New methodology of hyperspectral information extraction and accuracy assessment based on a neural network

Abstract

Spectral analysis is an important research topic in the field of hyperspectral remote sensing. In this paper, the pixel vectors of hyperspectral imagery are taken as research object. A new method of spectral analysis and hyperspectral image information extraction were studied based on a Neural Network (NN). An application of this method in disaster measurement of wheat stripe rust disease is described. The experimental data was hyperspectral imagery of wheat-infected with stripe rust and came from the Push-broom hyperspectral imager (PHI). The NN model of Self-organizing Feature Maps (SOFM) was introduced to cluster and analysis of the disease severity. By means of computing the spectral index data (SID) and spectral angle data (SAD) as well as spectral average reflectance data (ARD) of hyperspectral imagery in experimental area, the three kinds of two-dimensional data matrixes were used as the inputs of the SOFM model. After iterative learning and self-organized clustering, the outputs of the model were mapped in 3-dimensional severity space of wheat stripe rust. The computed data and image data in the experimental area were simulated through net-training. The trees and soil in the experimental area are easily interpreted by comparison of sample’s spectral curves. The disease and healthy sampling points were identified by “Red-edge” positions. Then, the degrees of disease severity were established by spectral threshold analysis of samples in near-infrared bands. Finally, a cross correlogram spectral matching technique was used to assess the accuracy. The study results show that the disease of wheat stripe rust was severe. The whole trial spot was divided into six feature classes, including four severity degrees of wheat stripe rust disease, they are healthy, light, mid and heavy. These results are reliable and accurate.