What does Satellite Infrared data tell us about the evolving Russian Strategy in its Ukraine invasion?
The Ukraine-Russia war began with Russia’s coordinated invasion across multiple areas of Ukraine. Geared as an initial decapitation strike against the Ukrainian government, Russia found it had brought insufficient forces in the wake of unexpected Ukrainian resistance. The northern thrust of the war ended with Russia’s retreat and refocus on the Donbas region (Luhansk and Donetsk oblasts to Ukraine’s far east). Russia broke through Ukrainian lines at Popasna in late May/early June of 2022, but made limited gains after capturing key cities in July. In early September Ukraine counterattacked and took back portions of Kharkiv and Luhansk oblasts. Russia withdrew from Kherson city to the south after it’s ability to supply forces across the Dnipro were limited following the damage to at least one rail span of the Kerch bridge. As of this writing (early March 2023) the front has stayed relatively consistent in a line from Kupyansk — Svatove — Kremina — Bakhmut — Avdiivka — Vuhledar.
A key tool to visualize the conflict has been infrared satellites — namely VIIRS 375m, as it sits at a convenient spot between low latency (updated twice daily), relatively narrow spatial resolution, and good-enough quality. Since the war began a superior Landsat product was announced with an astonishing 30m resolution, but it misses the key first months of the war. Discussion here will resolve around VIIRS.
The use of infrared is complicated by a number of factors. The first is cloud cover. Clouds are composed of water vapor, and water can and does attenuate light, including infrared, as anyone who has enjoyed the shade on a warm summer day knows. Furthermore the satellites’ ability to detect fires is not universally consistent — it does better and worse sometimes. Lastly the ability of satellites to pick up infrared depends on material to burn — one can imagine a static front sees limits to how far fire may spread. Nonetheless the core discussion here is necessarily physics — what photons are emitted during conflict, and to what extent are those photon signals attenuated? If we can address this question, then infrared data can answer many more about the nature of modern artillery warfare.
If you are interested in the nitty gritty detail of what infrared wavelengths are emitted, how much it is attenuated by cloud cover, and what wavelengths artillery can be detected, you’ll probably want to read the full discussion. If you are only interested in the portions relevant to the Ukraine War, you can read the next section, skip the following two, and go straight to the conclusions. The tl;dr for anyone who doesn’t want to read much at all is simply that yes, we can detect artillery to a limited extend through clouds and that the change in IR over the course of the war reflects changes in Russian strategy rather than changes in weather.
VIIRS-375m and the Ukraine War
Both the Kyiv and Donbas regions see significant IR activity in the first three months of war, with Kyiv being particularly active from mid to late March, and Donbas beginning to see significant activity with the renewed Russian offensive in the region starting in May. What is notable here is that while conditions supported detection of IR signals in the Donbas in May, we nonetheless see significant activity in the Kyiv region in the coldest parts of winter — why? Cloud cover should have precluded detecting much IR.
The Donbas, on the other hand, saw increased activity in late spring when fire conditions are generally more observable via VIIRS. Identifying VIIRS-375m pixels with known Russian BTG locations at the time presents an extremely coherent picure.
Supporting the argument that VIIRS-375m can detect artillery fire is present in the May 7th map of activity. Here fire is closely concentrated within the (then) known positions of various Russian battalion groups. This particular date is likely the high point of detectable IR emissions from the Ukraine war, baring unknown future events. This artillery barrage was fierce, but what is noteable is how detectable a similar pattern is months later.
While intensity had dropped, activity is still easily detectable across the front. Russian artillery attention had by this point began its long focus on the city of Bahkmut. Note that this is also peak fire season, so some of the far-flung IR detections from the front may be related to activity not related to war. Though then again, note that these fires are more common on Ukrainian-held territory to the west than Russian-held territory to the east. However, this particular snapshot reflects one of the last high-intensity observations of IR activity of the war. Because starting the first week of September, activity drops of precipitously.
There are two possible explanations for this sharp drop. The first is seasonal — that increased cloud cover has obscured most artillery. The second is that the successful Kharkiv offensive initiated by Ukraine on September 6th distorted Russian artillery use to date in the conflict. This could be because of ammo seized, or reflect growing ammunition problems for Russia.
