When the stubble burns: the fires on Romania's fields, day by day

How it was counted

The data come from the country archives of NASA’s FIRMS service: every thermal detection by the VIIRS instrument on the Suomi NPP satellite, at a resolution of 375 m, from 23 January 2012 to 30 December 2025. Within Romania’s administrative boundaries there are 200,961 records, each with its date, time, position, radiated power and type.

What was counted

Detections that NASA types as a vegetation fire (type 0) were kept. Version 2 of the data, the one in the country archives on the day of download, has a known defect: it sometimes leaves fixed sources labelled as vegetation fires (the product guide, section 6.1). So one more rule was applied: any 400 m square with detections on at least 8 different days in at least 4 different years was removed whole. The rule removes 127 squares with 18,037 detections and catches 99.8% of those NASA marks as fixed sources. It also removes 1,910 detections labelled as vegetation fires, 245 of them on arable land; if some of those were vegetation fires repeated in one place, they were lost. 182,545 detections remain.

Each detection was assigned a land type, from Corine Land Cover 2018 (a European land-use map, at 100 m), and its commune or town. “Farmland” here means arable land, mixed farmland, pasture and grassland, vineyards and orchards: 14.1 million hectares, with 151,479 detections (83% of the total). “Arable land” means arable land proper (Corine classes 211–213): 8.7 million hectares, with 112,922 detections.

Natural grassland (Corine class 321) is counted here as farmland, beside pasture (231), although Corine files it under semi-natural areas. It has 3,297 detections, 2.2% of the 151,479, and 72% of them fall in spring. The count on arable land, with its fall of 78%, does not include them. Without them, 1,566 of the 3,158 communes and towns ranked within their county keep their place, 2,741 move by at most two places, and first place changes in 3 counties.

The Corine map has a 100 m grid, but the smallest patch drawn on it is 25 hectares, larger than the square of one detection: a small field among pastures can be mapped as pasture, and the reverse. The commune comes from administrative boundaries, not from Corine. The fall appears with both maps tried and with no map. Counted per day with satellite data, what remained after the harvest is 22% on Corine arable land, 25% on all farmland, 24% on cropland in a different map made from different images (Copernicus Global Land Cover, described below; there without the fixed-source rule), and 29% for all retained detections with no land map at all. The map dates from 2018 and is the same for every year: land that changed use is miscounted, but the fall is there without any map.

The harvest season is counted from 1 July to 31 October and spring from 1 February to 30 April. A day counts as having no data when the archive has no detection in any of the seven countries, between April and October: 32 days, 15 of them in summer 2022. NASA says the satellite’s data were lost between 26 July and 20 August 2022 (the notice); in the archive, detections return from 11 August.

Rain and cloud

For each country, the 15 grid squares of half a degree of latitude and longitude with most detections on cropland from July to October in 2012–2017 were chosen. Daily rain and cloud for the centre of each come from ERA5, a reconstruction of past weather from measurements and a weather model (through Open-Meteo). In one square, a day is “dry and clear” when under 1 mm of rain fell on that day and on each of the two before it, and mean cloud cover was under 50%. A season’s number of dry, clear days is the average over the 15 squares, each weighted by its number of detections in 2012–2017; days without data are not counted. The rate is all the season’s detections in the whole country (not only those on dry, clear days) divided by that number of days; it is a weather-adjusted rate: for Romania, on Corine arable land, 151 a day in 2012–2017 and 32 in 2022–2025. In the comparison with the neighbours, detections on cropland in the shared map are divided by the same kind of day count. The thresholds were fixed before the count. The adjustment allows for rain and cloud, not for wind, temperature or the weather elsewhere.

The neighbours

Corine does not cover Ukraine or the Republic of Moldova, so the comparison between countries uses one map for all: Copernicus Global Land Cover at 100 m, class “cropland”. It also keeps every type-0 detection in each country’s file, without the fixed-source rule and without filtering detections against local administrative boundaries. That is why Romania’s numbers in the neighbours table differ from the arable-land numbers. Of Ukraine, only the part west of 32° east was counted.

