This paper investigates adequate selection of hyperspectral components for estimating the degree of soybean bacterial pustule infection. According to a filed report, the higher the disease damage level is, the less the yield is expected.Based on highresolution hyperspectral images for different levels of damaged soybean fields, we have investigated the relationship between the soybean weight averaged for 100 seeds and the hyperspectral changes.In order to estimate the degree of damages from the hyperspctral data, we have tested 3 methods: (i) Estimation using reflectance of a single band based on a linear regression analysis.The RMSE was 2.0g. (ii) Estimation using normalized vegetation indices also based on a linear regression analysis.The RMSE was 2.1g. (iii) Estimation using neural network (NN) of a single-layer perceptron which has input nodes corresponding to the hyperspectral bands and a single output node.The RMSE was 1.5g, for estimating 100 seeds weight ranged over 16.7g to 22.7g.This paper indicates that NN is the most accurate among three methods to estimate the degree of soybean bacterial pustule infection.
