A Deep Learning Based Approach for Canola and Weed Segmentation in Precision Agriculture

A Thesis Submitted to the Faculty of Graduate Studies and Research In Partial Fulfillment of the Requirements for the Degree of Master of Applied Science in Electronic Systems Engineering, University of Regina. xvi, 91 p. Weeds are unwanted plants that can be eliminated using herbicides. Traditionally, these herbicides are sprayed uniformly which consumes an unnecessary amount and leaves adverse e ects on the environment. We can signi cantly cut this usage by selective spraying. It can be achieved using a decision map produced using eld images. Firstly, for all the images, classi cation is performed at the pixel level in either crop or weed. Then, the percentage of each class is calculated. Finally, these percentages are mapped using Global Positioning System (GPS) locations to generate a nal decision map for selective spraying. This thesis focuses on the classi cation part of the plants at a pixel level. Currently available models are not suitable as they need more storage. We aim for model optimization without performance loss. To accomplish this objective, we start with preparing data. We use Maximum Likelihood Classi cation (MLC) and image processing techniques to label eld images in three classes: background, crop, and weeds. Then, we use ResNet-50 based U-Net like classic architecture and modify it to reduce the number of parameters. In the modi cation we convert the normal convolution into two sub operations: point-wise and depth-wise convolution. This optimization drops the performance which we maintained by introducing the rich low-level features. The variety in low level features is achieved with the help of a parallel feature extraction module which uses dilated lters. Dilated lter requires a small number of parameters to extract a bigger picture and connectivity of the features. In other words, a small kernel is used to approximate a large size kernel. The result is an improvement in semantic accuracy with reduced memory requirement. While training the model, we include data diversity using image augmentation. It also helps solve challenges in the dataset. With extensive experiments over objective function, we choose DICE loss to train the model which provides solution to the data imbalance problem. A comparison is made for quantitative and qualitative analysis with state-of-the-art. We achieve 89.18% mean Intersection Over Union (mIOU) with 86.33%, 83.19%, and 98.01% individual IOU for the crop, weeds, and background, respectively. We also use Average Precision (AP) to report numerical results. Our proposed model is more con dent in classifying weed and background with a score of 92.83% and 99.57%, respectively. Overall, our proposed methodology reports mAP 95.54%, 0.7% more than HRNet MScale. In terms of memory requirement, our proposed model uses only 15M parameters which are 57M less than the state-of-the-art models with a compromise of 1% mIOU score. It only requires 60 Megabytes (MB) of storage which is 10 less than HRNet MScale. Student yes