Rice Plant Disease Diagnosis Using SqueezeNet and Deep Transfer Learning

Abstract

Rice serves as a fundamental food source for around 50% of the world's population, mostly in Asia, where agriculturalists have difficulties due to several rice illnesses that may result in substantial crop losses. Timely identification of these illnesses is essential to avert such losses; yet swift and precise diagnosis continues to be challenging owing to constrained knowledge and resources. This research investigates the use of deep transfer learning for the automation of identifying and classifying rice leaf diseases, including blast, brown spot, blight, sheath blight and tungro. We have sourced dataset consisting of 2550 image samples divided into five categories from the Kaggle. Each category has 510 images of infected leaves. By using contrast stretching for image enhancement and data augmentation for data enrichment, we applied a modified SqueezeNet pre-trained deep network on processed dataset, achieved 99.30% accuracy in disease recognition. The final convolutional layer (conv. layer 10) of the pre-trained SqueezeNet is modified by applying multiscale feature aggregation (MFA) in place of 1 × 1 standard convolution. MFA consists of two parallel convolution paths with different kernel size to captures diverse features of the infected lesions. The model's proficiency is highlighted by precision values ranging from 0.972 to 1.000 and recall values between 0.980 and 1.000, whereas maintaining an extremely low error rate between 0.0% and 0.3%, highlighting its high effectiveness. In a comparison with state-of-the-art (SOTA) models under a similar experimental setup, the proposed model demonstrates superior performance in terms of precision, recall, F1-score and accuracy. The proposed method offers a fast, cost-effective and accurate solution to assist farmers in disease detection, even with small datasets and complex backgrounds.