Analysing the number of images needed to create robust variable spray maps

The targeted treatment of weeds is an expanding part of precision farming in many countries. Targeted weed treatments, using precision spray maps, reduce herbicide consumption, whilst still maintaining long term weed control. Assembling accurate spray maps is a vital part of this process. However, acceptable accuracy in spray maps is difficult to quantify, due in part to rapid technological advances in cameras, weed recognition software, and herbicide decision support systems (DSS). This research applied a DSS to repeated samples from field gathered weed data. Variability in the herbicide recommendations when different numbers of images were used for the same areas (polygons) within a field were examined. Type 2 errors (not recommending herbicide where it was needed), were analysed separately to type 1 errors (recommending herbicide where it was not needed). Type 2 errors were more common than type 1 errors in diagnosing herbicides to control Viola arvensis in Field 1, and Poaceae species in Field 2, and were also more common with systematically dispersed images compared to randomly dispersed images. In contrast, type 2 errors were less common than type 1 errors for Poaceae species in Field 1. Variability in herbicide recommendations differed for herbicides but was generally reduced (1) with greater numbers of images per polygon; (2) by using regularly arranged images; and (3) for datasets with greater ratios of ‘empty’ (not needing spray) polygons. Targeted treatments reduced herbicide use to 3–11% of the rate recommended for blanket spraying of the same weeds. High numbers of ‘empty’ polygons gave better results with lower relative percentages of type 1 errors. These results highlight the need to focus on reducing type 2 errors in spatial herbicide recommendations.