Evaluating The Efficacy Of Hybrid Deep Learning Models In Rice Variety Classification

Abstract

In this comprehensive study, we have advanced the field of agricultural technology by developing and comparing multiple deep learning models for the classification of rice varieties. Conducted in the agriculturally rich regions of Southern Bangladesh, our research utilized a diverse dataset comprising 20,000 high-resolution RGB images representing five principal rice varieties. The study primarily focused on a custom-engineered hybrid deep learning model, designed specifically for this agricultural application. This model's architecture encompasses an initial convolutional layer, zero-padding, batch normalization, and max pooling, followed by residual blocks that address the vanishing gradient problem, and concludes with Global Average Pooling leading into a Support Vector Machine (SVM) for final classification. Additionally, we incorporated and evaluated the performance of two renowned deep learning models: MobileNetV2 and VGG16. These models were adapted and fine-tuned to suit the specific requirements of our dataset and task. Across various metrics, including precision, recall, and F1-score, our hybrid model demonstrated superior performance, achieving an exceptional 99% accuracy. This was notably higher compared to the 95% and 93% accuracy achieved by VGG16 and MobileNetV2, respectively. Various optimizers, including SGD, RMSprop, Adam, and Nadam (all with a learning rate of 0.001), were employed to refine our models, with the Adam optimizer emerging as the most effective across all models.