The outstanding problems of understanding VIIRS-375m IR data are as follows:
1) Why are we able to see so much IR activity in Kyiv during the short period of winter/early spring fighting?
2) Why is there a consistent Donbas front IR signal? Didn’t they run out of things to burn?
3) Why does detected IR collapse at the start of the Kharkiv offensive, and why do these low readings persist to today (March 2023)?
To address these, we will first look at the physics of infrared radiation, specifically what wavelengths/energies are involved and which of those wavelengths/energies might see signals survive through clouds. Then we will look at the IR emissions at each stage of artillery, from gun to secondary fires started. Then we will revisit the 3 above questions to conclude our discussion.
Physics of infrared radiation
Infrared radiation (IR) exists along the electromagnetic spectrum, alongside other types of light such as the color blue, X-rays, and AM Radio.
A general rule to understanding light is that it is both a wavelength and a particle with a specific energy. Large wavelengths of light have very low energy, and can be used to communicate across large distances. AM and FM radios are longer and shorter wavelengths that can be read. AM, being the longer wavelength and lower energy transmission can be read over a larger region than FM. Wi-Fi can be thought of as a type of radio with a much smaller range than FM, but because it is higher energy it can carry more information (this also is why FM was higher quality than AM, and why 5 GHz Wi-Fi is faster than 2 GHz but works over a shorter region). The longer the wavelength, the lower the energy, the farther it can travel. Microwaves work over a shorter distance, and are particularly effective at getting water to move, thus translating wavelength into kinetic energy into a warmer breakfast burrito. Infrared conveys heat, as anyone who has taken a bite into that breakfast burrito would know. What we see as visible light — red to violet — is only a small part of that spectrum. As we get to vanishingly smaller wavelengths and much larger energies we enter the world of UV, X-rays, and finally gamma rays. But all these types of light are just two related phenomena — the wavelength and energy of photons.
VIIRS satellites pick up 5 375m bands of wavelengths/energy ranging from 12 micrometers (0.10 eV) to 600 nanometers (2 eV, also known as the color orange). This IR range covers a wide range of physical properties, because as anyone who has gotten into a hot car in July knows, infrared light isn’t stopped by windows. VIIRS gathers a wide band because different wavelengths/energies have different efficiencies traveling through substrates such as water.
The mid-range of IR that VIIRS can detect, 6 — 7 micrometers (~0.2 eV) is completely attenuate by water. However the lower wavelengths of ~600 nanometers (2.0 eV, color orange) and higher wavelengths of 12 micrometers (0.10 eV) both see transmittence efficiency as high as 80% through 17mm of water. The peak transmittance efficiency of nearly 90%) is at 3.5 micrometers (0.35 eV). If 100 photons of light with a wavelength of 3.5 micrometers enters a column of water 17mm tall, 90 of those photons would come out the top.
If this is surprising to you, keep in mind you can test this in your own home. Find an orange object (600 nm, 2.0 eV, VIIRS band I1) or simply use paper with orange on it, perhaps from a marker. Then see how much water you must put into a glass on top of it to obscure the orange color completely. What is more challenging for the transmission of light through water isn’t the attentuation through water, but the scattering due to multiple discrete droplets. This is why most clouds appear white to us — light scatters and randomizes across its surface. When a cloud is more grey, it is because light is transmitting through it from above. This is why the clouds overhead are sometimes less white than the clouds you see at an angle. The problem for VIIRS is that both are relevant — not all of its measurements are top down.
It is essential to note that Chen’s calculations are major generalizations — most such models must make many assumptions about cloud characteristics and they have tremendous variance. But part of the problem can be resolved by inverting it — what proportion of infrared radiation makes it through clouds to us? They include, after all, the same photons we are interested in detecting from VIIRS. However here we run into fundamental limitations — calculating every molecule of H2O in the various types of clouds and how their movements scatter or absorb various infrared bands exceeds human computational capability (still). But you don’t need a computer to draw observations from this — how often have you been able to see on a completely overcast day? Some sunlight is making it through. And we still see temperature maxima after noon on cloudy days, so some infrared waves are making it as well. The answer to how much infrared light can pass through a cloud is an unsatisfying and imprecise “some” (Trenberth et al. 2009).
