Genetic algorithm based new sequence of principal component regression (GA-NSPCR) for feature selection and yield prediction using hyperspectral remote sensing data

Recently, hyperspectral images are used to estimate the yield of food crops. The images consist of a large number of bands which requires sophisticated method for its analysis. One approach to reduce computational cost and to accelerate knowledge discovery is by eliminating bands that do not add value to the analysis. In this paper, a genetic algorithm based new sequence of principal component regression (GA-NSPCR) method is proposed and tested using 116 band HyMap airborne hyperspectral data and yield data collected from paddy fields. The proposed method uses GA to select an initial subset of hyperspectral bands, and subsequently generate a more accurate subset by measuring the minimum error of prediction model defined by principal component regression (PCR). Unlike standard PCR methods which order the features based on singular values, in each generation NSPCR orders the features based on squared multiple correlation coefficient R². Yield data and spectral data are used to generate a separate training and testing dataset using 8 times bootstrap resampling (8-rounds BSR) to deal with limited number of samples in training data. Differed from standard GA impelementation, the fitness function evaluates three L_p-norms to obtain the best prediction model.