Sequential support vector machine classification for small-grain weed species discrimination with special regard to Cirsium arvense and Galium aparine
- Publikations-Art
- Zeitschriftenbeitrag
- Autoren
- Till Rumpf and Christoph Römer and Martin Weis and Markus Sökefeld and Roland Gerhards and Lutz Plümer
- Erscheinungsjahr
- 2012
- Veröffentlicht in
- Computers and Electronics in Agriculture
- Verlag
- Elsevier
- Band/Volume
- 80/
- ISBN / ISSN / eISSN
- 0168-1699
- DOI
- 10.1016/j.compag.2011.10.018
- Seite (von - bis)
- 89-96
Site-specific weed management can reduce the amount of herbicides used in comparison to classical broadcast applications. The ability to apply herbicides on weed patches within the field requires automation. This study focuses on the automatic detection of different species with imaging sensors. Image processing algorithms determine shape features for the plants in the images. With these shape descriptions classification algorithms can be trained to identify the weed and crop species. Since weeds differ in their economic loss due to their yield effect and are controlled by different herbicides, it is necessary to correctly distinguish between the species. Image series of different measurements with plant samples at different growth stages were analysed. For the classification a sequential classification approach was chosen, involving three different support vector machine (SVM) models. In a first step groups of similar plant species were successfully identified (monocotyledons, dicotyledons and barley). Distinctions within the class of dicotyledons proved to be particularly difficult. For that purpose species in this group were subject to a second and third classification step. For each of these steps different features were found to be most important. Feature weighting was done with the RELIEF-F algorithm and SVM-Weighting. The focus was on the early identification of the two most harmful species Cirsium arvense and Galium aparine, with optimal accuracy than using a non-sequential classification approach. An overall classification accuracy of 97.7% was achieved in the first step. For the two subsequent classifiers accuracy rates of 80% and more were obtained for C. arvense and G. aparine.
Highlights
► Image processing combined with machine learning for separation of weeds and crops.
► Weed species are grouped with regard to separability and economic thresholds.
► Classification of hardly separable dicotyledons using a sequential classification.
► Different features prove to be relevant for successive classification steps.
► Specialised classifiers lead to robust weed separation.