Fig 1.
Architecture of the CNN-BiLSTM model for leaf disease detection and classification.
Fig 2.
Sample plant leaf images from PV database.
Fig 3.
Block diagram of the BiLSTM architecture.
Table 1.
Summary of the existing works.
Table 2.
Crop disease recognition using CNNs.
Fig 4.
Accuracy on crop diseases using DL models.
Fig 5.
Dimensionality reduction/feature extraction.
Table 3.
Crop disease recognition using CNNs.
Fig 6.
Proposed model for CNN-BiLSTM architecture.
Fig 7.
Schematic diagram of proposed methodology.
Table 4.
Parameter settings.
Fig 8.
Model accuracy and loss over epochs.
Fig 9.
Training vs. validation loss for different learning rates.
Fig 10.
Grad-CAM visualization.
Fig 11.
Grad-CAM visualization.
Fig 12.
t-SNE feature embedding.
Fig 13.
Feature importance/activation maps.
Fig 14.
Various pepper leaf disease (a, b, c, d) and Maize leaf disease (e, f, g, h) images.
Table 5.
Disease classes with assigned numbers and image counts.
Table 6.
Disease classes with assigned numbers and image counts with training and testing.
Fig 15.
Example photos from the collection (top row represents pepper and bottom row represents maize).
Fig 16.
Qualitative Representation (a) input image, (b) Resized image, (c) Pre-processed image and (d) Segmented image.
Table 7.
Ablation study on various ImageJ image enhancement filters.
Table 8.
Ablation study: effect of transformer layers, activation functions, pooling layers, and stride size on leaf disease classification.
Table 9.
Ablation study effect of transformer layers, activation functions, pooling layers, and stride size on leaf disease classification.
Table 10.
Ablation study on changing kernel size, pooling layer kernel size, loss function, batch size.
Table 11.
Ablation study: effect of optimizer, learning rate, and image size on model performance.
Table 12.
Ablation studies clearly indicate that model architecture.
Table 13.
Configuration of proposed CNN-BiLSTM architecture after ablation study.
Fig 17.
Improvement in test accuracy over 11 ablation studies.
Fig 18.
Performance with standard errors.
Fig 19.
Comparison of suggested approach with existing works.
Table 14.
State-of-the-art works performance comparison with proposed model.
Table 15.
Comparison of performance analysis of existing methodologies with the proposed methodology.
Fig 20.
Training set class distribution.
Fig 21.
Test set class distribution.
Fig 22.
BS found in test dataset.
Fig 23.
BS and HP unaffected region.
Fig 24.
Bacterial Spot and unaffected region identified from a Google picture.
Fig 25.
Accuracy levels for feature extraction.
Fig 26.
Accuracy levels for feature extraction.
Fig 27.
Time levels for feature extraction.
Fig 28.
Time levels for feature labelling.
Fig 29.
Accuracy levels for feature labelling.
Fig 30.
Feature dimensionality reduction time levels.
Fig 31.
Accuracy levels for feature dimensionality reduction.
Fig 32.
Confusion matrix for different models.
Fig 33.
F1-score levels for feature dimensionality reduction.
Fig 34.
Recall levels for features dimensionality reduction.
Fig 35.
Precision levels for features dimensionality reduction.
Fig 36.
Precision, Recall, F1-score comparisons of proposed model with existing models.
Fig 37.
Accuracy comparison of the proposed model to existing models.
Fig 38.
Testing accuracies and training parameters with the suggested work.
Table 16.
Table 17.
Comparison of state-of-the-artworks test accuracies and training parameters with the suggested work.