ROBUSTNESS EVALUATION OF TEXTURE-BASED LEMON LEAF DISEASE CLASSIFICATION UNDER COMMON IMAGE DEGRADATIONS

Abstract

The automation of plant disease diagnosis is an emerging research area as it can benefit crop production and minimize crop losses. While many image-based methods for disease classification have been suggested, few studies have examined the performance of handcrafted texture descriptors in realistic images degradation. The study in this paper works to evaluate the robustness of a framework for the classification of lemon leaf disease based on the texture features extracted from the image using Local Binary Pattern (LBP) and classification using the Random Forest (RF) classifier. The experiments were carried out with a six-class lemon leaf disease dataset on Anthracnose, Bacterial Blight, Citrus Canker, Curl Virus, Dry Leaf, and Sooty Mould respectively. To simulate real world image acquisition, six common degradations were applied, namely Gaussian Noise, Gaussian Blur, Salt-and-Pepper Noise, Low Brightness, High Brightness, and Motion Blur. Experimental results show that the proposed framework achieves a mean accuracy of 69.41 ± 2.63% on the original dataset using five-fold stratified cross-validation and is robust to different degradation conditions. The analysis reveals the advantages and disadvantages of texture-based descriptors for practical plant disease recognition and offers valuable insights into the development of a more reliable agricultural computer vision system. The ROC analysis also showed good class separability in each of the six disease categories, with AUC values between 0.890 and 0.963.