🔥 Multimodal Wildfire Burned Area Prediction

Abstract: This project implements a comparative deep learning framework to evaluate whether pre-fire multimodal data (satellite imagery, terrain, weather, and infrastructure) can predict final wildfire burned areas more accurately than traditional Sentinel-only baselines. The study highlights the impact of encoder design on segmentation performance and is optimized for efficiency on CPU-constrained environments.

---

📌 Project Objectives

The primary goals of this research are:

• Predict final burned areas using only pre-fire information.
• Compare a standard Sentinel-only baseline with a novel multimodal deep learning model.
• Analyze the effect of encoder design choices (e.g., ResNet vs. EfficientNet).
• Estimate the specific contribution of each input modality to prediction performance.
• Produce reproducible, slide-ready results for academic presentation.

---

❓ Research Questions

1. Feasibility: Can pre-fire multimodal data predict final burned area with sufficient accuracy to support early emergency decisions?
2. Optimization: Do auxiliary objectives (e.g., land-cover segmentation) improve burned-area prediction?
3. Feature Importance: Which input data modalities contribute most to predictive performance?

---

🧠 Methodology Overview

This project contrasts two modeling approaches to isolate the value of data fusion.

1. Baseline Model

• Architecture: U-Net / FPN-style segmentation
• Input: Sentinel-2 imagery only (12 spectral bands)
• Encoder: ResNet-34 (ImageNet pretrained)
• Purpose: Provide a strong, interpretable reference baseline.

2. Multimodal Model (Final Proposed Solution)

• Architecture: MultiModalFPN
• Inputs:
  • 🛰️ Sentinel-2 imagery
  • 🛰️ Landsat imagery
  • 🏔️ DEM + Road network rasters
  • ☁️ ERA5 Weather (Raster + Tabular)
  • 🔥 Ignition point map
• Encoder: EfficientNet-B4 (ImageNet pretrained)
• Fusion: Attention-based feature fusion blocks
• Purpose: Leverage complementary pre-fire signals to improve segmentation accuracy.

---

🧩 Encoder Strategy

Two encoder strategies were strictly investigated:

1. Multiple encoders with different capacities for each modality.
2. Single unified encoder shared across all modalities.

Decision: The final model adopts a unified EfficientNet-B4 encoder, which achieved:

• Higher IoU and F1 scores.
• More stable training dynamics.
• Reduced architectural complexity.

---

📊 Results Summary

• The multimodal model consistently outperforms the Sentinel-only baseline.
• Encoder choice has a significant impact on segmentation quality.
• Multimodal inputs provide complementary information beyond optical imagery.
• Strong performance is achieved even under CPU-only constraints (MacBook Air).

Detailed quantitative comparisons and plots are stored in the docs/ directory.

---

🧪 Experimental Notes

To ensure transparency and reproducibility:

• Hardware: All experiments were conducted on CPU (MacBook Air).
• Optimization: Auxiliary land-cover loss was disabled to avoid negative transfer.
• Evaluation: Some evaluations use reduced validation subsets for efficiency.
• Reproducibility: All results are clearly labeled in the docs/ folder.

---

✅ Conclusion

This project shows that:

1. Pre-fire multimodal data can meaningfully improve burned-area prediction.
2. Encoder design is a critical performance factor.
3. A unified high-capacity encoder outperforms more complex multi-encoder setups.
4. Careful architectural choices enable strong results under limited resources.

---

👤 Author

Yousef Fayyaz

---

🗂️ Project Structure

WildFire/
│
├── data/                       # Raw and processed datasets
├── geojson/                    # Vector data
│
├── docs/                       # Documentation & Analysis
│   ├── Maps_Graphs/            # Generated inference maps
│   ├── modality_ablation/      # Ablation study results
│   ├── model_comparison/       # Baseline vs Multimodal metrics
│   ├── slide_figures/          # Figures for presentation
│   └── output_result_paper_comparison.txt
│
├── inference/                  # Inference Scripts
│   ├── compare_baseline_vs_multimodal.py
│   ├── deploy_inference.py
│   ├── inference_2.py
│   └── inference_map.py
│
├── src/                        # Source Code
│   ├── Baseline_model/         # Baseline Implementation
│   │   ├── dataset.py
│   │   ├── augmentations.py
│   │   ├── train.py
│   │   ├── main.py
│   │   └── unet_sentinel_best.pth
│   │
│   ├── checkpoint_2/           # Saved Models
│   │   └── best_model_3.pth
│   │
│   ├── figures_tables/         # Visualization Scripts
│   │   ├── export_slide_table.py
│   │   ├── make_qualitative_panels.py
│   │   ├── modality_ablation_quick.py
│   │   ├── paper_comparison.py
│   │   └── plot_threshold_sweep.py
│   │
│   ├── dataset.py
│   ├── model.py
│   ├── train.py
│   ├── utils.py
│   └── main.py                 # Main Training Entry Point
│
├── inference_output/
├── runs/                       # TensorBoard Logs
│
├── .gitignore
├── .gitattributes
└── readme.md



-----