|Towards the automated mapping of weed patches in arable fields
A.J. Murdoch, R.A. Pilgrim, P. de la Warr, J. Edwards, P.J.W. Lutman, B. Magry, P.C.H. Miller, S. Morton, T. Robinson, N. Walters
University of Reading, Department of Agriculture,
Earley Gate, P.O. Box 237, Reading RG6 6AR, U.K
Weeds such as black-grass (Alopecurus myosuroides), wild-oat (Avena fatua), barren brome (Bromus sterilis) and cleavers (Galium aparine) often occur in patches in arable fields. Farmers, however, frequently spray whole fields to control patches of such weeds. Given a geo-referenced weed map, technology exists to confine spraying to these patches. In spite of environmental and economic benefits, adoption of patch spraying by arable farmers has, however, been negligible. A major reason for low adoption is the difficulty of constructing weed maps. This paper describes a machine vision system designed to automate the weed mapping process. The primary focus was on identifying the weeds mentioned which typically occur in patches and which can be difficult or expensive to control.
Hypotheses tested include (1) the accuracy of weed identification by machine vision based on one or several field surveys at different growth stages will be adequate to identify weeds and weed patches with the precision needed to create herbicide application maps and (2) images required for mapping can be captured as an adjunct to normal farming operations in each field. The approach adopted was to create maps off-line for use by farmers and advisers in the subsequent growing season rather than for real-time control. Work carried out to date has sought to prove the concept that geo-referenced images captured using a computer-controlled camera mounted on farm machinery during conventional farming operations such as spraying and combine harvesting can be used to automate the process of weed mapping.
A prototype machine vision system was attached to farm machinery (sprayers and combine harvesters) on two farms from June to December 2009, images being captured at times of chemical application and harvesting. Optimal times for black grass identification in wheat in the UK were from flowering until before seed shedding, detection after seeding and of early season seedlings requiring very high resolution images. Late season mapping clearly highlights failure of control and places where seed bank replenishment is likely. Accuracy of identification is being verified by geo-referenced 'ground truth' assessments in the same fields. The potential of automated the process of weed mapping and the value of this approach in comparison to real time weed detection and control, will be discussed.