NDWS Wildfire Spread Prediction

Upload .npy/.npz shaped [12|13, 64, 64] or [N, 12|13, 64, 64]. Channel order: PrevFireMask, elevation, NDVI, vs, th, pr, sph, tmmn, tmmx, pdsi, erc, population.

Sample or batch file

Threshold (τ)

Probability cut-off τ. A pixel is marked 'will burn' only above τ. Lower = more area flagged; higher = more conservative.

0.05 0.95

Ensemble Members

How many independent predictions to average. The model is stochastic, so averaging gives a smoother confidence map. More = steadier but slower.

1 10

DDPM Steps

Denoising steps per prediction. 1000 matches training and is most accurate; 50–200 is a quick preview.

50 1000

Input Channels

Prediction

Samples shipped with this Space (samples/) plus every .npy in amar-atoum/NDWS-Weights. Batch files get an index slider so you can walk through your whole test set.

Threshold (τ)

Probability cut-off τ. A pixel is marked 'will burn' only above τ. Lower = more area flagged; higher = more conservative.

0.05 0.95

Ensemble Members

How many independent predictions to average. The model is stochastic, so averaging gives a smoother confidence map. More = steadier but slower.

1 10

DDPM Steps

Denoising steps per prediction. 1000 matches training and is most accurate; 50–200 is a quick preview.

50 1000

Input Channels

Prediction vs. Ground Truth

Set each conditioning channel to a constant value to build a synthetic scenario. Good for sensitivity analysis.

Elevation (m)

0 4000

NDVI (vegetation index)

Fuel proxy. Higher = more to burn.

-1 1

Wind speed (m/s)

0 30

Wind direction (deg)

Bearing the wind blows toward (0=N, 90=E).

0 360

Precipitation (mm)

0 100

Specific humidity (kg/kg)

Drier air favours faster spread.

0 0.03

Min temperature (K)

250 310

Max temperature (K)

250 320

Drought (PDSI)

Negative = drier = easier spread.

-10 10

Energy release component

0 120

Population density

0 10000

Threshold (τ)

Probability cut-off τ. A pixel is marked 'will burn' only above τ. Lower = more area flagged; higher = more conservative.

0.05 0.95

Ensemble Members

How many independent predictions to average. The model is stochastic, so averaging gives a smoother confidence map. More = steadier but slower.

1 10

DDPM Steps

Denoising steps per prediction. 1000 matches training and is most accurate; 50–200 is a quick preview.

50 1000

Input Channels

Prediction

🔥 NDWS Wildfire Spread Prediction