What is important to note about transmittance efficiency is that it is probabilistic. More water lowers the proportion of photons capable of exiting that medium, and thus attenuates its signal. So the options are not binary — more cloud cover decreases the chances you see IR emissions from the Earth, but the decrease is probabilistic based on how tall and dense the cloud is. But this in turn simplifies to how many water molecules the photon interacts with, and the the chance of absorption or scatter for each.
Different clouds have different extinction coefficients for IR photons based on their density and thickness. Generally most clouds are thick enough to block IR, but note that longer wavelengths start to transmit more efficiently. This is because if the wavelenght of light is larger than the particle blocking it, it can “skip” moving through it. This sounds weird, but this relates to how things move at the speed of light. If the wavelength is larger than the particle interacting with it, then it is more likely the light has already passed it without interaction. Spooky physics, but it’s also why your AM stations were more reliable than your FM ones back in the day (for zoomers, why 2 GHz WiFi covers a larger area than 5 GHz, but at the cost of slower data).
When a satellite has a given resolution, from MODIS at 1km to Landsat-8 at an astonishing 30m, it doesn’t mean that the entire 1km2 or 30m2 area has to be engulfed in flame, only that somewhere in that area signal intensity is strong enough to be detected. So when VIIRS-375m indicates heat activity, it is within the area it covers, but not across the entire area that it covers. In the same way that crossing your eyes doesn’t actually change what is happening in front of you, the lower resolution is simply about our ability to see. Landsat-8 is very obviously the tool of choice moving forward, but data availability began in June 2022, so to fully test the hypothesis “is Russian artillery in 2023 the same as it was in 2022”, we have to rely on VIIRS-375m, particularly for winter-to-winter comparisons.
Note — turns out one of the best tools to monitor cloud cover is the 750m VIIRS data. This can capture reflectance, as clouds have a strong albedo effect.
To round our discussion of photon attenuation and VIIRS, it is important to remember the following two points:
- Detection of any fire requires high intensity somewhere in a VIIRS-375m pixel
- Clouds can attenuate that signal, and thick enough clouds attenuate it to extinction.
Normal fires can be extinguished this way — but are artillery fires normal? What is their IR signal like?
Artillery Infrared Emissions
Artillery can emit infrared radiation in three ways:
- The firing of a shell
- The detonation of a shell
- Secondary fires caused by detonation
Each of these IR sources are capable of being detected by satellites, but they also will have different wavelengths/energies and thus be impacted by cloud attenuation differently.
The first point of IR emission is from firing the round, and would be primarily visible in a plume stretching 8m from the initial shot. While this doesn’t cover a great area, its intensity will be quite high given how far these heavy rounds are made to travel. They will almost certainly be visible in Landsat-8, whose 30m resolution will neatly enclose the plume and gun alike. However if intensity is significant enough and firing regular enough, the cruder VIIRS-375m can (and has) pick up routine fire as well.
152mm howitzers emit primarily in the 400 — 800 nm portion of the electromagnetic spectrum. This region should see around 60–70% transmittence efficiency through 17mm of water; meaning it can but won’t necessarily travel through cloud cover. The tallest peaks at 780 nm correspond to the ignition of potassium and would unfortunately be missed by VIIRS bands I1 and I2.
The second IR emissions from artillery are from the detonation of the shell — and this gets interesting. Munitions can have different explosive agents, and those explosive agents can have fingerprints in the IR spectra emitted.
Contaminants in artillery shells can expose themselves in different wavelengths as the munition burns. For example, copper molecules will show IR peaks at 550, 590, 610, and 640 nm. Potassium will have discrete peaks at 780 and 794 nm. Thus, depending on how well the IR emissions are tracked, a considerable amount of information about the munitions used will be exposed to the right spectrometer when they are fired.
Finally, artillery will cause secondary fires upon detonation, particularly in the dry summer months from May to August 2022 in the Donbas. This spectra will be larger, and more easy to identify.
