Automatic derivation of weed densities from images for site-specific weed management

Publikations-Art

Kongressbeitrag

Autoren

Weis, Martin and Gerhards, Roland

Erscheinungsjahr

2009

Veröffentlicht in

Proceedings of the Joint International Agricultural Conference

Herausgeber

C. Lokhorst and J.F.M. Huijsmans, R.P.M. de Louw

Verlag

Wageningen Academic Publishers , Wageningen, Netherlands

Band/Volume

JIAC2009/

Serie/Bezeichnung

ECPA (European Conference on Precision Agriculture)

ISBN / ISSN / eISSN

978-90-8686-114-9

Seite (von - bis)

349-354

Tagungsname

Joint International Agricultural Conference

Tagungsort

Wageningen, Netherlands

Tagungsdatum

6.-8. Juli

Site-specific herbicide applications can save large amounts of herbicides and improve management practices. One crucial part of a system for site-specific weed management is the measurement of the spatial variability of weed densities. A system was developed to identify different weed species from images taken in the field. The automation has the potential to increase the spatial density of weed sampling points. A manual sampling with high density of points is unfeasible due to the costs. Image processing algorithms are used to generate a shape description for each plant in the image. A classifier can be constructed that assigns weed and crop classes to the plants based on the shape features. Weed density maps are generated using the results of the classification. The weed maps are transformed to application maps, which are used for the site-specific herbicide application.

The shape of the plants vary with their growth stage and may be segmented into parts, e.g. single leaves, in the image processing. Therefore different classes for each species need to be introduced into the process. To avoid over- or underestimation of the actual number of weeds some of the classes are aggregated using weight factors. To derive transformation functions the results of the automatically derived weed counts are compared to manual measurements of weed densities, which were derived from the images and in the field. The results show that the raw classification results are linearly related to the actual number of manually counted weeds and the results can be used as input for a site-specific decision component.