The year of the fall

For each year from 2014 to 2024 the series was divided in two, before and after that year, and the share of its year-to-year variation removed by the division was computed. The series is the logarithm of detections per dry, clear day, on cropland; the logarithm makes the test compare proportional changes. The division at 2022 explains 84%, the most of the 11 years. With 2022 left out, the best division is at 2023, the first year left after 2021.

Traces in the fields: the check with Sentinel-2

The fire service sees no fall, and the satellite with the heat sensor passes only a few times a day. To find out whether less land was burned or only at other hours, the traces were counted: the black patches left on fields, which show in Sentinel-2 images (10–20 m, a pass every few days) whatever the hour of the fire.

60 one-kilometre squares were chosen at random, centred on arable land, in the eight counties with most detections after harvest in 2012–2017 (Dolj, Teleorman, Olt, Mehedinți, Călărași, Constanța, Giurgiu, Timiș). For 2018, 2019 and 2021, then 2023, 2024 and 2025, every pair of consecutive clear images between 25 June and 5 November, at most 20 days apart, was compared. A 20 m point counts as burnt when it darkens abruptly in the infrared and takes on the signature of ash: as bright at 2.2 microns as at 1.6.

The rule and the threshold for the verdict were written down before the result. The first rule, without the ash signature, failed: it also marked freshly ploughed soil as burnt. The second was chosen from spectra measured in 22 windows around detections and 5 windows of ploughed soil; those windows touch 4 of the 60 squares.

The second processing

The first processing of the images, the one first published, had faults that an outside review found. The cloud mask was used only to accept or reject a whole image, not point by point. For 2018–2021 the image catalogue holds two versions of most images (the processing of the time and one redone in 2023–2024), and the script took whichever it met first; it also read at most a hundred records a season, so in the years with two versions the start of the season was sometimes missing. The correction of 1,000 units that newer images need was decided from a catalogue label known to be wrong at times: 6 images were corrected twice.

All 60 squares were read again under rules written first: one image tile per square, the newest version of each image, the correction decided from the image’s own values (none of the 11,382 images read needs it), a cloud and shadow mask point by point at both ends of each pair, and the arable map brought onto the image grid. After the check below showed a control field marked because of thin cloud the mask did not see, a cloud filter on the blue band was added, before any result for the squares. The filter removed 520 of the images read, 262 of them from 2025.

The result of the second processing: 51.4 blackened hectares per thousand hectares of arable land per 100 days of observation in 2018, 2019 and 2021, 19.3 in 2023–2025. 38% remained, with a 95% interval from 21% to 62% when the squares are resampled and from 18% to 69% when the years are resampled too. The intervals do not include the rule’s own errors. By year: 60.0, 41.4 and 53.6, then 19.1, 21.3 and 17.3; 2020, measured afterwards, has 53.3. The first processing had given 54.2 and 20.0, that is 37%.

None of the corrections moves the result: with or without the point-by-point mask, with or without the cloud filter, with the catalogue label or the image’s values, with the old arable map or the new, between 36% and 38% remained. Without the 4 squares touched by the choice of rule: 37% (from 19% to 65%). With the first rule, the one that also takes ploughed soil for a burn: 50% (from 36% to 68%). The area is concentrated in few places: five squares hold 46% of everything measured in the first years, and without them 47% remained (from 27% to 75%). In 42 squares the area fell, in 15 it rose. Two of the squares, drawn at random, lie less than 300 metres apart and largely overlap, so they are not independent; without either one the result is the same.

How often the rule is wrong

The check on detections, read again in the same way (there a pair of images may be more than 20 days apart; one case has 25): of 54 detections on arable land chosen at random, with clear images before and after, 39 have a marked patch of at least two hectares around them (19 of 27 in 2018–2021, 20 of 27 in 2022–2025); of the 26 detections with the highest power, 23. But on 79 control fields, points on arable land 5–40 km from a detection with no detection within 3 km in those days, the rule marks a few points in 6 and a patch of at least two hectares in 2 (in the first processing, in none). Looked at, the two are a ploughed field and an image under thin cloud. The absence of a detection does not prove that nothing burned on the others.

