By Duncan Steel 31/08/2019


The numerous fires now burning in Brazil have been much-discussed of late, with world leaders complaining that the nation’s authorities allowing such clearing of land is highly detrimental to international efforts to limit the release of more carbon dioxide into the atmosphere, with the potential to exacerbate anthropogenic global warming/climate change. In this post I illustrate how such fires may be identified, and analyses conducted of the areas that have been (or are being) burnt, using freely-available satellite imagery collected at various wavelengths across the visible, near-infrared and shortwave-infrared parts of the spectrum. 

Over the past couple of weeks there has been intense media coverage of the widespread fires burning in Brazil. Last Friday (August 22nd) I was asked by the good people at the NZ Science Media Centre to provide comments on the utility of satellite imagery in scrutinizing what is going on in Amazonia, and elsewhere in Brazil, a request with which I was happy to comply. It took me fifteen minutes to find, download, clip and annotate the image below, showing just one of the areas currently burning.

False-colour (red-green-blue colour-coding of images collected in the red part of the visual spectrum plus two bands in the near-infrared) of a burnt and burning region of Mato Grosso province in Brazil. To find the area in an online satellite imagery source such as Google Maps, enter latitude 9.9 south, longitude 60.2 west (here is a link to make it easy). At the time of writing, the high-resolution satellite imagery (which is true-colour) in Google Maps still shows the area as mostly-forested.

In this blog I will present imagery of another specific fire location, not so far away from the above. In order to find the detailed area in question, again you could utilise Google Maps, and again I provide a clickable link so as to make it easy for you; close-enough, the region surrounds latitude 10.34 degrees south, longitude 59.60 degrees west.

My intent here is not to conduct any scientific analysis of the imagery, nor to provide any commentary on what is going on in Brazil (with similar burnings now underway in central Africa, and Indonesia, and elsewhere). My aim here is solely connected to the domain which I know a few things about: the satellite imagery that is available to all-comers, and how such imagery can quite-easily be manipulated so as to investigate the areas recently burnt, or now burning. That is, my intention here is largely educational, an attempt to show what can be done.

All the imagery that I use in this post was collected by one of the two European Sentinel-2 satellites (oh, plus a few screen grabs from Google Maps); in a later post I will look into complementary imagery available from the US Geological Survey’s Landsat-8 satellite. These Sentinel satellites are part of the European Copernicus Programme.

The spectral bands of the optical detector on board the Sentinel-2 satellites are as summarised in the table below; the GSD (Ground Sampling Distance) is essentially the pixel size in the images collected.

In terms of what our eyes can detect, only bands 1, 2, 3 and 4 are accessible. Band 1 is essentially violet light; band 2 is blue light; band 3 is green light; and band 4 is red light. Band 5 (at around 0.7 microns or 700 nm wavelength) is on the limit of what the human eye might pick up.

The reason for there being six bands (5, 6, 7, 8, 8A and 9) between 0.70 and 0.95 microns (in the near-infrared, NIR) is largely that vegetation is highly-reflective at such wavelengths. The reason that vegetation appears green to our eyes is that plants generally absorb much of the sunlight at the blue and red ends of the visible spectrum (powering photosynthesis), but not so much in the green, and so a greater proportion of the green sunlight is reflected. Once one moves into the NIR, however, although our eyes cannot detect such wavelengths in fact the NIR is very strongly reflected by vegetation. The balance between leaf reflectivities at different NIR wavelengths can enable much to be known about what plants are growing where, and also how well they are thriving. That’s why many satellite sensors have multiple bands across the NIR.

Returning the table above, the final three spectral bands are in the shortwave infrared (SWIR) part of the spectrum. The flux of sunlight is highest in the visual spectrum (between about 0.4 and 0.7 microns), dropping off in the NIR, and reducing still further in the SWIR. Nevertheless, reflectivities of the solar SWIR by the ground and vegetation can be useful in various scientific studies. In addition, items on the ground that are particularly hot will also emit a substantial flux of SWIR, and this can be detected by suitable satellite sensors. Considering band 12 of the Sentinel-2 MSI, at that wavelength (near 2.2 microns) an object at a temperature around 1000°C will be strongly-emitting; that is, Planck’s law indicates a peak in the emission flux at that wavelength (compared to the peak in the visual spectrum for sunlight, with the solar photosphere being at a temperature near 5,500°C).

