, is carried out on the principle of new target function. It is proven by experiments that better classification performance about band combination image can be obtained with evolutionary algorithm on the premise of the new target function.(2)Besides, on the basis of evolutionary algorithm, we propose hyperspectral feature selection based upon Memetic algorithm, which is incorporating local search operator. Memetic algorithm is researched by way of online optimization and offline optimization. It is shown by the experiment that the Memetic hyperspectral feature selection algorithm incorporating the local search operator increases the access to feature selection as well as classification.(3). At last, we apply multi-target evolutionary algorithm to hyperspectral feature selection with an effort to avoiding the randomness of some parameters in single-target evolutionary algorithm.