The rule written at the start had a third clause: if wrong marks alone could produce more than a third of what was measured in the first years, the verdict is “cannot tell”. The control fields cannot settle this clause. They are windows larger than the squares, read without the arable map and with a single pair of images each, and most of the area marked on them comes from 2020, a year outside the comparison. An earlier version of this page gave an estimate here (73%); it mixed different years and different ways of measuring, and has been withdrawn. The control fields show that the rule makes mistakes, not how often.

The verdict does not depend on that clause. The rule required the upper end of the interval to be below 60%; it is 62%, just over the line, and 69% when the years are resampled too. The verdict of the second processing is “cannot tell”. The marked area fell; whether, and by how much, the area truly burnt fell, the test cannot establish.

The images were then looked at, one by one. The three largest blackenings in the sample (between 54 and 99 hectares) look like burns: straw-coloured stubble gone black within two to five days, in one case with streaks left by the wind. Of 12 cases of at least two hectares drawn at random, six from each period, 4 look like burns (2 before, 2 after), 5 look like ploughing, cloud or shadow (2 before, 3 after), and 3 cannot be judged. The judgement was made from images, not on the ground, and was not repeated by a second reader.

What is left: of the 146 blackenings of at least two hectares, those of more than 20 hectares at once number 12 in 2018, 2019 and 2021 and 1 in 2023–2025. This count was made after the results were seen; it is not a test fixed beforehand. It too rests on the same rule. It points the same way as the detections; by how much the burnt area fell, these images cannot say.

Other weaknesses. The squares are few, and three years on each side are not enough to measure the variation between years. The set of ploughed fields on which the rule was chosen is small and only from June–August. The arable map is from 2018. Cloud shadow that the mask does not see is not filtered. The comparison with detections is not quite like for like: detections cover whole counties, from 1 July to 31 October.

What the count cannot say

A detection records heat within one square of ground. It does not show who lit the fire and does not prove an offence. Commune totals do not show individual plots.

The count does not catch everything. In summer the satellite passes at about 2 pm and about 4 am; 16% of detections are at night. A fire lit in the evening and out by night can be missed.

The land map is from 2018. A field since left unworked or built on is still counted as arable.

The 3,186 units are 2,862 communes, 216 towns, 102 cities and the six districts of Bucharest. For each, 2012–2017 is compared with 2022–2025. The first version of this page gave the labels “fewer” and “more frequent” from a test on detections, with no correction for the thousands of comparisons: 1,112 falls and 81 rises. That test was too lenient, because several detections can come from one fire. The present rule was written down before its result. It counts days with fire (one commune, one day, at least one detection: 83,199 across 2012–2025). It allows for their varying from year to year more than a purely random count would (the variance is 1.3–2.0 times larger). Then it corrects for the 2,426 comparisons made at once (the Benjamini–Hochberg method, 5%). 147 falls remain and no rise: none of the 81 rises passed the test. For 2,279 communes and towns the test gives no firm conclusion, which does not mean that nothing changed; 760 have fewer than five days with fire in the periods compared, and 114 of those have no detection at all. The fall is clear for counties and for the country, where the numbers are large; for a single commune, most often, it is not. Days with fire in a row can belong to one fire, and neighbouring communes can share a fire: the labels are flags, and the 5% level is approximate, not guaranteed.

The fall coincides with the year the fines were raised, and falls as large exist in countries with other laws. How much of it is due to the fines, the subsidies, satellite monitoring, the price of straw or habit, the data do not show.

The law, step by step

Fines are for an individual unless stated, in Romanian lei. For the fine raised in 2022, the date in the table is that of publication in the Official Gazette; the date from which it could first be applied was not verified.

