An empirical ablation study on model design choices for efficient, deployable crop disease classifiers — trained on a 4GB consumer GPU, ready for edge inference.
Most plant disease detection projects pick one model and report one number. This project asks a different question: what actually drives accuracy — and what breaks it?
Three factors were systematically isolated across two datasets representing opposite ends of the data spectrum:
1. Network depth must match dataset size — or the model collapses. A 22-layer CNN trained on 2,152 potato images didn't just underperform. It collapsed to 33.3% accuracy — random guessing across three classes. The 7-layer model hit 96.67% on the same data. Depth isn't better; it's contextual.
2. Augmentation + attention recover what data scarcity loses. Adding squeeze-and-excitation (SE) attention blocks and a targeted augmentation pipeline to the 7-layer potato model pushed accuracy from 96.67% → 97.8% (Macro F1 = 0.96) — within 7.39M parameters and 53ms inference on a 4GB GPU.
3. Background removal creates a domain alignment requirement. U²-Net background removal on the tomato dataset produced a model that dropped from 85.99% to 78.9% when tested on raw field images (BR→Orig). The same model, when test images were also preprocessed (BR→BR), recovered robustness — demonstrating that the preprocessing protocol is as important a design decision as the architecture.
| Dataset | Volume | Architecture | Accuracy | Observation |
|---|---|---|---|---|
| Potato | 2,152 images | 7-layer (Shallow) | 96.67% | Optimal capacity — prevents over-memorization |
| Potato | 2,152 images | 14-layer (Medium) | 94.22% | Slight variance inflation from excess capacity |
| Potato | 2,152 images | 22-layer (Deep) | 33.33% | Model collapse — collapses to random guessing |
| Tomato | 18,160 images | 7-layer (Shallow) | 80.33% | Underfits 10-class morphological diversity |
| Tomato | 18,160 images | 14-layer (Medium) | 85.99% | Data-capacity balance achieved |
| Crop | Architecture | Preprocessing | Accuracy | Macro F1 | Parameters | Inference |
|---|---|---|---|---|---|---|
| Potato | 7-layer CNN + SE Attention | Heavy augmentation | 97.80% | 0.96 | 7.39M | 53.3 ms/img |
| Tomato | 14-layer CNN + SE Attention | U²-Net BR → Orig | 78.90% | 0.73 | 6.52M | 55.0 ms/img |
| Training Domain | Inference Domain | Accuracy |
|---|---|---|
| Original images | Original images | 85.99% |
| Background-removed (BR) | Original images | 78.90% ← domain mismatch |
| Background-removed (BR) | Background-removed (BR) | Superior robustness ← aligned deployment |
The full inference pipeline operates in three sequential stages:
Input Image
│
▼
[Stage 1] YOLOv8 Leaf Localization
│ Crops the leaf bounding box, removes extraneous scene clutter
│
▼
[Stage 2] U²-Net Background Segmentation
│ Pixel-wise saliency map isolates leaf structure from soil/shadows
│ (Optional — enables BR→BR deployment alignment)
│
▼
[Stage 3] CNN + SE Attention Classification
7-layer (Potato) or 14-layer (Tomato) with mid-level Squeeze-and-Excitation blocks
Grad-CAM confirms attention focuses on lesion regions, not background
Hardware: All training and experiments conducted on NVIDIA RTX 2050 (4GB VRAM), Intel Core i5-11260H, 8GB RAM — demonstrating feasibility on constrained hardware.
| Study | Crop | Model | Accuracy | Notes |
|---|---|---|---|---|
| This work | Potato | 7-layer CNN + SE Attention | 97.8% / F1 0.96 | SOTA, 7.39M params |
| This work | Tomato | 14-layer CNN + SE Attention | 78.9% / F1 0.73 | Competitive under domain alignment |
| BSPP Journal (2025) | Tomato | CNN baseline | 75–85% | Realistic field conditions |
| AgriFusionNet (2025) | Tomato, Grape | CNN + multimodal fusion | 93–95% | SOTA, but computationally heavy |
| Nagpal et al. (2024) | Wheat, Barley | CNN–RNN hybrid | 92% / F1 0.90 | Sequential modeling, heavier |
| Springer Review (2024) | Potato, Tomato | CNN variants | 85–99% | Dataset-dependent |
Our potato model matches or exceeds SOTA while remaining under 7.5M parameters. The tomato model is competitive with field-condition benchmarks, without multimodal sensor inputs.