What this means is that burning or smouldering material on the ground below might well be best-detected in the SWIR bands noted above, and band 12 in particular (notwithstanding the title of a famous book which suggests that paper combusts near a temperature of just 233°C).

Let me move now, then, to a region of Mato Grosso in the municipality of Aripuanã. Here is a Sentinel-2B ‘true-colour’ image that I built up using red=band 4, green=band 3, blue=band 2, and then fiddling a little with the brightness and contrast to make the overall frame look better. The black area to the right/east was outside of the instrument’s scan track. The data were collected on a satellite pass around 10:30 in the morning local time on August 16th.

True-colour image of part of the Mato Grosso obtained using the Sentinel-2B satellite on August 16th. The smoke from two main fires is obvious.

Most of the media coverage of the fires in Brazil have shown imagery similar to the above (i.e. true colour). Particular examples can be found, for example, in the Washington Post, and in the New York Times. (There is a heap of other pertinent information derived from satellite data in those articles, I should note.)

By using other wavelengths/other spectral bands, there is much else that we can discover. The composite image below I constructed from data collected using Sentinel-2B on the same pass as above. This time, though, I am presenting a false-colour image: the coding is such that blue=band 3, green=band 7 (NIR), and red=band 12 (SWIR). A number of changes are now clear, for example the Aripuanã river snaking down through the middle of the frame, and the dark burned areas to west and east.

False-colour image built up by mixing blue, green and red based on intensities in the green part of the visual spectrum, a near-infrared band, and a short-wave infrared band. The two main concentrations of smoke are still apparent, though not so obvious as in the preceding true-colour image: the scattering of electromagnetic radiation (light) is more efficient for shorter wavelengths (Rayleigh scattering by atmospheric molecules, Mie scattering by smoke particles if we assume them to be spherical) and that is why the smoke appears blue-ish in this image.

Next I present an enlarged and annotated view of part of the image above, including the two main burning areas in white boxes. It is that to the west/left that I will concentrate on below. Previously I indicated that this is around latitude 10.34 degrees south, longitude 59.60 degrees west.

Here is a true-colour zoom-in on that area at left: 

That is a R,G,B composite of bands 4,3,2 respectively. Now let us look at the individual bands as greyscales.

The above four images, then, are the result of data collection across the visual spectrum, from the violet through the blue, then green, then red. There are obvious differences between them. In the table above setting down the characteristics of the Sentinel-2 Multi-Spectral Instrument bands one might have wondered why band 1 has a pixel size of 60 metres, whereas the other three visual bands use pixels of 10 metres. The images above give an answer, in part: in band 1 (violet) the amount of atmospheric scattering (Rayleigh scattering, which varies as the inverse fourth power of the wavelength) naturally blurs any satellite image; there is no point in trying to collect imagery with high spatial resolution in the violet.

Quite apart from the alterations in brightness and darkness in the blue, green and red images above, the decreasing fuzziness as the wavelength increases is due to the reduction in atmospheric scattering. You can easily see this effect for yourself… When you look at distant hills (say the Southern Alps from Christchurch, looking across the Canterbury Plains) they appear blue, due to Rayleigh scattering, just as the sky is blue. If you try looking through a simple blue filter (such as used in the lighting at your local theatre) and then through a red filter, the clarity is far better through the latter. It’s obvious. Heck, in Christchurch you can just peer at the Port Hills through blue then red filters, and see the effect. Physics in action.

