Automatic classification of grassland herbs in close-range sensed digital colour images

Abstract

The broad-leaved dock (<em>Rumex obtusifolius </em>L. (RUMOB)) is one of the most harmful and persistent weed species on European grassland and it has been spread into the temperate grassland regions throughout the world. Large dry matter contributions of <em>Rumex</em> obtusifolius L. reduce the quality of the standing forage considerably because of the poor palatability of leaves and tillers and withdraw water and nutrient from surrounding plants. For Central Europe it is estimated that more than 80% of all herbicides used in conventional grassland farming are used to control <em>Rumex</em> species. Until today, herbicides are applied over the whole field, even if <em>Rumex</em> plants are not homogeneously distributed area-wide.<br /> Recently developed precision farming techniques based on weed mapping that use mainly image processing, enable site-specific spraying of weeds in arable crops. Until today those techniques have not been applied to grassland weed sensing. Compared to the identification of isolated individual plants on a rather uniform soil background in arable crops, image processing for a more complex environment as grassland requires a different approach.<br /> The aim of the thesis was to develop an image processing procedure for automatic detection of grassland weeds using close-range digital colour images, focussing on the detection of RUMOB. A field experiment has been established with grassland plots populated with RUMOB and the other typical broad leaved grassland weeds <em>Taraxacum officinale</em> Web. (TAROF) and <em>Plantago major</em> L. (PLAMA). Digital colour images have been taken from around 1.5 m above ground at three dates in 2005. Image acquisition was done automatically by a vehicle driven on rails alongside to the experimental plots, whereby nearly constant recording geometry conditions were guaranteed. Images were taken during cloud cover in order to avoid direct sunlight.<br /> Using the images from 2005 an object-oriented image classification has been developed. Thereby, the leaves of the weeds were separated from the background using parameters of homogeneity and morphology, resulting in a binary image. The remaining image objects in the binary image were contiguous regions of neighbouring pixels related to the object classes of the weed species, soil, and residue objects. Geometrical-, colour and texture features were calculated for each of these objects. Discriminant analysis exhibited that colour and texture features contribute most to the discriminating of objects into the different classes. In a Maximum Likelihood classification these features were used to differentiate the objects into their respective classes. High overall accuracies and even higher RUMOB detection rates were achieved. The algorithm has been modified and applied to images of varying image resolutions. High classification accuracies have been achieved with all image resolutions, whereby the processing time could be improved for images with lowest resolutions.<br /> Images were taken at 13 dates over the two grassland growths in 2006. In all the images the plant species were classified automatically using the developed image classification integrated in a graphical user interface software. The coordinates of the objects classified as RUMOB were transformed into Gauss-Krueger system to generate distribution maps of this weed. The combination of object density and area further decreased its misclassifications. RUMOB classification rates across the season were analysed and phenological stages have been identified on which classification performed best.<br /> The results demonstrate high potential of machine vision for weed detection in grassland. A classification procedure based on image analysis and Geographic Information System (GIS) post-processing has been developed for detecting<em> Rumex obtusifolius</em> L. and other weeds in grassland with high accuracy. Future projects might focus on the application to real grassland conditions and the derivation of RUMOB distribution maps. Thus, herbicide application maps can be calculated, utilized for site-specific weed control. The development of an image acquisition unit to be mounted on a driving vehicle along with a standardization of image recording is going to be the main focus.