Method

How it works

The calendar shows burning activity: how much fire NASA satellites detected in an area, measured the same way from 2000 to today. It is not a count of individual fires and not burned area.

1. The problem: two sensors that don't agree

NASA detects active fires from space with two instruments. MODIS (on Terra and Aqua) has watched since 2000 with 1 km pixels. VIIRS (on Suomi NPP, from 2012) has 375 m pixels, so it sees more and smaller fires. Over Ghana, VIIRS records about 3.9 times as many hotspots as MODIS for the same months. Joined naively, the record shows a huge jump in 2012 that has nothing to do with burning. MODIS is also ageing and will end, so a long record must be carried forward by VIIRS.

2. What we do

1. Same quality rules. Keep only vegetation fires (no volcanoes, gas flares or factories) with nominal or high confidence. MODIS confidence 0–100 is mapped to low / nominal / high the way FIRMS recommends.
2. Same grid. Each hotspot is placed in a ° grid cell (about km). We count active cell-days: a cell with at least one fire on a day. Several VIIRS pixels on one fire collapse into one cell, which removes most of the resolution difference.
3. Measure the remaining gap. From , both sensors were flying. One MODIS cell-day equals on average VIIRS cell-days, but the ratio depends on the season (table below): in the main fire season fires are big and both sensors see them; outside it, small fires only VIIRS can see.
4. One record. Before VIIRS started () the calendar uses MODIS × the factor for that month. From then on it uses VIIRS directly. Units: VIIRS-equivalent active cell-days. Before Aqua's fire data begins () only Terra was flying, so a separate Terra-only factor is used and those months are shown faded as lower confidence.

3. How we chose the method, and how accurate it is

We tried 4 grid sizes and 4 ways of setting the factor (one for all of Ghana, one per month, one per zone, one per zone and month). Each combination was tested by leaving out one year at a time, fitting on the rest, and predicting the missing year's VIIRS activity from MODIS. The winner is the simplest method within half a point of the best: on the ° grid.

• Average error in a year's total:
• Typical error for a single fire-season month:
• Terra-only period (2000–02), yearly error:

4. Normals, unusual months and seasons

• Normal: the average for that month over the reference years (all full years by default, or the last 15 years).
• Usual range: the middle 80% of those years (10th to 90th percentile). A month outside it is flagged unusually high ▲ or unusually low ▼, but only in months that normally hold at least 1% of the year's burning, so a single fire in the rainy season doesn't raise an alarm.
• Fire season: from the date 10% of a normal year's burning has happened to the date 90% has. The "half of burning" window is 25% to 75%. Years start at the quietest point of the year, so a dry season is never split across two rows.
• Last 30 days: compared with the same 30 days in every reference year.

5. Outlook: the months ahead

For any area and any period in the next 12 months, the app gives the burning it expects, a likely range, and whether burning is rising or falling. It uses the harmonized record only.

• Forecast: for each month, half of the same month's average over the last 5 years, plus half of the trend line through the last 15 season totals, spread over the year with the usual monthly pattern of the last 10 years. The recent average follows where burning is now; the trend line follows where it is heading.
• How we chose it: we made forecasts from every month between 2011 and 2025, for Ghana and each of its 16 regions, using only the data available at the time, and compared them with what happened. The typical error on a 12-month total was about 10%, against about 15% for the long-term normal. Recent averages alone, the trend alone, and adjusting for the last few months were all tried and did no better.
• Likely range: from the same test. We take the past forecasts made at this time of year (within a month either side) for the same months ahead, and see how far the truth landed from them. The range covers the middle 80% of those misses.
• Trend: the median-of-slopes line through season totals (robust to one odd season), with a Mann-Kendall test. A trend is called "falling" or "rising" only when it is clear (p < 0.05); otherwise it is reported as no clear trend.
• What it can't see: the forecast knows only past fires. A very wet or very dry year, a burning ban or land-use change will move burning away from it. In 2025/26, the lowest season on record, actual burning fell well below the range forecast for it.

6. Limits

• Satellites pass about twice a day. Fires that start and end between passes, or burn under thick cloud or smoke, are missed. Single days are noisy; weeks and months are reliable.
• Off-season months before 2012 multiply small MODIS counts by large factors, so they are less certain. They hold under 5% of a year's burning.
• June and July have too little MODIS activity for their own factor and use the overall one.
• The latest weeks use near-real-time data, which is replaced by quality-checked data a few months later.

7. Data

NASA FIRMS: MODIS Collection 6.1 and VIIRS S-NPP 375 m active fire products, standard (archive) and near-real-time. We acknowledge the use of data and imagery from NASA's Fire Information for Resource Management System (FIRMS), part of NASA's Earth Science Data and Information System (ESDIS). Ghana boundaries from geoBoundaries. Data currently runs to .