Browsing by Author "Mahajan, Sukanya"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
ThesisItem Open Access Evaluating XAI for Weed Localization in Crop Images: Survey Based Hypothesis Testing(Sher-e-Kashmir University of Agricultural Sciences & Technology of Jammu, 2024-04-18) Mahajan, Sukanya; Rizvi, S.E.H.Weeds are considered as undesirable vegetation that competes with cultivated plants for essential resources such as light, water, and nutrients. This competition can lead to a decrease in the yield and quality of the desired crops. When examining the potential agricultural losses caused by various factors, it is evident that weeds are the most significant, accounting for approximately 34 percent of the potential loss. In comparison, animal pests and pathogens are of lesser concern, contributing to losses of 18per cent and 16per cent respectively. The impact of weeds on agricultural production is substantial, with an estimated average loss of 5 percent in developed countries, 10 percent in developing countries, and a significant 25 percent in the least developed countries.Weed management involves the elimination of weeds and control its growth through micro-spray, cutting, thermal, electrocution, automated weeding system etc. Presently, weed management and control is executed using mostly herbicides.However, the extensive usage of herbicide causes harm to the environment, raises concern about risks to human health and is expensive.To address these challenges, the present investigation entitled “Evaluating XAI for Weed Localization in Crop Images: Survey Based Hypothesis Testing” have been focused on building an automated weeding solution in the research with the objectives to encompass the development of a binary classifier, generating visual explanation for the results predicted by the classifier and finally evaluating these visual explanationsthrough hypothesis testing. To conduct this study cotton (Gossypium sp.) crop and weeds image dataset was taken from Kaggle. Then“TinyVGG” model was trained, validated and tested on a customized image dataset. Evaluation of this model overtest set resulted in an accuracy of 95.52per cent. F1 score of the model was 95.46per cent. Then, two techniques of Explainable Artificial Intelligence (XAI) i.e. local interpretable model-agnostic explanations (LIME) and gradient-weighted class activation mapping (Grad-CAM) to generate visual explanations on the output predicted by the classifier were adopted. Further,a survey was created with the aim of evaluating these visual explanations by humans, called human-centric evaluations which refers to assessments conducted based on human criteria, evaluating how explanations align with human perceptions and whether they fulfill human needs or not, using hypothesis testing.Based on human-centric evaluation, a significant difference between LIME explanations and Grad-CAM explanations was discovered. And, finally Grad-CAM technique resulted in superior performance when presented to participants.