The black line represents IR emissions from a simple wood fire; with a massive peak at 780 nm representing potassium. Which again is going to be in the undetectable middle between VIIRS IR bands I1 and I2. While detailed diagnosis of artillery munitions and vegetation burning is in theory possible, it is not with our current tool of choice. We will need to zoom out to the broader IR range to see what we can learn with some confidence.
For broader infrared light, what kind of flames can we expect as different materials burn? The answer very much depends on the type of material which burns. For example, a bunsen burner (fuel source alcohol) used in laboratories provides a simple starting point. We see peaks localized at 2.9 and 4.5 micrometers — both of which are missed by VIIRS. Note that the peak at 4.5 micrometers is the signature for CO2. An alcohol fire would not be readily detectable by the VIIRS sattelite. How about for a fuel like gasoline?
Gasoline, a more complex mixture, has a correspondingly more complex emission spectrum in the infrared range, with each peak associated with the molecules combusting. C-H bonds and their scattering likely represent the VIIRS I4 band (3.5–3.9 micrometers), and again they are not the peak intensity, despite the efficiency with which they transmit through clouds.
In sum, most IR emissions from artillery from gun to collateral damage should be capable of penetrating light cloud cover, but would be obscured by constant dense cover. VIIRS satellites were designed to contend with the fact that the infrared wavelengths which transmit most efficiently through clouds (Chen 1975) are not the same as those which are most likely to be emitted during burning (Underhill-Shanks and Hudson 2000, Liu et al. 2007).
However there is opportunity here — detailed study of individual bands with discrete wavelength ranges affords the ability to positively identify burning material in cloudless skies. And the diversity of IR emissions from burning material combined with the uncertainty of cloud transmittance mean that absolute statements such as “infrared cannot be detected through cloudy skies” override a cacophony of independent and interacting variables.
Fire Radiative Power
Infrared, like any other portion of the electromagnetic spectrum, is energy. The quantity of energy over a period of time represents power. As such, a timed observation of IR emissions can be used to estimate the power, or wattage, emitted. This translates IR pixels into a time series of power emitted over an area. Provided conditions are clear, this can relate to the amount of vegetation burned (Wooster and Strub 2005).
In this classic study looking at fire radiative energy and biomass burned, a clear relationship exists between biomass lost and total energy released. FRP represents the mass turned to energy upon combustion detected via remote sensing. Assuming clear measurement on a sunny day, IR emissions will correspond to mass lost in a given location. And summing pixels can provide total energy release. With this foundation of certainty (physical principles) and uncertainty (interactions and unknowns) we can return to the Russian invasion of Ukraine and understand what the data do and don’t say.
Outstanding Questions
In the introduction, we outlined 3 key questions about the IR signals from the Russia-Ukraine conflict. Answers are not conclusive, but to hint that we can learn quite a bit from IR data. We will dive into the historical weather associated with these IR readings as well.
1) Why were we able to detect heavy IR activity around Kyiv in cloudy conditions?
The answer to this question, in part, may be that cloud cover in the first few days of the conflict may have been enough to obscure some of it, particularly if we look at the first months.
In this image, we see small activity in early March in the Kyiv region, but a tremendous amount in the last two weeks of the month prior to Russia’s withdrawl. Let’s take a detailed look at weather that day.
Maximum IR detected was on March 16th and 21st, and both of these were sunny cloud-free days. However cloudy days are interspersed through this month, and some IR clearly made it to the satellites. Those most likely to survive the journey through the clouds were the 3.5–3.9 nm wavelengths that form VIIRS band I4. Note that there were 11 sunny days and 16 cloudy days — and clouds are not a complete impediment to the IR signals being measured here. They did however likely obscure most IR signals from early March due to precipitation.
2) Why is there consistent Donbas IR signal in spring/summer 2022?
One would be tempted to say it was because of consistently favorable IR conditions, but that is not the case.
Cloudy and rainy days were more common than sunny IR-detection-friendly ones. We can find some evidence that rainy conditions does indeed suppress VIIRS-375m signal. Note that May 18th, 2022 was a rainy day in Sieverodonetsk. Let’s look at it’s IR map for that day.
