Rice Leaf Nutrient Deficiency Classification System Using CAR-Capsule Network

Abstract

Rice, a worldwide grown grain, frequently suffers production issues caused by nutrient imbalances, particularly potassium, nitrogen, and phosphorus. Identifying nutrient deficiencies in rice plants proves challenging due to variations in leaf colour and form. Visually classifying nutritional shortages based on leaf characteristics, such as colour and shape, becomes a complex and resource-intensive task. The intricacies involved make the identification of nutrient deficiencies in rice both time-consuming and expensive. This study presents a computer vision-based deep learning system termed CAR-CapsNet, an upgraded capsule network (CapsNet) that uses contextual attention routing (CAR) to classify rice crop nutrient deficiencies. CAR-Capsnet’s innovative use of contextual attention routing significantly enhances the model’s ability to navigate and interpret complex visual features and patterns, leading to improved classification accuracy compared to previous routing methods. The training and evaluation datasets are sourced from Kaggle, a freely accessible data platform. The dataset consists of 1,155 images of rice leaves, divided into three distinct classes representing deficiencies in nitrogen, phosphorus, and potassium. The dataset undergoes pre-processing using a Wiener filter and adaptive Otsu segmentation. The proposed model was evaluated against CNN and the original CapsNet. CAR-CapsNet outperformed both baseline models in the experiments. CAR-CapsNet classifies rice crop nutrient deficiencies with 97.1% accuracy. Additionally, the model exhibits an impressive recall of 96.9%, an exceptional Kappa score of 95.4%, and an F1-score of 96.9%, highlighting its overall effectiveness. The classifier’s performance was compared with three prior approaches, including Random Forest Regression with an accuracy of 81.82%, SVM with C-means clustering at 92%, and VGG19 at 91.8%. The results demonstrate that the proposed method more effectively classifies rice crop nutrient deficiencies than these methods.

Keywords

• Crop nutrient
• deep learning
• contextual attention routing algorithm
• capsule network
• nutritional imbalances