Positive discrimination of minority classes through data generation and distribution: A case study in olive disease classification

Abstract

Deep learning models have achieved remarkable success in various tasks, especially in classification. This success is particularly evident in the precise classification of plant diseases, which is crucial for effective agricultural management. However, accurate classification faces challenges, particularly in data collection, where certain classes are underrepresented, namely the minority classes. This issue can significantly impact model performance. To tackle this challenge, this paper introduces a novel methodology that differs from existing approaches. We focus on addressing the issue of minority classes in image-based classification tasks, particularly for olive diseases. We employ data generation methods, including basic transformations, to produce augmented data and utilize Deep Convolutional Generative Adversarial Networks (DCGAN) to produce generated data. Next, we apply the Frechet Inception Distance (FID) to the generated dataset to select the highest-quality images. We then distribute varying percentages (25%, 50%, 75%, 100%) of this new data into the minority classes of the original dataset. Our data distribution strategies involve incorporating specific amounts of (1) augmented data, (2) generated data, and (3) a combination of both augmented and generated data to achieve target percentages (T.P) in the resulting dataset. Our experiments focus on classifying olive diseases into seven distinct categories using a pre-trained Convolutional Neural Network (CNN) architecture. We observe significant improvements in the model’s performance, particularly in the accurate classification of minority classes. This approach enhances diagnostic accuracy and optimizes data distribution, which is crucial for effectively addressing the challenges posed by minority classes.