Potato — 2,152 images across 3 classes (Early Blight, Late Blight, Healthy) Augmented to 18,605 images for training using spatial rotations (±25°), horizontal/vertical flips, zoom (±20%), brightness/contrast shifts, and Gaussian noise.
Tomato — 18,160 images across 10 classes (Bacterial Spot, Early Blight, Late Blight, Leaf Mold, Septoria Leaf Spot, Spider Mites, Target Spot, Yellow Leaf Curl Virus, Mosaic Virus, Healthy) Background-removed variant created using U²-Net (~18,000 images) to enable domain alignment experiments.
Both datasets split 70% train / 15% validation / 15% test.
| Setting | Value |
|---|---|
| Optimizer | Adam |
| Learning rate | 1e-4 with plateau decay |
| Loss | Categorical cross-entropy |
| Batch size | 32 |
| Max epochs | 25 (early stopping on val accuracy) |
| Image size | 224 × 224 |
| Frameworks | TensorFlow/Keras (CNNs), PyTorch (U²-Net) |
A full interactive interface is included for exploring results, uploading leaf images through the pipeline, and visualizing Grad-CAM outputs.
🏠 Home Page & Landing Dashboard
🥔 Potato Pipeline & Diagnosis
🍅 Tomato Pipeline & Diagnosis
📊 Model Analysis & Grad-CAM Dashboards
git clone https://github.com/Tarun995/Lightweight-CNN-Attention-for-Plant-Disease-Detection.git
cd Lightweight-CNN-Attention-for-Plant-Disease-Detectionpython -m venv venv
venv\Scripts\activate # Windows
# source venv/bin/activate # macOS/Linux
pip install -r requirements.txtWeight files (.keras, .pth, .pt) are hosted on Google Drive due to size.
Automated download:
# PowerShell
.\scripts\download_weights_gdown.ps1# Bash / macOS / Linux
bash scripts/download_weights_gdown.shManual: Google Drive folder → extract into models/
models/
├── disease/ potato.keras tomato.keras maize.keras
├── yolo/ yolo_leaf_best.pt yolo_lesion_best.pt
└── u2net/ u2net.pth
streamlit run app.py├── app.py ← Streamlit entrypoint
├── pages/
│ ├── 2_Detection.py ← Image upload & pipeline diagnosis
│ ├── 3_Model_Analysis.py ← Confusion matrices, evaluation reports
│ ├── 4_Research_Dashboard.py ← Baseline vs. attention comparison
│ ├── 5_GradCAM.py ← Grad-CAM interpretability heatmaps
│ └── 6_About.py ← References and contributors
├── utils/
│ ├── attention_layer.py ← SE block implementation
│ ├── disease_info.py ← Class labels and treatment guides
│ ├── model_loader.py ← Cached model loading
│ ├── pipeline.py ← Full inference pipeline
│ └── u2net_bg_removal.py ← U²-Net preprocessing helper
├── models/ ← Weights directory (gitignored)
├── results/
│ ├── confusion_matrices/ ← Per-model validation matrices
│ ├── evaluation/ ← Classification report CSVs
│ └── gradcam/ ← Pre-computed Grad-CAM outputs
└── requirements.txt
- Don't default to deep models. On datasets under ~5K images, a 7-layer CNN with augmentation will outperform a 22-layer model by 60+ percentage points.
- Preprocessing protocol is a hyperparameter. If you train on background-removed images, you must deploy on background-removed images — or accept the accuracy penalty.
- SE attention adds almost nothing to parameter count. Going from 7-layer baseline to 7-layer + SE attention costs negligible parameters but recovers measurable F1.
- Edge deployment is viable. Sub-60ms inference on a 4GB GPU means this pipeline runs on hardware accessible to agricultural AI applications.