Enhanced Classification of Coffee Leaf Biotic Stress by Synergizing Feature Concatenation and Dimensionality Reduction

Abstract

Significant yield challenges are posed by biotic stress on coffee leaves, which has a negative effect on the revenue generation of this highly utilized commodity. Numerous studies have proposed techniques for the early detection and classification of biotic stress in coffee leaves. In this study, we propose a technique called extracted feature ensemble (EFE) for classifying healthy and infected classes. Transfer learning-based convolutional neural networks (CNNs) and custom-designed features are used to improve classification performance. Under the concept of EFE, three methodologies are proposed for evaluating various extracted feature combinations and determining the effect of dimensionality on the performance of the model. In addition, a semi-segmentation approach is used to guide the extraction of informative foreground details, while non-segmented inputs are used to improve the model's robustness against complex background noise. By improving three open-source datasets for biotic stress categorization in coffee leaves, a new dataset was created and employed. The first proposed method, ECNN, focused on the effective concatenation of five CNNs and obtained a classification accuracy of 93.45% using a decision tree classifier, exceeding the maximum individual accuracy of 86.07% from Mobile-Net v3 features. In addition, the HLGGM method was investigated, which demonstrated an enhanced accuracy of 99.16% by combining dimension-reduced Mobile-Net v3 features with handcrafted features. HLGCM, the final approach represented, aimed at extracting features from dimensionality-reduced handmade and CNN-based data, and ultimately succeeded in accomplishing an accuracy of 99.49 percent by using decision tree model. The obtained results demonstrate the efficacy of feature concatenation in enhancing the classification model's discriminative capabilities and classification accuracy. The appropriate combination of hand-made and CNN-based features gives better accuracy and interesting insights into the effect of feature reduction on model classification efficiency. The article offers dimensionality reduction, directed learning, and feature concatenation techniques for identifying coffee leaf diseases. This work can aid in the development of computationally efficient and accurate disease control and coffee plant sustainability strategies.