The neighbours in numbers

Detections on cropland from July to October, divided by the season’s number of dry, clear days. The last column is the year at which a single division best describes the country’s series.

The fire service and the satellite, year by year

The call-outs are those in the inspectorate’s annual reports, under “dry vegetation and other fires”, a heading that also covers rubbish fires and whose make-up changed in 2022. Between 2012 and 2021 their total rises and falls with all retained detections (correlation 0.74). In 2012–2017 there were 1.4 detections for every call-out; in 2022–2025, 0.4. Call-outs to stubble fires are 6.8% of the total.

The counties

Detections on farmland over the fourteen years, then how many that makes per year per 1,000 hectares of farmland. The last two columns: the yearly average of detections on arable land after the harvest, in 2012–2017 and in 2022–2025.

The communes and towns with most detections

The first 50 by number of detections on farmland. The middle columns are the totals for 2012–2017 (six years) and 2022–2025 (four years). The last column: detections per year per 1,000 hectares of farmland.

The data, to download

The tables the page is made from, in three CSV files and one JSON file: detections by day (farmland and arable land, with the days without data marked), the counties by year, the communes and towns by year (with days with fire and the label) and a sheet with the rules, the sources and the version. Column names are in Romanian. The raw detections can be downloaded from NASA, at the address in the sources.

What else exists on the subject

On vegetation fires in Romania there is a series of reports with the fire service’s counts by county (PressOne, “România arde”, 2023, in Romanian) and a conference paper that combines NASA’s fire data with the Corine map for 2001–2022 (Mattes, Marzolff and Feurdean, 2024) and finds a fall in fires on arable land after 2007. For Ukraine there is a measurement on arable land (Hall and colleagues, 2021) and a journalistic map (Texty).

None of them gives the number of detections by commune, a comparison of Romania with its neighbours around a change of law, or the series of detections after 2022.

Where Romania burned shows something else: the burnt perimeters mapped by the European service EFFIS between 2016 and 2026. The largest are in the Danube Delta; Tulcea county has 58.6% of the mapped burnt area. That page has the perimeters of large fires, with their area; this one, the daily detections in the fields.

Sources

• Detections: NASA FIRMS country archives, VIIRS S-NPP 375 m (product VNP14IMG, collection 2) and MODIS (MCD14ML), firms.modaps.eosdis.nasa.gov/country, downloaded on 11 October 2026. In the wording NASA asks for: “We acknowledge the use of data and/or imagery from NASA’s Fire Information for Resource Management System (FIRMS) (https://earthdata.nasa.gov/firms), part of NASA’s Earth Science Data and Information System (ESDIS).”
• Land: Corine Land Cover 2018, European Environment Agency, Copernicus Land Monitoring Service; Copernicus Global Land Cover 100 m, 2019 (Buchhorn and colleagues, CC BY 4.0).
• Weather: ERA5 (Copernicus/ECMWF), through Open-Meteo, CC BY 4.0.
• Images: Sentinel-2, Copernicus data, read from the Earth Search catalogue.
• Commune boundaries: geo-spatial.org, based on data from ANCPI, Romania’s land registry agency, CC BY-SA 4.0, version of 26 March 2025. Borders on the regional map: Natural Earth.
• Fire service: the General Inspectorate for Emergency Situations (IGSU), its yearly statistical analysis of emergency response, one report for each year from 2012 to 2025 (in Romanian): 2012, 2013, 2014, 2015, 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025.
• Laws (in Romanian unless noted): Emergency Ordinance 195/2005, Law 265/2006, Emergency Ordinance 38/2022 (Official Gazette 344 of 7 April 2022), Order 30/147/2010, Order 54/570/32/2023; the rule on fires against the farmer’s will: APIA Bacău’s statement of 12 March 2025; Ukraine: Law 556-IX of 13 April 2020 (in Ukrainian); Bulgaria: the Agricultural Land Protection Act, art. 41a (in Bulgarian).