Advanced clustering and transfer learning based approach for rice leaf disease segmentation and classification.

Abstract

Rice, the world's most important food crop, requires an early and accurate identification of the diseases that infect rice panicles and leaves to increase production and reduce losses. Most conventional methods of diagnosing diseases involve the use of manual instruments, which are ineffective, imprecise, and time-consuming. In light of such drawbacks, this article introduces an improved deep learning and transfer learning method for diagnosing and categorizing rice leaf diseases proficiently. First, all input images are preprocessed; the images are resized to a fixed size before applying a sophisticated contrast enhanced adaptive histogram equalization procedure. Diseased regions are then segmented through the developed gravity weighted kernelised density clustering algorithm. In terms of feature extraction, EfficientNetB0 is fine-tuned by subtracting the last fully connected layers, and the classification is conducted with the new fully connected layers. Also, the tent chaotic particle snow ablation optimizer is added into the learning process in order to improve the learning process and shorten the time of convergence. The performance of the proposed framework was tested on two benchmark datasets and presented accuracy results of 98.87% and 97.54%, respectively. Comparisons of the proposed method with six fine-tuned models show the performance advantage and validity of the proposed method.

Code and data availability

The paper uses two public rice leaf disease image datasets (Kaggle and Mendeley) and releases its authors' source code on Zenodo. The Kaggle dataset URL and Zenodo DOI are explicitly given in the Data Availability statement and match allowed URLs; the Mendeley dataset DOI is not among allowed URLs so it is excluded.