Plant classification for weed detection using hyperspectral imaging with wavelet analysis

Abstract

The goal of this study is to develop a new weed detection method that can be applied for automatic mechanical weed control. For successful weed detection, plants must be classified into crops and weeds according to their species. In this study, we employed a portable hyperspectral imaging system. The hyperspectral camera can capture landscape images that include crops, weeds, and the soil surface, and can provide more extensive information than conventional red, green, and blue (RGB) images. Although RGB images consist of red, green, and blue wavebands, the obtained hyperspectral images consist of 240 wavebands of spectral information. Hyperspectral imaging is expected to provide powerful technology for agricultural sensing. In the initial step of this study, the image pixels of the plants (crop or weeds) were segmented from the background soil surface using Euclidean distance as the discriminant function. In the next step, the image pixels of the crop (sugarbeet) and weeds (four species) were classified using the difference in the spectral characteristics of the plant species. In this process, classification variables were generated using wavelet transformation for data compression, noise reduction, and feature extraction, and then stepwise linear discriminant analysis was applied. The validation results indicate that the developed classification method has potential for practical use.