Abstract

Rice demand is increasing with the rise in population worldwide, but this crop production is negatively affected due to different fatal diseases. Reported rice disease diagnosis models are imprecise, inefficient, and Taylor made. Hence, this research presents an efficient hybrid model of different rice disease diagnoses to support the agricultural industry's economic growth. The proposed hybrid model is composed of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Self-Attention (SA) modules. The fitness level of the proposed model is evaluated using a test dataset, 5-fold cross-validation (CV), Hosmer- Lemeshow test, Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Four rice leave diseases (Bacterial blight, Blast, Tungro, Brown spot) are diagnosed from the benchmark dataset. From the 5-fold CV metric, the proposed model attained 100% with a 0.001 average accuracy and loss for training samples. Similarly, got 97.51% with a 0.110 average accuracy and loss for validation samples. The proposed model also achieved the higher Receiver Operating Curve (ROC) with Area Under Curve (AUC) rate, precision, recall, and F1-score. The model also obtained minimum RMSE, MAPE and Hosmer Lemeshow test values, revealing that the proposed model is well-fitted. The proposed model also got 100%accuracy, precision, recall, F1-score as well as for testing samples. The performance metrics exhibited that the proposed model's overall performance was perfect and could be used in agriculture for disease identification of rice leaves. The present investigation achieved a high diagnosis rate for rice leave disease identification without over and under-fitting issues. The model has a 97.5%–100% confidence interval for detecting rice leaf disease. Finally, the proposed model will support the agriculture industry in diagnosing rice leaf diseases and monitoring their growth.