Now let’s move on to light outside of the visible spectrum, to bands 5 and beyond. I wrote previously that the bright-green-looking images above were composites – false-colour composites – made up using bands 3, 7 and 12. So, here are greyscale images of the detailed region of interest in bands 7 and 12 (we have already looked at band 3, above):

Quite different, are they not? — Both compared to each other, and also contrasting with the band 3 image. The smoke is obvious in band 3 (green wavelength detection), still apparent in band 7 (NIR), but really has little effect in band 12 (SWIR). In a previous post about the Nelson-Tasman bushfires in February I emphasized this fact: to a large extent, one can ‘see’ through the smoke (or thin clouds) at wavelengths somewhat longer than the visible spectrum.

In the middle of the band 12 frame, the line of white patches is due to thermal emission from actively burning areas, as mentioned earlier. Most of what we see in such images is just reflected sunlight, but if things are hot enough then there is emission occurring in the SWIR; and if hotter still, in the NIR and indeed in the visual spectrum (yes, you can see a fire burn – no surprises there). What should be obvious to the reader is that the SWIR is of pivotal importance if we want to trace the fire-fronts: where there is smoke there is fire, the adage goes, but where beneath the smoke is the fire? This band 12 SWIR image shows us where.

Next let’s look at a composite false-colour image made up from the above greyscale frames for bands 3, 7 and 12, just as we did previously but now zoomed in to show the burning area in more detail.

False-colour image of the burning area that is the focus herein, with blue in the frame coding for the intensity detected in band 3 (green wavelengths), green coding for band 7 (NIR), and red coding for band 12 (SWIR). The parts of this image appearing bright green are heavily-vegetated: leaves strongly-reflect NIR solar radiation. Where the plants have already been burned, the charcoal and bare soil absorb most of the solar radiation across the visible spectrum (as we all know) but also across the NIR and SWIR, and so those parts appear dark/black here. Small particles scatter short-wavelength radiation more effectively, and therefore the smoke plume appears blue here. The vegetation that was burning when this satellite pass occurred emits a substantial flux of radiation in the SWIR, but less in the NIR and less again in the visible spectrum, and so in this colour-coding the active fires appear orange-red; close examination of the mostly-dark area in the lower half of the frame will show that in fact there is an overall reddish hue that is uneven, and that will be due to smouldering materials such as tree trunks which have an elevated temperature and therefore significant emission in the SWIR/band 12.

How long have these fires been burning, so as to produce such a large area denuded of vegetation? The preceding frames result from data collected when Sentinel-2B made an overhead pass on August 16th. It happens that passes by one or other of the Sentinel-2 pair repeat every ten days (and in some locations interleaved or overlapping coverage occurs such that there is an A/B alternation every five days, though not in the specific location being examined here). The preceding pass by Sentinel-2B took place on August 6th; here is the equivalent image from that date, demonstrating that this fire is quite recent (i.e. it was started between August 6th and the 15th).

False-colour image of the same region as that in the preceding image, with the same coding as described above, but from a pass on August 6th (i.e. precisely ten days earlier). There is no sign of any fire or burnt area, giving us a limit on when the fire was initiated.

Our next step is to look in ever-more detail at the area that has been burnt/is still burning, as in the zoom-in frame below.

The burnt and burning area on August 16th, in more detail. The letters A to E have been added for ease of referring to specific parts of this frame.

In the above frame I have labelled five features, which I will now address in turn. Letter A simply indicates trees that are growing along a meandering creek. This may not be obvious from the above image, to readers not yet familiar with the ‘bright green means vegetation’ coding, and so below I include a screen grab from Google Maps that shows a higher-resolution true-colour satellite image, from which it should be obvious that feature A indeed comprises trees growing along a creek or rivulet.

Screen grab from Google Maps satellite imagery showing locations A and B from some earlier time, presumably within the past six months or so.

Feature B also appears in the above true-colour frame, which derives from imagery collected long before the fires burned through this area. Thus in my August 16 false-colour image made using Sentinel-2 data feature B appears dark (it’s been burnt), whereas in the earlier true colour imagery (from Google Maps) it is far brighter than the adjacent wooded area. Looking at the image directly above it would seem that location B was sparsely-covered with grass, with animal tracks (presumably cattle) crossing it. I point this out because many media reports claim that the fires in Brazil are generally “rainforest” being destroyed, and location B is not forest at all, nor has it been for some years.

