A Soft-Voting Hybrid Ensemble Deep Learning Framework for Robust Rice Disease Detection Using Transfer Learning

Abstract

Rice disease caused severe losses in productivity, quality of grain, sustainable agricultural production, which requires accurate and automatic disease identification. This study introduces an innovative Soft-Voting Hybrid Ensemble Deep Learning Framework for disease detection in rice through transfer learning. The proposed framework involves a combination of DenseNet121, EfficientNetB7, Xception to learn complementary visual representation from the image of rice leaves and then fused the class-probability output using a weighted soft-voting approach. The methodology uses image quality evaluation, resizing, normalization, augmentation, partitioning of the datasets, transfer learning, fine-tuning with categorical cross-entropy and optimization. The dataset used for the experiments consists of 5,130 rice-leaf images across nine disease classes, with the images undergoing preprocessing and augmentation for increased data diversity and model generalization. The performance of the individual CNN models is comparatively analysed based on accuracy, precision, recall, F1 score, Matthews correlation coefficient and confusion matrix analysis. The ensemble model outperformed the individual models, with an accuracy of 97.3%, a precision of 0.972, a recall of 0.971, an F1-score of 0.971 and an MCC of 0.964. The results show the effectiveness of complementary transfer-learning models for fusion of probability level to classify rice disease for automated agricultural monitoring with strong and reliable results.