Drone based weed monitoring with an image feature classifier

Abstract

Site specific weed management needs detailed weed information down to the species level. Then herbicides can be used more specifically according to weed occurrence and their spatial distribution. The accurate identification of weeds is one of the major prerequisites to generate weed maps. Next to predominant implementations of online monitoring approaches on agricultural machinery, unmanned aerial vehicles (UAV) platforms will be used in future to generate weed maps of different species by using high-resolution imagery. While colour-based indices are already applied for mapping nutritional deficits or water deficiency, they have failed to identify different weed species. In contrast, object-based image analysis looks much more promising to separate plant characteristics by means of form and morphology yet are much more complex.
This study proposes a new computer vision approach to discriminate weed species based on a bag-of-visualword (BoVW) framework using high resolution aerial images. BoVW is an object-based image classifier that has recently gained interest in agricultural research. In our trials this technology has been applied in laboratory tests and field trials for automatic weed sampling with digital cameras.
The results showed that the BoVW model allows the discrimination between Matricaria recutita L., Triticum aestivum L., Papaver rhoeas L. and soil with good accuracy. For providing consistent weed maps in terms of precise herbicide applications in the future, the robustness of the classifier must be evaluated with more crops and weed species acknowledging the natural plant variability observed in the fields.