Emberline

About, NASA Space Apps Challenge 2026

Built in 48 hours, for people who study fire.

Emberline answers the challenge “MODIS and VIIRS hot spots”: harmonize the satellite fire record to reduce the discontinuity between the MODIS and VIIRS records, and turn it into a burning activity calendar for any fitted country or drawn area.

Challenge brief → Emberline

Each ask in the official brief, where it is answered, and how completely.

#The brief asks forWhereStatus
R1Harmonize MODIS and VIIRS into one comparable seriesMethod: per-country detection model, tested on held-out yearsdoneregional fits; smaller countries borrow a pooled curve and the sparsest have no fit (counts on Method)
R2A burning activity calendar from active-fire hotspotsExplorer: month × year calendardone
R3Number of hotspots from the early 2000s to the presentExplorer: raw 2001–2025, harmonized 2003–2025partial2026 only as provisional raw counts in “This month so far”, for the countries loaded
R4Which months and years had the highest activityExplorer: calendar and busiest-year readoutdone
R5Current month against the historical medianExplorer and Seasons: “This month so far” for 166 of 171 fitted countries and any drawn area, against the median of the chosen yearsdoneprovisional, raw, near-real-time, no vegetation-type filter, platform by platform (NOAA-20 vs Suomi-NPP; drifting MODIS); never a harmonized anomaly
R6Historical patterns, unusual conditions, critical periodsSeasons (each month against its history), Trends (season timing)done
R7Select an area of interestExplorer: any country, a drawn box or polygon, or typed boundsdonean area resolves to the ~45 km grid cells whose centre lies inside it (shown, with the country files they come from)
R8Select a specific date rangeExplorer: start and end month; points download by daydonecalendar, totals, this-month median and every download follow the range
R9Downloadable CSV of hotspots from a searchExplorer: individual detections for the area and any dates (up to 31 days per file, archive or near-real-time), cell-level CSV for any month range, 7-day cellsdonepoints come from NASA FIRMS through this site's server; each file says points or cells, raw or harmonized
R10Accessible colour rampsEvery ramp and palette checked for colour-vision deficiency, lightness order and text contrast (scripts/check_color_access.py, docs/color-accessibility.md)done
R11Draw the area on a map or enter latitude and longitudeExplorer: draw a box or polygon on the map, or type the four boundsdone
R12Download analytics and visualizationsExplorer: calendar and chart as SVG or PNG, analytics JSON, calendar and cell CSV, HTML brief; open Parquet and JSON files on Datadone

The problem

From 2012 a sharper satellite sensor began detecting small fires the old one could not. Raw fire records jumped — by a median ×3.41 (2012–14 over 2009–11) across the 130 countries with their own fit — much of the apparent increase reflects improved detection sensitivity rather than an equivalent increase in fire activity.

What we built

A detection-probability model, fitted per country on 2012–2018 and tested on 2019–2021, that converts every VIIRS detection into what MODIS would have seen. Regional fits, no global step correction: 171 countries harmonized: 130 with their own fit, 41 with a pooled fit; 66 shown raw (too few detections to fit). Then a calendar, insights, a live globe and open data on top.

Who it is for

A regional analyst checks whether a past season was unusually active, compares its timing with earlier years, and exports comparable evidence as CSV. Historical analysis, not an operational fire alert.

Evidence

Every number on this site comes from our own fit of NASA FIRMS data, computed during the event: trained on 2012–2018 and tested on the held-out 2019–2021 years. The per-country accuracy, the step before and after correction and the validation design are on the Method page; the per-country tables and downloads are on the Data page.

Limitations

  • The fit uses 2012–2021 only. Terra and Aqua left their frozen orbits in 2022, so later years are a holdout and carry an orbital-drift caveat.
  • The value is regional, not global: the sensor break may be undetectable above natural variation at global scale.
  • Countries without enough data use a borrowed (pooled) curve, and cells without a published fit show raw sums flagged as uncorrected.
  • The live layer is NOAA-20/21 VIIRS, a different platform from the historical Suomi-NPP record. Historical analysis, not an operational fire alert.

Data & credits

MODIS Collection 6.1 active fires (Terra, Aqua)NASA FIRMS country archive, 2001–2025, and near-real-time feed · FIRMS download · country archive
VIIRS 375 m active fires (Suomi-NPP)NASA FIRMS country archive, 2012–2025 · FIRMS download
VIIRS 375 m active fires (NOAA-20)NASA FIRMS country archive (viirs-jpss1), 2018–2023: fits and tests the NOAA-20 to Suomi-NPP bridge · country archive
VIIRS 375 m (NOAA-20, NOAA-21)NASA FIRMS near-real-time feeds: live layer, and the current month · FIRMS active-fire data
MODIS-to-VIIRS transitionNASA LAADS overview of the sensor change · LAADS PDF
Orbital elements (TLEs)Public CelesTrak elements via tle.ivanstanojevic.me · CelesTrak · TLE API
Fire diurnal cycle parametersAndela et al. (2015), Atmos. Chem. Phys. 15, 8831–8846, Table 1 — illustration only · doi:10.5194/acp-15-8831-2015
MODIS FRP detection thresholds (9–11 MW)Laurent et al. (2019), Biogeosciences 16, 275–288, citing Roberts & Wooster (2008) and Schroeder et al. (2010) — “thresholds of FRP detection”, a sanity check only · Laurent et al. 2019
Current month to dateNASA FIRMS area API (MODIS_NRT, VIIRS_NOAA20_NRT), provisional; raw, or harmonized through the NOAA-20 bridge where it passed · FIRMS API
Satellite imagery (optional map layer)NASA GIBS / EOSDIS, NOAA-20 VIIRS corrected reflectance true colour · NASA GIBS
Wildfire events (optional globe layer)NASA EONET API v3, open events in the Wildfires category — curated, a different source from FIRMS · NASA EONET
Country outlines and areasNatural Earth 1:110m (public domain): hotspot and month-to-date membership, approximate land areas · naturalearthdata.com
BasemapOpenFreeMap / OpenMapTiles, © OpenStreetMap contributors · OpenFreeMap · OpenStreetMap

Use of AI and work history

Shishir Bhat and Ridhan Teli have worked on this problem for about a year. Every scientific choice — the method, what to fit, the validation design and how to interpret the results — was made by the team. We used Claude, by Anthropic, through Claude Code to build it under our direction: during the event it wrote the code, ran the data pipeline and the fit we specified, tested the results and drafted documentation; before the event it also assisted with literature research and planning. Before the event we prepared planning and research documents, verified data schemas, the FIRMS ingest package and download scripts, and an empty app scaffold; these were imported into this repository in its first commit at the start of the event. The harmonization, the fit, the validation and the web app were built during the event, 14–15 Nov, as the git history shows. No AI-generated images, video or audio are used.

Team Modulark

Shishir Bhat and Ridhan Teli.

A small lab in Tumkur, Karnataka. Emberline is open source under the MIT license.

Not affiliated with or endorsed by NASA. No NASA logos or insignia are used.