Statistics-based segmentation using a continuous-scale naive Bayes approach Morten Stigaard Laursen a,⇑ , Henrik Skov Midtiby a , Norbert Krüger a , Rasmus Nyholm Jørgensen b a University of Southern Denmark, Niels Bohrs Allé 1, 5230 Odense M, Denmark b Faculty of Agricultural Sciences, Finlandsgade 22, 8200 Aarhus N, Denmark article info Article history: Received 2 January 2014 Received in revised form 6 October 2014 Accepted 9 October 2014 Keywords: Naive Bayes Segmentation Plant–soil discrimination Vegetation indices RGB + NIR abstract Segmentation is a popular preprocessing stage in the field of machine vision. In agricultural applications it can be used to distinguish between living plant material and soil in images. The normalized difference vegetation index (NDVI) and excess green (ExG) color features are often used in the segmentation of images with multiple color channels. In this paper, a Bayesian method is used to combine existing color features into a common color feature. This feature is then used to segment images into separate regions containing vegetation and soil. The common color feature produces an improved segmentation over the normalized vegetation difference index and excess green. The inputs to this color feature are the R, G, B, and near-infrared color wells, their chromaticities, and NDVI, ExG, and excess red. We apply the developed technique to a dataset consisting of 20 manually segmented images captured under artificial illumination. The results show that our combined feature enables better segmentation using the individual color features. Better segmentation allows for more robust vision-based weeding, thereby allowing for lower safety margins within cell-sprayers and lower herbicide usage. Ó 2014 Elsevier B.V. All rights reserved. 1. Introduction In agriculture, increased yield requirements combined with decreases in profit per acre and permitted pesticide usage demand more intelligent solutions. The use of site-specific crop manage- ment to make fine-grained spatial decisions can result in signifi- cant savings. However, this requires a decision support system, to prevent the labor requirements increasing. The decision support system in turn requires data about the field conditions. One method of obtaining this data is to use machine vision systems. In many machine vision applications, segmentation is applied as a preprocessing step, dividing the scene into areas of interest (vegetation) and areas that are of no interest (soil). This allows the analysis to focus on the objects of interest. The results of seg- mentation can be used directly to calculate shape-based features such as area, perimeter length and symmetry descriptions, or indi- rectly as a mask for analyzing color and texture features. The objects of interest will be referred to as the foreground regions, whereas objects that are not of interest are classed as the back- ground. The problem of segmenting vegetation from background areas in agricultural images has been described by several papers. Lei (1995) proposed a segmentation approach based on a naive Bayesian assumption, using the chromaticities of individual colors as input data, but reported problems with changes in illumination. Woebbecke et al. (1995), Meyer and Neto (2008) examined several combinations of red, green, and blue pixel values to distinguish between vegetation and different background types. Excess green (ExG) was found to be effective in locating vegetation, but had a high false positive rate; excess red (ExR) could assist in recognizing some of these false positives. The combination of ExG and ExR was found to be suitable for detecting vegetation. Guo et al. (2013) pro- posed a segmentation algorithm based on a decision tree model using rgb, yCbCr, HSL, HSV, CIE L / a / b, and CIE L / u / v color schemes as input data. They reported similar results as for ExG– ExR in non-sunny conditions, and better performance under sunny conditions. Andersen (1999) proposed a dichromatic reflection model for soil and vegetation, but did not investigate the impact of other objects present in real-world scenes, such as rocks and dead plant material. In this paper, we investigate the problem of distinguishing between vegetation and soil based on inputs from separate color channels and indices, such as ExG (Woebbecke http://dx.doi.org/10.1016/j.compag.2014.10.009 0168-1699/Ó 2014 Elsevier B.V. All rights reserved. ⇑ Corresponding author. Tel.: +45 2440 8401; fax: +45 6550 7354. E-mail addresses: msl@mmmi.sdu.dk (M.S. Laursen), hemi@mmmi.sdu.dk (H.S. Midtiby), norbert@mmmi.sdu.dk (N. Krüger), Rasmus.Jorgensen@agrsci.dk (R.N. Jørgensen). Computers and Electronics in Agriculture 109 (2014) 271–277 Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag