Study On K-Nearest Neighbor Algorithm Orentied To Weed Identification

It has been an inevitable trend to develop precision agriculture greatly, while weed is one ofthe main restriction factors of agriculture. In order to solve weed problem with variable volumespraying, the study of weed identification based on image processing, in which the identificationmethods decide the accuracy of the results, has become a reseach hotspot both in domestic andoverseas. K-Nearest Neighbor (KNN) has been widely used because of its simplility and easyimplement. Furthermore, KNN has high classification accuracy without building any model, soKNN is used in weed identification with low complexity and high identification rate. Therefore,the study on KNN orentied to weed identification has great theoretical and realistic significances.KNN algorithm orentied to weed identification was discussed in this paper, and the mainresults could be concluded as follows:Firstly, this paper made image segmentation and feature extraction for plants. The reaearchobjects were leaf images of rice and its four kinds of accompanying weeds including calystegiahederacea, curly bristlethistle herb, abutilon hybridum, and chenopodium album at experimentalfarm in Northeast Agricultural University. In segementation, weighted average method was used inimage graying and median filtering method was used in filtering gray images; Then OTSU methodwas used in threshold segmentation of gray images, and the binary images were obtained; ThenRoberts operator, Sobel operator, Prewitt operator, Gauss-Laplacian operator and Canny operatorwere introduced and used to detect edge, and the results showed that Gauss-Laplacian operatorwas the best. Finally, Some feature, including R, G, B, H, S, I and their combination wereextracted, in which eight features with great distinctions were extracted to classify usinghistogram statistical method.Secondly, according to traditional k-nearest neighbor (TR_KNN), two new methods, k-nearest neighbor algorithm based on maximum entropy (ME_KNN) and k-nearest neighboralgorithm based on I-divergence (ID_KNN), were proposed. In TR_KNN, Euclidean distance isused as the distance metric between different samples, which leads to all feature components withthe same weight and the same contribution to classification result. However, maximum entropyand I-divergence, involving weight adjustment coefficient, are distance metrics essentially.ME_KNN and ID_KNN performed best when feature values were positive, while color featurevalues of green plants were positive. So ME and ID were used as distance metrics instead ofEuclidean distance in TR_KNN, which avoided the affect of subjective factors and improvedclassification performance. Finally, ME_KNN and ID_KNN were tested on UCI datasets andartificial datasets, and the results showed that ME_KNN and ID_KNN were superior to TR_KNNin recall, precision, F1measure and accuracy. Lastly, experiments on rice and weed identification were carried out with TR_KNN,ME_KNN and ID_KNN according to the8color features. The results showed the two proposedmethods ME_KNN and ID_KNN were superior to TR_KNN in accuracy and macro F1measure,while ID_KNN was superior to ME_KNN in stability. So ID_KNN was more suitable to weedidentification than ME_KNN and TR_KNN in this paper. The optimal identification accuracy ofrice with ID_KNN was93.33%, the accuracy of calystegia hederacea was86.67%, the accuracy ofcurly bristlethistle herb was93.33%, and the accuracy of chenopodium album was93.33%.Rice and weeds identification were realized sucessfully with ME_KNN and ID_KNNaccording to color features. The computation complexity was low while identification accuracywas high. The study showed that ME_KNN and ID_KNN were superior to TR_KNN; it also hadcertain reference significance to both weeds identification and KNN in future.