🔥 Wildfire Spread Prediction

Visualizations

• Pre-Fire Input — false-color composite of the input image using SWIR / NIR / Red bands. Healthy vegetation appears green; bare soil pinkish-red; water blue.
• Model Prediction — the predicted post-fire scene rendered in the same SWIR/NIR/Red space. Look for darkening or color shifts in vegetated areas (signal of burn).
• NBR Map — Normalized Burn Ratio derived from SWIR and NIR. On the upload tab this shows predicted NBR change (red = potential burn). On the test tab this shows the GT-vs-prediction NBR error (red = model under-predicted burn, blue = over-predicted).

Metrics (test-sample tab only — needs ground truth)

• SSIM (Structural Similarity, 0–1) — measures perceptual structural similarity. > 0.85 is strong, > 0.90 is excellent. Penalizes blur and structural distortion more than pixel error.
• Pixel MSE — mean squared error in [-1, 1] space. Lower is better. < 0.01 indicates the prediction is very close pixel-wise.
• PSNR (Peak Signal-to-Noise Ratio, dB) — log-scale pixel fidelity. > 25 dB is good, > 30 dB is very high quality.

Why three metrics? Generative models can produce visually plausible outputs that score poorly on pixel MSE (because they invent plausible but slightly-different details). SSIM and PSNR together give a fairer picture of how well the model captured the actual fire event.