The fire has passed. Now comes the harder question: exactly what burned, and where?

Why 2025 changed how we think about wildfire

A landmark review published in Nature Reviews Earth & Environmentin June 2026 summarised the 2025 global fire season with a finding that should restructure how every government, insurer, and land manager thinks about fire risk. Global burned area in 2025 was 335 million hectares — 16% below the long-term average, the second lowest since 2002. Global fire-related CO₂ emissions were the third lowest on record.

And yet 2025 was the most expensive wildfire year in history. Fires accounted for 38% of all insured natural hazard losses globally. Over 300,000 people were evacuated. More than 90 people died. Europe experienced its first-ever wildfire classified as a "megafire" — exceeding 10,000 hectares.

Sources: van der Werf et al. (2026), Nature Reviews Earth & Environment; UNDRR (2026); University of East Anglia wildfire review 2026.

The satellite science of burn scars

When vegetation burns, it undergoes a transformation that is visible from space — not in visible light alone, but in the spectral signatures that carry far more diagnostic information. Healthy green vegetation strongly reflects near-infrared (NIR) radiation and absorbs shortwave infrared (SWIR). (For a deep dive into vegetation reflectance dynamics and Sentinel-2 band processing, see our companion guide on building an automated NDVI alert pipeline).

Burned vegetation does the opposite: charred plant matter and exposed soil dramatically increase SWIR reflectance while NIR drops. This spectral contrast is the physical basis of every satellite-based burn scar detection method.

Spectral indices: NBR, BAIS2, and RBR

Three indices are workhorses of operational burn scar mapping, each capturing a different aspect of the fire-affected landscape:

Why Sentinel-2 and Landsat are the right instruments

Sentinel-2 provides 10m and 20m multispectral imagery with a 5-day revisit cycle, making it the highest-resolution freely available satellite appropriate for burn scar mapping. Its 13 spectral bands include the Red Edge region (Bands 5, 6, 7) that enables BAIS2 computation, and two SWIR bands (B11, B12) that are critical for fire severity assessment. A 2025 study of the Los Angeles wildfires confirmed that Sentinel-2 achieved 88% overall accuracy in burned area classification.

Landsat 8 and 9, operated by USGS and NASA, provide 30m imagery with a 16-day revisit. Landsat’s longer archive — extending back to 1972 — makes it irreplaceable for change detection and historical baseline analysis. (See our research on Landsat TIRS land surface temperature retrieval for deeper thermal calibration methods).

Atmospheric Radiative Transfer: BOA (L2A) vs. TOA (L1C)

A fatal mistake in operational burn severity pipelines is calculating NBR directly on Top-of-Atmosphere (Level-1C) reflectance. Rayleigh scattering in the atmosphere preferentially scatters shorter visible wavelengths, while aerosol optical depth from residual post-fire smoke haze and atmospheric water vapor absorption bands (940 nm and 1130 nm) distort SWIR transmission:

Sentinel-2 Level-2A products package a dedicated 20m Scene Classification Layer (SCL). Robust burn scar detection algorithms must mask out invalid pixels before difference modeling:

From index to classified map: object-based vs pixel-based classification

Applying spectral indices to satellite imagery produces a raster of continuous values. Converting these values to a classified burn scar map requires a threshold decision or a machine learning classifier. Pixel-based thresholding is fast and interpretable but produces characteristic "salt-and-pepper" noise at fire edges and within heterogeneous burned patches.

A December 2025 study found that object-based classification using SNIC (Simple Non-Iterative Clustering) segmentation before machine learning classification markedly improved boundary delineation accuracy, reducing fragmentation artefacts and producing more reliable burned area estimates without requiring labelled training data. When training GeoAI models for wildfire boundary detection, accounting for spatial autocorrelation is critical to avoid evaluation bias — as explored in our guide on spatial data leakage in machine learning.

NASA FIRMS: real-time active fire intelligence

Burn scar mapping answers the question of what burned. But emergency responders, fire managers, and evacuation planners need to know what is burning right now. NASA’s Fire Information for Resource Management System (FIRMS) provides near-real-time active fire detections from four satellite sensors.

Introducing GeoPlatform Wildfire: The Automated Satellite Intelligence Suite

Understanding the science is one thing. Having operational access to it — without configuring a Google Earth Engine account, managing Sentinel Hub credentials, or writing custom Python pipelines — is another. Infryne TechWorks has engineered GeoPlatform Wildfire, a production-grade Earth Observation suite decoupled into a high-performance FastAPI GeoAPI backend and an interactive Next.js 15 MapLibre GL web client.