One can see an absence of IR detected around Sieverdonetsk on rainy May 18th, while some signal is detectable from drier but still cloudy Izyum. This illustrates that while precipitation likely extinguishes any chance at detecting IR emissions, some signal still survives cloudy skies. Let’s look at Sieverdonetsk on the cloudy day of May 25th.
Here signal returns to Sieverdonetsk, but it is a bit weaker than we see for the sunny day of May 7th (see below). What this tells us is that cloud cover attenuates, but does not eliminate, the signature of artillery fire. As long as the combined moisture of clouds above a region are less than 17mm vertically, some IR signal will make it through them even given the scattering effects. In theory if we knew enough about cloud-cover, we could compensate for missing signal.
Revisiting our highest signal day (~3,500 MW FRP) in the Donbas on May 7th, 2022, one can see conditions were ideal and sunny for VIIRS-375m detection across the length of the front. This confirms that while sunny cloud-free days provide us with our clearest view, cloudy days attenuate, but do not fully obscure, our picture of fighting. Rainy days, however, very much can. While there is lots of uncertainty about how clouds transmit IR, we are on much firmer ground with FRP calculations on sunny days — the energy released is equivalent to the mass lost and IR is a good measure to conveying that information.
The answer to our question — “why was Donbas signal so clear” — was because artillery was heavy and constant throughout the late spring and summer of 2022 in the region. Cloudy days attenuated, but did not eliminate, that signal. This is due to the interaciton of a number of factors — including but not limited to what materials were burning, consistency of cloud cover, and the interaction effect between absorption and scatter of specific IR bands emitted by fires and detected by VIIRS. Which leads us to our next question — what changed in September?
3) Why does IR in the Dobnas in particular and war in general stop precipitously at the start of the Kharkiv Offensive in September 2022?
This is perhaps the most interesting question as it has implications for Russia’s ability to currently prosecute the war. The first stop is to look at weather in the Donbas, specifically Izyum, around the critical time window.
Just like Sieverdonetsk in May 2022, sunny days in Izyum are outnumbered by cloudy days in September 2022. But while the Donbas shows a strong IR signal in May, this period collapses to near 0 IR emitted after Ukraine launched its successful counteroffensive. Sunny days like the 19th are still near-0 readings of IR activity, and this is simply because the Russian artillery machine stopped. It is fair to say based on what we saw for May in Sieverdonetsk that there isn’t a compelling climatic explanation for clouds to block all IR. But we know the answer here — Ukraine stopped the Russian war machine with a rapid advance that collapsed their lines. That is why IR activity drops off.
But is that the reason that is stays persistently low for the rest of the war to today (early March 2023)? Is it perhaps the case that the entire country of Ukraine has been covered universally by clouds for the past 6 months? Well, let’s see what October looks like for the same region.
If anything, October 2022 is sunnier than September 2022 in Izyum by one day. But the combat line had shifted to Svatove-Kremina by this date, perhaps cloud cover was universal for that area?
It is the case that Svatove was cloudier and rainier that month, but there are still sunny and partly cloudy days for some IR signal to emit — but still 0. In fact, the picture gets muddier the closer we look. For example, in early 2023 Russia launched an offensive across the entire line of contact in the Donbas. And what’s peculiar is that it is very sunny in Svatove, one of the main points of conflict despite it being the middle of winter.
In fact, it is exactly as sunny in Svatove in January 2023 as it was in Sieverdonetsk in May 2022 when record high IR emissions were detected. There are multiple windows of opportunity to detect IR emissions from artillery and secondary fires. What did we see?
On sunny January 25th, 2023, preceded by two other sunny days, we still see very low IR activity across the Svatove-Kremina frontline. This is qualitively and quantitively different than the pattern observed for the Donbas in May 2022. The Institute for the Study of War noted that Russian spoiling attacks occurred across Ukraine on this date — but they left very little IR signature as past Russian attacks. It should be noted that the Russian offensive went into full swing only a few weeks ago, and Svatove is a bit cloudier for February, but sunny days don’t show IR emissions near what we saw in Donbas in the late spring/summer of 2022. Something has fundamentally changed about Russian artillery since the Kharkiv offensive.
