Rose Leaf Diseases Detection Using Convolutional Neural Networks with Texture-Based Patch-Wise Selection

Authors : Birkan Büyükarıkan, Keziban Yalçın Dokumacı

Pages : 949-959

Doi:10.24012/dumf.1771419

View : 61 | Download : 91

Publication Date : 2025-12-30

Article Type : Research Paper

Abstract :In agricultural, early detection of leaf diseases enables control targeted at less chemical input, environmental impact, and expense. Towards this end, the see-and-spray strategy has been sought more intensely in recent decades in applications such as spot/precision spraying and UAV-based treatments. Because roses are converted into high-value medicinal and aromatic products, early identification of leaf diseases prevents quality losses and yield reductions, allowing selective pesticide application and timely intervention. In turn, this optimizes input use and lowers production costs. Convolutional neural networks (CNNs) are widely used for leaf disease detection; however, resizing images to meet fixed input dimensions can cause information loss and hinder the capture of subtle disease cues. To address this limitation, we propose a new approach that combines texture-based patch selection with CNN models for classifying rose leaf diseases. Texture-based patch selection provides a patch-wise pipeline that preserves fine lesion patterns otherwise lost when full images are resized for CNN input. A public rose leaf disease dataset was used. Each dataset image was initially divided into four patches of 224×224 pixels, and the patches with the highest information content were selected using Gray-Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP), and Histogram of Oriented Gradients (HOG). The selected patches were classified using nine different CNN models. These CNN models were evaluated using transfer learning and five-fold cross-validation. Experimental results show that LBP-based patch-wise selection combined with DenseNet121 model achieved an accuracy of 95.73 ± 0.87%, while GLCM-based selection with DenseNet201 achieved 95.48 ± 0.70%. The findings indicate that texture-based patch selection can aid the detection of diseased regions on rose leaves and enable targeted pesticide application.
Keywords : Gül yaprak hastalığı, yama-bilgi seçimi, doku analizi, evrişimli sinir ağı, tespit, hassas püskürtme

ORIGINAL ARTICLE URL