Feature extraction for the identification of weed species in digital images for the purpose of site-specific weed control

Publication Type

Contribution to conference

Authors

Weis, M. and Gerhards, R.

Year of publication

2007

Published in

Precision agriculture '07

Editor

J.V. Stafford

Pubisher

Wageningen Academic Publishers , Wageningen, Netherlands

Band/Volume

6/

Series/labeling

Conference on Precision Agriculture (ECPA)

Conference name

6th European Conference on Precision Agriculture (ECPA)

Conference location

Skiathos, Greece

Conference date

Juni

Automated weed detection and classification allow a high spatial density of measurements and
can therefore be used for site-specific application of herbicides at variable rate. A system for the
detection and classification of different crops and weed species is presented. Near-range images
were taken with a bi-spectral camera (IR+VIS) mounted on a vehicle driving at a speed of about
8 km/h. The techniques used analyse the images including pre-processing steps to reduce noise
and to obtain comparable results. A segmentation of green plants and background is achieved by
binarisation. The shapes of all plants were extracted and shape parameters, contour and skeleton
features were calculated. The features were used to classify different weed and crop species.
Their discriminant abilities were tested using data mining and classification algorithms, including
discriminant analysis. Different feature sets were compared to each other and the most promising
were selected for classification. The classification of an image series taken in a field with Hordeum
vulgare in 2006 resulted in a correct classification of 98%. Additionally an image database with
weed and crop samples was created, which can be used as prototypes to set up and test different
evaluation approaches. This database helps to develop new approaches and makes them comparable
to each other.