Analyzing the Effectiveness of Image Augmentation for Soybean Crop and Broadleaf Weed Classification

Data is the key for every artificial intelligence (AI)-based application irrespective of the type of data (numerical, categorical, image) being used. Quality and the depth of information in the data determine the performance of the AI model. Before the data is given as input to the classifier, it must be cleaned using pre-processing techniques. The data must also be sufficient enough to produce satisfactory results. The data considered for this work is images and thus emphasizes image augmentation techniques. This paper focuses on analyzing the best image augmentation techniques for deep learning classifiers. It is essential to analyze the effective data augmentation technique for a particular dataset. In this work, 2382 observations (images) from a crop-weed dataset are used to build the classifier. To expand this dataset, 11 image augmentation methods are applied to the training images. Six out of 11 methods show a high level of effectiveness and are chosen for further process. The outcome of every augmentation method is depicted for an in-depth understanding of augmentation techniques. Sixteen convolutional neural network (CNN)-based pre-trained models are built for evaluating the results. However, MobileNet outperformed other models by resulting in an overall accuracy of 99.58% and F1score of 1.0. Moreover, the performance of the model is evaluated using 24 metrics, and the formulas used for calculation are also tabulated in detail. Tables and graphs are represented for understanding the outcome precisely. Future works in image processing with deep learning are also discussed before concluding.