But what? Telegram channels highlight “shell hunger” as a very real problem for Russia, with the Wagner group publicly shaming the Russian Ministry of Defense (MOD) for more fire support and supplies. There are also reports of embarrassingly old artillery ammunition being used. But there is still active artillery across this area — why can’t VIIRS see it? There are a few hypotheses:
1) Artillery fire has fallen below detectable levels for VIIRS
2) Nothing is left to burn
3) Artillery ammunition is different, and IR is more readily attenuated by water vapor
4) Artillery is too dispersed
Of these, 3) is the least likely since even on sunny days we don’t see much IR activity since the fateful first week of September 2022. But the possibility can’t be permanently discounted, as 6–7 micrometer IR emissions would be attenuated and partially extinguished even by humidity in the air. 2) is also unlikely as heavy fire along the Svatove-Kremina region was never observed. Combat footage from the region indicates that it, and Kremina in particular, are still heavily forested. This is plenty of wood to burn that can penetrate low cloud cover easily.
This leaves a combinatio of hypothesis 1) and 4), that artillery fire is now too infrequent or too dispersed for VIIRS to identify via IR emissions. Dispersal cannot be a complete explanation, since the offensive at Bahkmut still shows very little IR signature proportional to the fighting then occurring. In general, this would suggest there is a critical value of rounds expended at which satellites provide useful data. It would also indicate that Russian strategy of firing artillery has changed profoundly in response to Ukraine’s Kharkiv offensive.
In the first two months of 2023, many asked “when will the Russian offensive start” when experts of the Russian military answered “it has already started”. It may be the case that the same is true of the Russian ammunition crisis — it is already here, and has been here for some time.
References Cited
Chen, C. 1975. Attenuation of electromagnetic radiation by haze, fog, clouds, and rain.
Hird, K., Bailey, R., Stepanenko, K., Philipson, L., Mappes, G., Kagan, F.W. 2023. Russian Offensive Campaign Assessment, January 25th, 2023. Institute for the Study of War. https://www.understandingwar.org/sites/default/files/Russian%20Operations%20Assessments%20January%2025%202023.pdf
Kaaret, P., Tammes, S., Wang, J., Fuller, C.A. 2022. On the potential of flaming hotspot detection at night via multiband visible/near-infrared imaging. Remote Sensing 14: 5019
Liu, Z.G., Kashef, A., Crampton, G., Lougheed, G., Almand, K.H. 2007. Research progress of the international road tunnel fire detection project. National Research Council Canada. https://www.researchgate.net/publication/44091749_Research_progress_of_the_international_road_tunnel_fire_detection_project
Schoeder, W. and Giglio, L. 2017. Visible Infrared Imaging Radiometer Suite (VIIRS) 375 m & 750 m Active Fire Detection Data Sets Based on NASA VIIRS Land Science Investigator Processing System (SIPS) Reprocessed Data — Version 1. Product User’s Guide Version 1.3. Prepared for NASA. https://lpdaac.usgs.gov/documents/132/VNP14_User_Guide_v1.3.pdf
Steward, B.J., Perram, G.P., Gross, K.C. 2011. Visible and near-infrared spectra of the secondary combustion of a 152mm howitzer. Applied Spectroscopy 65(2): 1363–1371.
Trenberth, K.E., Fasullo, J.T., Kiehl, J. 2009. Earth’s global energy budget. Bulletin of the American Meteorological Society 90(3): 311–324. doi: 10.1175/2008BAMS2634.1
Underhill-Shanks, K., and Hudson, M.K. 2000. Fixed and scanning infrared radiometers for combustion studies. Journal of Pyrotechnics 12: 57–67
Wooster, M.J. and Strub, N. 2002. Study of the 1997 Borneo fires: Quantitative analysis using global area coverage (GAC) satellite data. Global Biochemical Cycles 16(1): doi:10.1029/2000GB001357
Wooster MJ, Roberts G, Perry GLW, Kaufman YJ (2005) Retrieval of biomass combustion rates and totals from fire radiative power observations: FRP derivation and calibration relationships between biomass consumption and fire radiative energy release. J Geophys Res 110:D24311. doi:10.1029/2005JD006318
