Spatial prediction of weed intensities from exact count data and image-based estimates

Collecting weed exact counts in an agricultural field is easy but extremely time consuming. Image analysis algorithms for object extraction applied to pictures of agricultural fields may be used to estimate the weed content with a high resolution (≈ 1m), and pictures acquired at a large number of sites can be used to get maps of weed contents over a whole field at a reasonably low cost. However, these image-based estimates are not perfect and acquiring also exact weed counts is highly useful both for assessing the accuracy of the image-based algorithms and for improving the estimates by use of the combined data. We propose and compare different models for image index and exact weed count and we use them to assess how such data should be combined to get reliable maps. The method is applied to a real data set from a 30 ha field. We show that using image estimates in addition to exact counts allows us to improve significantly the accuracy of maps. We also show that the relative performances of the methods depend on the size of the data set and on the specific methodology (full-Bayes versus plug-in) implemented.