An Evolutionary-Strategy-Based CHC Genetic Algorithm and Its Application to Rock Spectrum Discrimination

Data mining from field rock spectrums is important for hyper-spectral remote sensing modeling. The characteristics of the field spectrum data are used to reform and combine evolutionary strategies with CHC (cross generation elitist selection, heterogeneous recombination and cataclysm mutation) and to design a genetic algorithm to conduct rock spectrum discrimination: to build a linear discriminating equation with many variables (wavelength intervals). Floating encoding is used for the coefficients of the equation (genes). Two optimization objectives are compared with each other: one is to maximize the right-judgment ratio of known samples, and the other is to minimize the ratio of in-class versus between-class distances. The results show that the former is briefer and faster while both of them are effective. Monte Carlo sampling is employed to adjust searching spaces. Uniform and Gaussian distribution models are employed to produce new-generation genes, showing that the former model has better performance, because the genes have various unknown distributions. An experimental study is presented based on the data of 1 823 wavelength intervals from Beiya gold deposit, Yunnan, China, obtained with the FieldSpectr Fr equipment (ASD Co. US). The algorithm proves high efficiency in identification of altered dolomitic limestone, a potentially gold-bearing rock, from other rocks.