Hyperspectral classification of poisonous solanaceous weeds in processing Phaseolus vulgaris L. and Spinacia oleracea L

Poisonous weeds can occasionally unintentionally be co-harvested and pose a threat to human health as separation techniques during processing are not sufficient. Hence, elimination prior to harvest is required. For this reason, an exploratory study is performed to investigate the possibilities of an automatic detection system. The objective of this article is, firstly, to know if Phaseolus vulgaris and Spinacia oleracea are hyperspectrally separable from Solanum nigrum, Solanum tuberosum and Datura stramonium using spectrometer measurements. Secondly, the influence of different varieties/populations and of different pedohydrological and climatic conditions on this classification is investigated. Finally, it is examined whether it is possible to appoint discriminative wavelengths. To this means, the following analyses were performed: I and II) crop and weed species, and different populations or varieties of these species in varying conditions, were classified using hyperspectral spectrometer measurements and regularized logistic regression (RLR), III) data of consecutive years were investigated for similarities in order to indicate robust important regions in the electromagnetic spectrum with the use of RLR and IV) a subset of commercial off-the-shelf (COTS) filters was created for further research in the field. Results showed that the poisonous weed species D. stramonium, S. nigrum and S. tuberosum are hyperspectrally separable from the investigated crops. The accuracy of the two-class classification of poisonous weeds with S. oleracea and of these weeds with P. vulgaris was 0.982 and 0.977, respectively (I). Inclusion of different crop varieties, weed populations or different growing conditions in the model with S. oleracea and poisonous weeds resulted in a small decrease in weed recall (0.95 vs. 0.99) and crop precision (0.93 vs. 0.97). In future research, care must be taken to proper sample fields to cover the genetic variation present within weed populations and crop varieties, and diverse growing conditions (II). The bands selected using RLR did not show any consistency when using data of consecutive years and, therefore, RLR is not a suitable method to select robust wavelength regions for detection of poisonous weeds in vegetable crops to guide future research (III). With the use of COTS filters it was possible to select ten filters that worked sufficiently for both crops and are recommended for further research in the field. In addition, the authors recommend the use of a high resolution RGB camera to benefit from object-based image analysis to increase classification accuracy (IV).