What the platform does, step by step

Technical architecture: what runs under the hood

The platform employs a decoupled, cloud-native architecture optimized for zero-footprint raster processing:

• Cloud-Native GeoAPI Backend: Built on Python, FastAPI, and Rasterio, the /v1 API handles STAC discovery, remote COG window reads via WarpedVRT, and vectorized NumPy spectral index math without downloading full multi-gigabyte satellite scenes.
• STAC Multi-Granule Mosaicking:Analyzes all intersecting granule footprints from AWS Earth Search STAC for the selected acquisition date, dynamically filling spatial gaps and enforcing a >=98% union coverage threshold.
• 4-Band Display COG Delivery: Produces a streamable Cloud-Optimized GeoTIFF combining Bands 1–3 (True-Color RGB) and Band 4 (USGS burn classification mask), served with HTTP range request support and optional AWS S3 presigned durable storage.
• Cartographic PDF Report Generation: An integrated ReportLab engine that reverse-geocodes incident coordinates, compiles high-resolution cartographic map viewports, and computes exact geodesic area statistics (hectares, km², acres) across all USGS severity classes.
• Enterprise Security & Observability: Hardened with X-API-Key authentication, OWASP security headers, client rate limiting, payload size verification, and sanitized production error logging.

Forest fire in South Asia: a growing threat demanding better tools

While the Los Angeles fires dominated global attention, South Asia faces its own escalating wildfire crisis. A 2025 study tracking burn scar extent across Punjab, India, using Sentinel-2 data found that fire monitoring systems relying solely on MODIS or VIIRS active fire detections systematically underestimate burned area in fragmented agricultural and mixed forest-agricultural landscapes — precisely the landscape type that dominates the Pakistan-India borderland regions where agricultural burning is both widespread and poorly monitored.

Pakistan’s forests — particularly the coniferous forests of Khyber Pakhtunkhwa and Azad Kashmir, and the subtropical broadleaf forests of the Himalayan foothills — are increasingly vulnerable to wildfire as temperatures rise and pre-monsoon dry seasons extend. The Margalla Hills National Park adjacent to Islamabad experiences fires nearly every year in the April–May window. Ground-based monitoring of these events is limited by terrain and resources. High-resolution satellite burn scar mapping provides the only spatially complete record of fire extent, severity, and year-on-year progression.

Infryne TechWorks’ burn scar detection platform covers these geographies without restriction — any coordinate, any date within the Sentinel-2 and Landsat archives. For Pakistani forest managers, conservation organisations, or disaster response agencies, this means on-demand access to satellite fire intelligence that was previously available only to organisations with satellite data infrastructure and GIS expertise.

Production Python Implementation: Vectorized Spectral Pipeline

Below is the production NumPy/Rasterio implementation used inside Infryne TechWorks pipelines to ingest Sentinel-2 Level-2A COGs, apply SCL quality masks, calculate dNBR, RBR, and BAIS2, and assign USGS severity bins:

import numpy as np
import rasterio

def compute_burn_severity(
    pre_b08: np.ndarray,
    pre_b12: np.ndarray,
    post_b04: np.ndarray,
    post_b06: np.ndarray,
    post_b07: np.ndarray,
    post_b08: np.ndarray,
    post_b8a: np.ndarray,
    post_b12: np.ndarray,
    scl_mask: np.ndarray
) -> dict:
    """Computes cloud-masked NBR, dNBR, RBR, BAIS2, and USGS severity tiers."""
    # Convert BOA reflectances to float32 with valid scale factor (0.0001)
    eps = 1e-6
    
    # 1. Normalized Burn Ratio (NBR)
    with np.errstate(divide='ignore', invalid='ignore'):
        nbr_pre = (pre_b08 - pre_b12) / (pre_b08 + pre_b12 + eps)
        nbr_post = (post_b08 - post_b12) / (post_b08 + post_b12 + eps)
        
        # 2. Differential NBR (dNBR)
        dnbr = nbr_pre - nbr_post
        
        # 3. Relativised Burn Ratio (RBR) - adjusts for pre-fire vegetation density
        rbr = dnbr / (nbr_pre + 1.001)
        
        # 4. Burned Area Index for Sentinel-2 (BAIS2)
        re_ratio = np.sqrt(np.maximum(0, (post_b06 * post_b07 * post_b8a) / (post_b04 + eps)))
        swir_diff = (post_b12 - post_b8a) / np.sqrt(post_b12 + post_b8a + eps)
        bais2 = (1.0 - re_ratio) * swir_diff + 1.0

    # 5. Filter invalid pixels using SCL mask (mask out 0, 3, 8, 9, 10, 11)
    invalid_mask = np.isin(scl_mask, [0, 3, 8, 9, 10, 11])
    dnbr[invalid_mask] = np.nan
    rbr[invalid_mask] = np.nan
    bais2[invalid_mask] = np.nan

    # 6. Vectorized USGS Severity Classification (1 to 5)
    conditions = [
        dnbr < 0.10,                     # 1: Unburned / Regrowth
        (dnbr >= 0.10) & (dnbr < 0.27),  # 2: Low Severity
        (dnbr >= 0.27) & (dnbr < 0.44),  # 3: Moderate-Low Severity
        (dnbr >= 0.44) & (dnbr < 0.66),  # 4: Moderate-High Severity
        dnbr >= 0.66                     # 5: High Severity
    ]
    choices = [1, 2, 3, 4, 5]
    severity_map = np.select(conditions, choices, default=0).astype(np.uint8)
    severity_map[invalid_mask] = 0

    return {
        "dnbr": dnbr,
        "rbr": rbr,
        "bais2": bais2,
        "severity": severity_map
    }

Frequently Asked Questions