DEVELOPMENT OF DISCRIMINANT MODEL FOR WEED DETECTION USING HYPERSPECTRAL IMAGING
Physical weed control is environmentally safer than chemical methods. However, most of the physical methods are manual. Therefore, development of a robotic weeding system is necessary to practice efficient weed control. The objective of this study was to develop a pixel discriminant model for image segmentation between crop and weed using hyperspectral imaging. A hyperspectral image was processed pixel by pixel. First, every pixel spectrum was extracted from the image. Next, each pixel spectrum was classified into soil and plant. If the pixel spectrum was identified as plant, it was classified into crop and weed. In the pixel discriminant model for soil and plant, simple NDVI thresholding was employed. The pixel discriminant model for crop and weed consisted of a normalizer, an explanatory variable generator and the discriminator. Development of the explanatory variable generator was tried with two different methods (RAW or PCA). Development of discriminator was also tried with two different methods (LDA or NN). In this study, four types of models (RAW-LDA, RAW-NN, PCA-LDA and PCA-NN) were developed for discrimination between crop and weed and validated. Finally, segmented images for crop, weed, and soil were generated from the images by applying these models. Pixel discrimination between soil and plant was performed with high accuracy. In the pixel discrimination between crop and weed, success rates of all discriminant models were more than 85%. As for the accuracy of the models, NN models were superior to LDA, and the PCA method was superior to RAW. However, with respect to processing speed, the RAW method was superior to PCA. As a result of image segmentation, most of pixel spectra in the images were identified correctly. In conclusion, this study showed the possibility for weed detection by using hyperspectral imaging.
Suzuki, Y., Okamoto, H. and Kataoka, T. (2009). DEVELOPMENT OF DISCRIMINANT MODEL FOR WEED DETECTION USING HYPERSPECTRAL IMAGING. Acta Hortic. 824, 67-74
machine vision, image processing, plant classification, remote sensing, soybean