1 Application of Vector Agents to Weed detection from UAV imagery K. Borna* 1 , A. Moore 2 , B. Bollard 1 and A. Ghobakhlou 3 1 The School of Science, Auckland University of Technology 2 The School of Surveying, University of Otago 3 Department of Computer Science, Auckland University of Technology *Email: borna_kambiz@ymail.com Abstract In the remote sensing field, weed detection algorithms usually use the segmentation process to classify weeds in an image. In this context, the results are subject to user-defined parameters (e.g. scale) and predefined assumptions (e.g. uniform distribution of crop), limiting the usefulness of results. This paper presents a new approach based on Vector Agents (VAs) to extract weeds, more specifically boneseed, from Unmanned Aerial Vehicle (UAV) imagery. VAs are objects that can construct their own geometry and interact spatially with other VAs in the context of Geographic Automata Systems (GAS). The dynamic structure of VAs allows them to directly address real-world objects in an image, such as weeds. In this case, the method can automatically draw the boundary of the real world objects without setting any user-defined parameters, e.g. scale or compactness. We test the proposed model against the ones conventionally used in weed detection, e.g. mean shift and multiresolution. The preliminary results show 8% and 30% improvement in the correctness value of the VA model compared to the mean shift and multiresolution methods, respectively. Keywords: Geographic Automata Systems, Segmentation, Unmanned Aerial Vehicle, Vector Agent, Weed Detection 1. Introduction As a foreign plant species, weeds not only threaten local native species but also ecosystem processes (Reid, 1998). To detect weeds, there are various technologies that can be used, including Unmanned Aerial Vehicle (UAV) imagery. It is a cost-effective tool (Knoth et al., 2013), especially when the need is for a very high spatial resolution or for collection of data at a specific time (Pena et al., 2013). From this detailed imagery, the process of weed detection usually includes two main steps, segmentation and classification. In the former step, we usually use a set of parameters such as scale or color, depending on the segmentation method, to segment an image or create image objects. The information of these objects, such as shape or location, is then applied via a classification algorithm to identify weeds. For example, Pena et al. (2013) used a multiresolution segmentation process to identify weeds. In the proposed method, they used objects ranging from very small to large scale to first find the orientation