Disaster Image Classification (EfficientNet-B0)
A fine-tuned EfficientNet-B0 model for high-accuracy disaster scene classification. Developed by Soujanya Subedi, this model leverages transfer learning and modern regularization techniques to identify natural disasters from visual data.
🚀 Model Details
- Architecture: EfficientNet-B0 (Pretrained on ImageNet)
- Framework: PyTorch
- Input Resolution: 224x224 (RGB)
- Target Classes: Cyclone, Earthquake, Flood, Wildfire
- Key Techniques: Mixup Regularization, Data Augmentation, Partial Fine-tuning
📊 Performance
The model demonstrates robust generalization across diverse environmental conditions.
| Metric | Score |
|---|---|
| Validation Accuracy | ~95.5% |
| Test Accuracy | ~92.0% |
Performance Insights:
- Highest Reliability: The model is exceptionally accurate at detecting Wildfires due to distinct color and texture patterns.
- Known Confusion: Minor overlap occurs between Flood and Earthquake scenes, often due to similar visual features like structural debris and environmental damage.
🛠️ Usage
To run inference with this model, use the following PyTorch snippet:
import torch
from torchvision import models, transforms
from PIL import Image
# 1. Setup Architecture
model = models.efficientnet_b0(weights=None)
model.classifier[1] = torch.nn.Linear(1280, 4)
# 2. Load Weights
state_dict = torch.load("disaster_model.pth", map_location="cpu")
model.load_state_dict(state_dict.get("model_state_dict", state_dict))
model.eval()
# 3. Preprocessing
preprocess = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
📂 Dataset & Training
- Source: Disaster Images Dataset (Kaggle)
- Training Setup: Optimized using Adam with CrossEntropyLoss.
- Augmentation: Utilized Random Cropping, Rotation, and Mixup to prevent overfitting and improve scene context understanding.
⚠️ Limitations
- Critical Systems: This model is intended for research and prototyping. It should not be used as the sole decision-maker in real-time emergency response scenarios.
- Visual Similarity: Performance may degrade in low-light conditions or when scenes contain ambiguous debris that mimics multiple disaster types.