WildFire

Multimodal wildfire burned-area prediction from pre-fire satellite, terrain, weather, and ignition data.

WildFire is a research-focused deep learning project that estimates wildfire burned area using only information available before the fire fully develops. The repository compares a Sentinel-only baseline against a multimodal segmentation pipeline that fuses Sentinel-2, Landsat, DEM, ERA5 weather, infrastructure context, and ignition priors for fire events in Piedmont, Italy.

This repository is the cleaned, portfolio-ready version of the project: code is organized by model family, experiment outputs are grouped under docs/, and analysis scripts point to stable project-relative paths.

Project Snapshot

• Problem: predict final burned-area masks from pre-fire observations.
• Core task: binary semantic segmentation of burned area.
• Main comparison: Sentinel-only baseline vs multimodal fusion model.
• Engineering focus: reproducible experimentation on limited hardware.
• Context: academic computer vision and geospatial AI project.

Key Outcomes

• Multimodal model validation performance at best threshold (0.95):
  • IoU: 0.4026
  • F1: 0.5741
  • Precision: 0.5761
  • Recall: 0.5722
• Baseline validation performance at best threshold (0.95):
  • IoU: 0.1515
  • F1: 0.2632
• Auxiliary-task case study:
  • Mean delta IoU vs previous model: +0.0514
  • Win rate across samples: 30.7%
• Modality ablation:
  • Full multimodal setup clearly outperforms Sentinel-only and most ablated variants.

Source artifacts:

• docs/experiments/model-comparison/comparison_metrics.csv
• docs/experiments/modality-ablation/ablation_results.csv
• docs/case-studies/auxiliary-impact/report/summary.md

Model Overview

Baseline

• Architecture: U-Net style burned-area segmentation
• Input: Sentinel-2 only
• Encoder: ResNet-34
• Purpose: strong reference model for comparison

Multimodal Model

• Architecture: MultiModalFPN
• Encoder: EfficientNet-B4
• Inputs:
  • Sentinel-2 imagery
  • Landsat imagery
  • DEM and infrastructure rasters
  • ERA5 raster and tabular weather features
  • ignition priors
• Purpose: improve burned-area prediction through feature-level multimodal fusion

Multimodal + Auxiliary Variant

• Adds auxiliary land-cover supervision during training
• Used for qualitative case studies and post-hoc evaluation
• Helps analyze where extra semantic context improves prediction quality

Architecture

Repository Structure

WildFire/
├── README.md
├── requirements.txt
├── data/
│   ├── fire_*/
│   └── geojson/
├── docs/
│   ├── architecture/
│   ├── exploratory-analysis/
│   ├── experiments/
│   │   ├── modality-ablation/
│   │   └── model-comparison/
│   ├── case-studies/
│   │   ├── auxiliary-impact/
│   │   ├── inference-report/
│   │   ├── landcover-demo/
│   │   └── postprocess/
│   ├── figures/
│   │   └── slide-deck/
│   └── summaries/
├── inference/
│   ├── compare_baseline_vs_multimodal.py
│   ├── deploy_inference.py
│   ├── inference_2.py
│   └── inference_map.py
└── src/
    ├── analysis/
    ├── baseline_singlemodal/
    ├── multimodal/
    └── multimodal_auxiliary/

Getting Started

1. Install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Expected data layout

The repository expects:

• fire samples inside data/fire_*
• vector metadata inside data/geojson/

Large checkpoints and generated outputs are intentionally ignored by Git.

3. Run core workflows

Train the multimodal model:

python src/multimodal/main.py

Train the baseline model:

python src/baseline_singlemodal/main.py

Generate comparison artifacts:

python inference/compare_baseline_vs_multimodal.py
python src/analysis/modality_ablation_quick.py
python src/multimodal_auxiliary/inference_report.py

Generate qualitative figures:

python inference/deploy_inference.py
python src/analysis/make_qualitative_panels.py
python src/analysis/plot_threshold_sweep.py
python src/analysis/export_slide_table.py

Documentation Map

• Exploratory figures: docs/exploratory-analysis/
• Model comparison tables and charts: docs/experiments/model-comparison/
• Modality ablation results: docs/experiments/modality-ablation/
• Auxiliary model analysis: docs/case-studies/auxiliary-impact/
• Validation diagnostics: docs/case-studies/inference-report/
• Presentation-ready figures: docs/figures/slide-deck/

Why This Repo Is Worth Reading

• It tackles a real geospatial prediction problem with a measurable baseline-to-multimodal uplift.
• It combines remote sensing, computer vision, segmentation, and multimodal fusion in one project.
• It includes ablation studies, qualitative analysis, and diagnostic reporting rather than only a single training script.
• It has been reorganized so readers can move from problem statement to model code to experiment evidence quickly.

Author

Yousef Fayyaz

Keywords

wildfire prediction remote sensing geospatial ai semantic segmentation multimodal learning computer vision pytorch earth observation

Hashtags

#wildfire #remote-sensing #geospatial-ai #semantic-segmentation #multimodal-learning #computer-vision #pytorch #earth-observation