AI Based Rice Leaf Disease Identification Enhanced by Dynamic Mode Decomposition.

This paper considers the task of rice leaf disease identification using transfer-learned deep learning models. The similarity between various symptoms and the inability to distinguish diseases at glance have become a major challenge in the area. Modern deep CNN-based models have acquired state-of-the-art performance and they frequently incorporate global image as an input to learn the model. The major drawback of such a system for rice leaf disease identification is that the diseases often affect small area; CNN models trained on global images may suffer from irrelevant noisy regions. This paper proposes a Dynamic Mode Decomposition based on attention-driven preprocessing for rice leaf disease identification. The four different categories of rice leaf diseases such as bacterial blight, blast, brown spot and tungro (3416 images in total) are considered for the study. Four types of experiments are conducted in this work, Initially the effectiveness of 10 transfer-learned Deep CNN (DCNN) models for rice leaf disease identification is analyzed with an accuracy of 93.87%, DenseNet121 outperformed other transfer-learned models. Secondly, 3 machine learning models are trained on the deep features extracted from the final fully connected layers of DCNN models. The simulation results indicate that the DenseNet121 deep feature with Random Forest classifier performs better compared to other deep feature and machine learning algorithms. Furthermore, Dynamic mode decomposition (DMD) based attention-driven pre-processing mechanism to localize the infected region is investigated. The sparse component obtained from the DMD algorithm is processed to produce a hard attention map which is further multiplied with the original images to generate hard segmentation map. The influence of hard segmentation maps which localizes the infected regions for rice leaf disease identification is investigated. The performance of 10 DCNN models was evaluated using transfer learning strategy and machine learning models learned on deep features both on original images and DMD preprocessed images. The XceptionNet deep feature with SVM classifier on DMD preprocessed images outperforms other models with a test accuracy of 100%. Finally, the performance of the proposed DMD-based attention-driven pre-processing was investigated on on-field rice leaf images. XceptionNet model achieved a better classification accuracy of 94.33, compared to other transfer-learned models. The Analysis in terms of Accuracy, Precision, Recall and F1-Score revealed that the models trained on DMD pre-processed images exhibit significantly better performance.