TITLE:
Rice Leaf Disease Detection Using EfficientNetB0 Transfer Learning with a Weather-Aware Rule-Based Advisory System
AUTHORS:
Md Mahmudul Hassan, Md Shahinul Alam Murad, Nadia Afrin Ritu
KEYWORDS:
Rice Leaf Disease, Deep Learning, Transfer Learning, EfficientNetB0, Convolutional Neural Network, Expert System, Weather-Aware Advisory, Precision Agriculture
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.10,
October
27,
2026
ABSTRACT: Rice is Bangladesh’s most important staple crop, and its yield is repeatedly threatened by disease outbreaks that are difficult to diagnose quickly given the limited number of agricultural extension officers available to smallholder farmers. This paper presents a hybrid diagnosis-and-advisory system for rice leaf disease that couples a deep-learning image classifier with an independent, rule-based weather-risk expert system. An EfficientNetB0 backbone, pretrained on ImageNet, is fine-tuned through a two-phase transfer-learning procedure on a 531-image, five-class, field-collected Bangladeshi dataset covering Bacterial leaf blight, Blast, Brown spot, Leaf smut and Tungro, using computed class weighting to counter marked class imbalance. On a held-out test split of 80 images the classifier reaches 96.25% accuracy and a 96.40% weighted F1-score, outperforming VGG16, ResNet50 and AlexNet fine-tuned on the same dataset and evaluated on the same test split, while using markedly fewer parameters. Independently of the classifier, a forward-chaining expert system evaluates temperature, humidity and rainfall against fifteen agronomically grounded IF-THEN rules, three per disease and each a graded test contributing a full or reduced weight, to produce a HIGH/MODERATE/LOW risk level and a disease-specific treatment advisory. The classifier’s diagnosis and the expert system’s risk assessment are then combined into a single interactive report. The results show that transfer learning with a compact classification head, strong data augmentation and computed class weighting can achieve high accuracy even on a small, imbalanced, real-world dataset, and that pairing a CNN-based diagnosis with an interpretable rule-based advisory layer yields a more directly actionable output for farmers and extension officers than a bare classification label.