Computer-vision-based weed identification under field conditions using controlled lighting

The methods of digital image analysis were used to develop an identificatio n system for weeds in crops. Two vegetable crops (cabbage and carrots) and a number of naturally occurring weed species were used to develop the class ification algorithms. Considering the rougher environment, special attentio n was given to the open-field experiments. The images were obtained with a device that provided controlled lighting conditions. The analysis was carri ed out off-line. Eight different morphological features and three colour fe atures were calculated for each single object to build a joint feature spac e. On the basis of sample data sets of each class, statistics were carried out to determine the features, which are suitable for discrimination. A mem bership function based on a fuzzy logic approach was formed and used for th e classification. The experiments showed that colour features can help to i ncrease the classification accuracy. Moreover, colour was used successfully for the segmentation procedure of plants and soil. Depending on growth sta ge, weed density and method of calculation between 51 and 95% of the plants were classified correctly. Problems still exists by separating and allocat ing single plants in plant stands where the plants have grown together. Com pared to other studies the plant identification system presented is an impr ovement, especially considering that the experiments were carried out under field conditions. (C) 2001 Silsoe Research Institute.