Aerial Hyperspectral Image Classification for Weed Map Development

A crop field weed map can provide good guidance for successful site-specific herbicide inputs that can reduce the negative environmental impact of agricultural chemical applications. Manually mapping the weed infestation condition (field scouting) is very time consuming, requires extensive labor input, and is subject to human errors. Remote sensing provides an alternative way for fast and cost-effective in-field variability mapping. When NASA commercializes its hyperspectral remote sensing technologies, hyperspectral images will be available on a regular time basis for agricultural applications. This study explored the potential use of hyperspectral images for agricultural field weed mapping. Images were processed and classified using a spectra-feature-fitting method. Classification was based on the extracted endmembers from the calibrated image. In order to verify the results, the classified image was compared with ground truth data collected from a ground truth machine - selective patchy sprayer. The results showed that hyperspectral image classification could successfully identify in-field high weed infestation areas. The maximum match ratio of 0.82 is observed at 0.9 weed density level, indicating good correlation between the classified image and high weed infestation areas.