Method for object-oriented feature selection based on genetic algorithm with multi spectral images

In the crop identification process using remote sensing, selecting the robust feature variable is a key to the correct identification. After comprehensive analysis of spectrum characteristics of the multi-spectral data, aiming at the goal ground objects, this paper proposed a new model for object-oriented feature selection based on Genetic Algorithm(GA) with multi-spectral images. In the model, after comparison and analysis were made in terms of the value of OIF (Optimum Index Factor) of band combination, the best feature number of them was determined. Then considering the principle of Maximizing-Minimum Distance as the basic theory, the formula of weighed JM Distance was obtained through improving Jeffries-Matusita (J-M) Distance. It can be used as the criterion for measuring the efficiency of feature to classification. The results were optimized by GA, in which fitting function was constructed with this distance, and the band combination was selected that was sensitive to classification. In this paper the study area was selected in Dehui county of selected that was sensitive to classification. In this paper the study area was selected in Dehui county of Jilin Province, Northeast China. The band selection was conducted based on the model proposed using Landsat-5 images. Better results were obtained by the model.