Zooming out a little, here is the Google Maps satellite image with all five lettered locations indicated.

Feature C has been labelled simply as a reference point, a hook-shaped indentation of grassland intruding into the forest. Area D is clearly a larger grassland area, most of which was not seen to be burnt-off in the August 16 image, though some of it at its northwest extremity was. Contrariwise, the area just northeast of that letter D seems from the picture directly above to have been scrubby, with small trees and bushes interspersed with grass, and by August 16 this area had been burned through, resulting in the blackest segment of the false-colour August 16 composites; an interpretation of this would be that the fire had quickly consumed all combustable material and there were no thick tree-trunks left to smoulder, consistent with my thought that this was sparsely-covered scrubland before the conflagration.

On now to the tree-covered area E, lying about a kilometre west of feature C, as in the zoomed-in Google Maps image below.

Forested area E, around one kilometre west of feature C, appeared to be actively burning in the August 16 imagery.

In the false-colour Sentinel-2B images from August 16 (shown previously) there is a fire-front just to the north and east of E which might be interpreted as being a ground fire with a canopy fire not yet having consumed the trees in toto; note the greenness that persists around that letter E. To the west and northwest of E, though, the fire has progressed to a greater extent, with the previously-scrubby or less-densely-tree-covered parts already having burned through.

In this blog I have been looking at just one example area that is burning in Brazil, out of thousands of such fires. As previously noted, many media reports give the impression that what is happening is a mass burn-off of pristine rainforest. The present example indicates that in this specific case there has been (and likely continues to be) a burn-off of grassy fields previously cleared for agricultural purposes, plus a burn-off of scrubby areas perhaps to prepare those for future use in farming of some description, but also there are some densely-forested areas being burned down, presumably for similar purposes.

A final matter I wanted to investigate here was whether the evidenced burn-off is just a repeat of what has occurred in preceding years. That is, we are generally familiar in New Zealand with arable farmers burning the stubble across their acreages to prepare the land for ploughing and planting in the next season, or year; is that all that is being done here?

Although some satellites have been in orbit for far longer, the Sentinel-2 pair from which I have been using imagery herein have only been aloft since June 2015 (2A) and March 2017 (2B); thus the imagery archive only goes back a few years. The set of a dozen images below, each covering our area of interest, stretch back over three-and-a-bit years (i.e. four sets of data: 2016, 2017, 2018, 2019). The column of six images on the left are all from the past few months (June, July, August). In the right column the six images comprise two each from August and September in 2018, 2017, and 2016. Cloudy weather during passes in 2017 limited my choice of frames to use, neither of those shown being free of such interference.

Imagery of the region of interest in Brazil from 2019 (six frames in the left-hand column), and in 2018, 2017 and 2016 (two frames each in the right-hand column). These are false-colour images coded from frames collected at the red end of the visual spectrum and also near-infrared bands.

Examination of the above dozen images indicates that: (a) There is no sign of any burn-off occurring earlier than August 6 during this calendar year; and (b) Relatively small areas were burnt off at some stage between mid-August and early September in both 2018 (dark patch in northwest of the frame) and in 2016 (patch in the west of the frame), plus some smaller dark patches in 2017 somewhat later in September (in the northwest of the frame).

This has been a long post, with lots of images presented. The intent has been restricted to indicating the sorts of things that can be done with satellite imagery that is freely available to anyone, in this case limited to the European Copernicus Programme Sentinel-2 pair of satellites (plus some grabs of higher-resolution satellite imagery available through Google Maps). In a second post I will turn to data collected by the US Geological Survey’s Landsat-8 satellite, which has inferior spatial resolution (30-metre pixels) but has the advantage of a sensor working in the thermal-infrared part of the spectrum (around 10 microns wavelength) as well as an instrument working across the visible/NIR/SWIR (as with the Sentinel-2 MSI). In that post I will also consider other ways of analysing data and presenting results pertaining to the tracking of fires such as those now burning in Brazil and elsewhere.