Colour-Based Feature Analysis for Aloe Vera Leaf Disease Classification using Deep Learning

Abstract

Accurate classification of Aloe vera leaf diseases is challenging due to overlapping colour patterns between healthy and infected regions under varying lighting conditions. Traditional methods often fail to capture these subtle colour features, emphasizing the need for robust colour based analysis to improve deep learning model accuracy. This research introduces an advanced automated system for classifying Aloe vera leaf diseases, categorising them as healthy, aloe rust, or leaf spot, to facilitate early detection and support sustainable agriculture in Sri Lanka’s expanding Aloe vera leaf sector. Utilizing a dataset of 847 images collected from farms in Jaffna and Peradeniya, the approach integrates sophisticated colour based feature analysis across multiple colour spaces (HSV, LAB, YCbCr, Normalized RGB, Opponent, and YUV) with Discrete Wavelet Transform (DWT) for enhanced representation. State of the art deep learning models, including DenseNet, EfficientNetV2B3, VGG16, and ResNet50, were employed for classification. DenseNet achieved superior performance with 98% accuracy (precision: 0.99, recall: 0.99 for aloe rust), underscoring the efficacy of colour based feature analysis in discerning visually prominent disease patterns. However, dataset imbalance (698 aloe rust, 139 healthy, 276 leaf spot images) led to biases. These results validate the integration of colour based feature analysis and deep learning for plant disease detection, advancing prior computer vision applications in agriculture while revealing challenges with imbalanced data. The framework enhances theoretical insights into colour based feature analysis and DWT for capturing disease specific cues, surpassing manual inspections and offering practical benefits for farmers in arid regions like Hambantota and Puttalam by minimizing crop losses. Limitations include the dataset’s modest size and regional focus, potentially limiting generalization to varied conditions or emerging diseases. Future efforts will prioritize larger, balanced datasets, incorporate texture or spectral features, and investigate ensemble models for greater robustness, with potential extensions to other crops.