Accurate and resource-efficient automated diagnosis is a cornerstone of modern agricultural expert systems. While Convolutional Neural Networks (CNNs) have established benchmarks in plant pathology, their ability to capture long-range spatial dependencies is often limited by standard pooling layers, and their high memory footprint hinders deployment on portable devices. This paper proposes a lightweight hybrid CNN-LSTM system for bean leaf disease classification. By integrating an LSTM layer to model the spatial-sequential relationships within feature maps, our hybrid architecture achieves a 94.38% accuracy while maintaining an exceptionally small footprint of 1.86 MB; a 70% reduction in size compared to traditional CNN-based systems. Furthermore, we provide a systematic evaluation of image augmentation strategies, demonstrating that tailored transformations are superior to generic combinations for maintaining the integrity of diagnostic patterns. Results on the ibean dataset confirm that the proposed system achieves state-of-the-art F1 scores of 99.22% with EfficientNet-B7+LSTM, providing a robust and scalable framework for real-time agricultural decision support in resource-constrained environments. The code and augmented datasets used in this study are publicly available on this Github repo.
Figures & tables
Figure 1: Three classes of bean leaf images in the ibean dataset
Figure 2: Augmentation effect
Figure 3: Custom architecture of lightweight models
Parameter
Bean-CNN
Bean-CNN-LSTM
No. of hidden units in the 1st and 2nd layer
64, 8
64, 16
Optimiser
Adam
Adam
Learning rate
1e-3
1e-3
Epochs
40
40
Optimiser
Adam
Adam
Batch size
32
32
Table 1: Configuration of custom lightweight models
Figure 4: Activation map by custom lightweight model
Figure 5: Training Bean-CNN on the original set
Figure 6: Training Bean-CNN-LSTM on the original set
Figure 7: Graph representation of the best test accuracy: Comparison of two custom lightweight architectures
Bean-CNN
Training set
Accuracy
Loss
F1 Score
MCC
Original
0.8410
0.5180
0.8410
0.7614
Brightness
0.8769
0.5180
0.8751
0.8163
Combination
0.8872
0.6561
0.8882
0.8329
Crop
0.8821
0.4597
0.8844
0.8267
Flip
0.8872
0.5089
0.8867
0.8307
Table 2: Top-1 Test performance of custom lightweight models : Results across training sets - the best shown in bold
Figure 8: The average test accuracy on each training set: The red line indicates the median value, and a box represents the distribution between 25 and 75 percentiles. Small hollow circles indicate outliers.
Figure 9: Box-plot representation: The average performance of our lightweight custom models (Bean-CNN and Bean-CNN-LSTM) from 30 training runs
Class
Precision (%)
Recall (%)
F1 Score (%)
Angular leaf spot
100.00
91.18
95.38
Bean rust
89.23
93.55
91.34
Healthy
94.12
98.46
96.24
Overall Acc.(%)
94.36
Overall F1(%)
94.38
Table 3: Performance of the best Bean-CNN-LSTM
Figure 10: Confusion matrix: The best Bean-CNN-LSTM model
Class
Precision (%)
Recall (%)
F1 score (%)
Angular leaf spot
97.73
100.00
98.85
Bean rust
100.00
97.67
98.82
Healthy
100.00
100.00
100.00
Overall Acc.(%)
99.22
Overall F1(%)
99.22
Table 4: Classification scores of both the EfficientNetB7+FC & EfficientNetB7+LSTM Models
Source
Method
Acc. (%)
F1 Score (%)
Abed et al., 2021 [ 3 ]
DenseNet121
91.02
–
Elfatimi et al., 2022 [ 11 ]
MobileNetV2
92.97
92.94
Singh et al., 2023 [ 39 ]
EfficientnetB6
91.74
–
Önler 2023 [ 30 ]
MobileNetV2
99.24
–
Suntoyo et al., 2024 [ 43 ]
DenseNet121
96.90
97.00
Jain & Aneja 2025 [ 21 ]
InceptionV3
91.00
91.00
Table 5: Comparison with previous works
Models
Model size (MB)
Number of parameters
Best accuracy (%)
Bean-CNN
6.11
527,171
89.23
Bean-CNN-LSTM
1.86
155,219
94.38
EfficientNetB7 + FC
260.56
67,825,595
99.22
EfficientNetB7 + LSTM
250.20
1,377,667 / 52,844,795
99.22
Table 6: Our model summary
Figure 11: GradCAM Feature maps. The heatmaps indicate that the model correctly prioritises the necrotic centres of the lesions rather than the leaf edges, validating the management reliability of the system.