Genetic algorithms for Hyperspectral Range and Operator Selection

Abstract

A novel genetic algorithm was developed using mathematical operations on spectral ranges to explore spectral operator space and to discover useful mathematical range operations for relating spectral data to reference parameters. For each range, the starting wavelength and length of the range, and a mathematical range operation were selected with a genetic algorithm. Partial least squares (PLS) regression was used to develop models predicting reference variables from the range operations. Reflectance spectra from corn plant canopies were investigated, with proportion of plants (1) with visible tassels and (2) starting to shed pollen as reference data. PLS models developed using the spectral range operator framework had similar fitness than PLS models developed using the full spectrum. This range/operator framework enabled identification of those spectral ranges with most predictive capability and which mathematical operators were most effective in using that predictive capability. Detection of operator locality may have utility in sensor and algorithm design and in developing breeding stock for other algorithms.