Deep-CNN Based Multi-Class Classification and Accurate Severity Assessment of Biotic Stress in Paddy Leaves

Abstract

Rice is a staple food for over half of the world's population, particularly in Asia. It's a major source of carbohydrates, providing essential energy for daily activities. Rice cultivation plays a crucial role in the livelihoods of millions of farmers worldwide. It contributes significantly to the GDP of many countries. If left unchecked, biotic stress can cause substantial yield losses, leading to food insecurity and economic hardship for farmers. Early detection and management are crucial for preventing these losses. CNNs are a class of deep learning models well-suited for image classification tasks and can be easily scaled to large datasets and complex classification problems. The Automatic stress severity assessment can save time and resources compared to manual assessments, allowing for more efficiency in terms of decision-making. We proposed a Deep-CNN model, that utilizes the Paddy Doctor dataset with nine stress classes and one healthy class. we also addressed the imbalance in the dataset to avoid overfitting and performed a targeted augmentation technique. Multiple classes were classified and predicted on the proposed model. Extensive experimentation was carried out for tuning the parameters of the model. The proposed model achieved high accuracy of 94.4% while EfficientNetB0 achieved 93.5%. Our findings demonstrate that the model outperformed classification for all the 10 classes of the dataset. Using the predicted image, for every stress class the stress characteristics vary with color, we define color thresholds and apply a mask on the image to evaluate the stressed area and generate a severity report. This research demonstrates a promising solution to combat biotic stress in rice. It offers the potential to revolutionize disease management and empower rice-growing communities worldwide to safeguard their livelihoods and contribute to global food security.