Wheat is a staple crop that is grown across the world due to its substantial contribution to human nutrition. Its significance is evident as it provides almost 20% of calories and protein required for daily human consumption. However, wheat yield is affected by rust disease that can reduce 30% of wheat production which is a serious threat to food security. In order to minimize the loss, it is crucial to identify precisely and localize the wheat rust disease and its infection types. For this purpose, several classification and segmentation techniques are used which are based on machine/deep learning models. This paper provides a realistic analysis and evaluation of various segmentation techniques including Watershed, Grab Cut, and U2-Net. These techniques are applied to the wheat stripe rust data to generate multiple datasets such as Watershed segmented data, GrabCut segmented data, and U2-Net segmented data. Subsequently, a pre-trained deep learning model, ResNet-18 is applied to these datasets to assess the impact of segmentation on classification accuracy. The highest classification accuracy (96.196%) is achieved on the dataset segmented by U2-Net. This research collates several state-of-the-art segmentation techniques in terms of correctness and their direct impact on classification accuracy which gives a pragmatic analysis for researchers to choose optimal segmentation technique. The research primarily focuses on the direct impact of segmentation on classification accuracy of wheat stripe rust, which has not been given sufficient focus in earlier researches.