Comparison of different classification algorithms for weed detection from images based on shape parameters

Publication Type

Contribution to conference

Authors

Weis, Martin and Rumpf, Till and Gerhards, Roland and Plümer, Lutz

Year of publication

2009

Published in

Image analysis for agricultural products and processes

Editor

Zude, Manuela

Pubisher

Leibnitz Institute for Agricultural Enineering (ATB) , Potsdam-Bornim

Band/Volume

69/

Series/labeling

Bornimer Agrartechnische Berichte

ISBN / ISSN / eISSN

00947-7314

Page (from - to)

53-64

Conference name

1st International Workshop on Computer Image Analysis in Agriculture and 15. Workshop Computer-Bildanalyse in der Landwirtschaft

Conference location

Potsdam, Germany

Conference date

27--28 August 2009

The measurement of weed infestation in the field is a necessary prerequisite for site-specific weed management. Since manual weed sampling does not provide a practical approach, we develop a system for an automatic weed sampling. The system uses bi-spectral images, which are processed to derive shape features for the plants. The shape features are used for the discriminiation of different weed and crop species using a classification step. In this paper we evaluate the derived shape descriptors and different classification algorithms with main focus of the classifier Nearest Neighbours, Decision Tree and Support Vector Machines. Data mining techniques are applied to select an optimal subset of the shape features, which should be used for the classification. Since the classification is the crucial step for the weed detection, the different classification algorithms are tested and their influence on the results is assessed. To find the best parameters for feature selection and classification a parameter optimisation using an evolutionary strategies approach is deployed. The plant shape varies between different species and also within one species in different growth stage. The training of the classifiers is done using prototype information, which is selected manually from the images. All information about the images, the image processing and its results are stored in a database, which can also be used to define the prototypes. To generate a robust, reusable classifier overfitting to the training data has to be circumvented. Some classifiers provide a confidence value for the classification result, which is used to weight the result. This way the inner accuracy of the classification can be used to rate the results and take into account, that some classes are difficult to discriminate. Performance measures for the classification accuracy are evaluated using cross validation techniques and a comparison of the results with manually assessed weed infestation is performed.