Effective Spatial Resolution for Weed Detection

Patch spraying herbicide to control weeds has distinct advantages. Machine vision using digital images can be used for patch spraying however, research has found varied results. Image processing requires considerable computational time when image resolutions are high and poses difficulties for real time application. This study was undertaken to study trade offs between image resolution and detection accuracy. The images were acquired using a digital color camera and then individual plant images of 128x128 pixels were extracted using Matlab software. Shape, green color and radial spectral energy features were selected to classify crop weed images and three resolutions were taken. Excess green method was used for segmentation. A Bayesian classifier was used for three classifications while using only shape features, only green color and spectral energy features, and all features combined. Green color and spectral energy features performed best. The classification accuracy using these features at different resolutions for weed detection varied from 80% to 87%. Their performance was unaffected by image resolution and shows potential for field applications.