WEED DETECTION USING HYPERSPECTRAL IMAGING

Abstract

The goal of this study is to develop the discrimination method between crop and weed which require in the automatic mechanical weeding. In this study, the hyperspectral images were used. As data for analysis and verification, the hyperspectral images were acquired in the field of the university farm. These images consisted of crops, weeds and soil surface. First, the image pixels of the plant (crop or weed) were extracted from the background soil surface. In this process, the difference of spectral patterns between plant and soil was utilized. As a result of the test, the very high-precise segmentation was achieved. Next, the image pixels of crop (sugar beet) and weeds (four species) were classified by the analysis of the difference of spectral characteristics between plant species. In this process, the classification variables were generated using wavelet transform for data compression, noise reduction and feature extraction, and then the stepwise discriminant analysis was done. As a result of the test, the success rate in the plant classification was about 80%. Finally, the technique using the spatial neighbor information (area information) was devised in order to improve the performance of the plant classification.