Weed detection using k-nearest neighbour algorithm

Lim, Arishah (2021) Weed detection using k-nearest neighbour algorithm. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

	Text (24 Pages) Weed detection using k-nearest neighbour algorithm.pdf - Submitted Version Download (5MB)
	Text (Full text) Weed detection using k-nearest neighbour algorithm.pdf - Submitted Version Restricted to Repository staff only Download (20MB)

Abstract

Weed management is one of the important aspects in agriculture sector but traditional weeding method leads to unnoticed growth of weeds which cause loss in production. Early weed detection for plantation health monitoring helps to improve the overall crops yielding. This can be achieved by approaching technology of supervised machine learning for weed detection. The objectives of this project are to define the procedures and methodology in using k-Nearest Neighbours for weed detection, to analyse the performance of the algorithm built and to study the relationship between value of parameter k and rate of accuracy. MATLAB coding and applications were used to develop the weed detection model for classification of weeded and non-weeded areas. Data acquisition was done through obtaining images of areas with soil, average weeds and abundant weeds. The Bag of Visual Words (BoVW) technique was then utilized for extracting features from the image dataset with grid method used for selection of feature key points. The nearest neighbour classifier was trained with the features extracted on training set of images and validated on test set. The analysis of performance was done based on confusion matrix chart which shows the accuracy, precision, recall and F-measure of the trained model. From the simulation, Fine k-Nearest Neighbour with one nearest neighbour had provide the highest accuracy which is 0.98 for training and 0.85 for validation of classifier. The precision and recall for training were 1.00 and 0.97 whereas for validation were 0.77 and 1.00. By changing the value of k for the k-Nearest Neighbour classifier, result showed that the accuracy of model decreases with the increasing of k. The validation accuracy was found to be slightly lower than the training accuracy, which could be due to insufficient work in hyperparameter tuning. Overall, k-Nearest Neighbour algorithm has shown potential of efficiency and reliability in weed detection. Future work should focus on the optimization of the model in order to further improve the performance.

Actions (login required)

View Item