CLASSIFICATION OF RICE LEAF DISEASES USING VGG-19 MODEL WITH GRADIENT-WEIGHTED CLASS ACTIVATION MAPPING TECHNIQUE
Keywords:
Brown Spot (BS), Bacterial Leaf Blight (BLB), Leaf Blast (LB), Convolution Neural Networks (CNNs), Gradient- Weighted Class Activation Mapping (Grad-CAM)Abstract
Pakistan which is the fourth biggest producer and fifth biggest exporter of rice, struggles to manage rice leaf diseases that seriously affect its crop output and quality. Yet, finding these diseases early and at the right time is vital, though traditional methods turn out to be slow, difficult and expensive. To fix this, we developed a simple, reliable and reduced-cost diagnosis system by applying the VGG-19 deep learning model in Python, TensorFlow and T4 graphic processing units for a faster, more efficient implementation. With very little loss values, the model is able to predict well, achieving 97% accuracy during training, 96.88% accuracy on validation and 95.88% accuracy on testing. We use Grad-CAM to point out in leaf images the regions that affect the outcome of the disease classification by the model. Being able to visualize these data increases our trust in the system and teaches us more about diseases. On the whole, the framework can handle real-time rice plant health monitoring and using it may reduce the need for manual inspections in farming.
