The map showing the damage is now ready; the rescue team required it six hours ago.
The sentence refers to a problem which is only weakly connected with the clarity of the image; it is possible to have a highly detailed satellite photograph of a town that has flooded and yet be unable to provide the response team with anything useful, since by the time the image arrives the water has shifted, the roads have altered, and the decisions have already been taken without taking the image into account.
Then what is the situation? A satellite damage map is the result of a series of steps. The event takes place; somebody then uses a mapping service; a satellite has to be in the correct position with clear sky conditions, the data has to be downlinked and processed, an analyst has to examine it, and it is not until then that the map gets to a person who has to make a decision. Each of these steps takes time, and that time is not included in the resolution figure on the data sheet.
In this article, I'm going to look at how the hours are spent, explain why a before-and-after comparison can mislead you, discuss what radar and aerial imagery are able and unable to provide, and examine what a damage map should say regarding its own limitations. The examples are taken from published Copernicus emergency-mapping missions, so you can verify each one for yourself.
Why does image resolution by itself not determine whether a map is useful?
The resolution deals with one question: how small an object can I see? The emergency manager, on the other hand, is asking a different question: what is currently happening and which area needs assistance first?
The two questions may have opposing effects. The Copernicus rapid-mapping team admits this. In the case of emergency response, it starts using the first satellite image that becomes available, and that image is often of poor quality since it might be only partly clear due to cloud cover, taken at a steep angle, or covering only part of the area of interest, as the service's own documentation states.
The main point is this: a sharply focused image which comes late or only shows half the town is worth less than a more basic one that arrives on time. In an emergency, the value of an image depends just as much on timing, the extent of coverage and a clear view as it does on the number of pixels. If any one of these factors is absent, then the detail in the other areas is of no use.
How long does it actually take to produce a satellite damage map?
Permit me to illustrate this with an actual example, since the timestamps provide more information than a diagram could.
A Real Timeline: The Tonga Eruption
Copernicus Rapid Mapping was put into action approximately 10 hours after the eruption in Tonga, at 16:45 UTC on 15 January. The first image obtained was on 16 January at 21:52 UTC. The initial product relating to Lifuka Island was issued within 48 hours of the system being activated, and subsequent updates to the assessment were made as clearer images became available. More information is given in the Copernicus EMS report on the activation.
If you do the math, you'll find that there are about 29 hours between the time of activation and the first image—this is my figure based on the published timings. When you add on the 10 hours before activation, the first image was produced about a day and a half after the event. The team is therefore working quickly, not failing.
Even so, if there was a persistent cloud over a particular area, it would be impossible to analyse parts of that area in any of the three images used, and the same report states that the damage figures could be underestimates.
Revisit Intervals Are Nominal, Not Promised
As stated in ESA's mission facts, the Sentinel-2 pair revisits the same location every five days at the equator. According to the ESA page for Sentinel-1, Sentinel-1 radar has a revisit time of six days with two satellites. The Disasters Charter states that Sentinel-2's five-day interval is equivalent to two to three days at mid-latitudes, although this is only true under cloud-free conditions.
They are the design intervals, not a guarantee that the satellite will pass over your town the day after. If you'd like to find out which of the free archives you can access, I have included a list of them in our guide to free satellite imagery sources.
Deliveries are also uneven throughout the map. A recent preprint relating to a 2023 earthquake showed that around 71% of the optical data from the affected area had arrived within one day and all of it was in by day five, but in remote areas it took three to five days since collection prioritised the large cities near the epicentre. The preprint has not necessarily undergone peer review, so the figures should be regarded as indicative.
Clouds and Viewing Angle
The mapping service was activated on 23 December for the 2018 Indonesian tsunami, but for almost a week optical imagery was prevented from being taken due to persistent cloud cover. The first image of the volcanic island obtained by Sentinel-2 was on 29 December. In the case of the Corsica fire, the initial delineation map included a big "Not Analysed" area because of the cloud, and it was not until 26 and 27 February—after the service had been activated on the 24th—that cloud-free images became available.
Why Can a Before-and-After Pair Deceive You?
Damage mapping involves a comparison: you need an image taken before and one taken after. So what ensures that the two images can be compared?
What you would like is to use the same type of sensor, one with a comparable viewing angle, during a similar season and at a similar time of day, and with clear skies in both cases. In reality, you usually can't achieve this. Although the Copernicus documentation states that imagery taken before the event is essential for comparison, this imagery is not always available shortly before the disaster occurs.
Consider the situation when the two images do not match. For example, a summer image from last year and a winter image taken at a steep angle after the event will display differences which are entirely unrelated to damage—such as shadows, the leaning of buildings, vegetation, and standing water caused by normal rain. The software will readily point out all of these.
It is the same kind of trap that I explained in our earlier post about change-detection maps. A change map shows you which pixels have changed; it does not indicate that the ground has changed. When dealing with a disaster, you have less time to distinguish between the two, and it is at that point that the errors become costly.
What do you use in situations when optical imagery is not available?
Radar (SAR): Sees Through Clouds, Still Needs Care
Radar is effective at night and in cloudy conditions. When Hurricane Harvey caused flooding in Texas, Copernicus selected radar for precisely these reasons, as it is good at distinguishing between flooded ground and bare soil. The extent of the flooding was determined by comparing with a pre-event image, as the activation notice states. Nevertheless, the data collection did not begin until around midnight on 27 August, after the system had been activated on the 25th, which meant a waiting period of about two days (once again, my arithmetic).
SAR is not a general solution. In the case of the Indonesian tsunami, the Copernicus team stated that optical imagery was required for producing the damage grading maps and that radar could not be used for this purpose. With regard to earthquakes, a study that used coherent change detection on SAR imagery found that the detected changes were concentrated around the buildings that had suffered heavy damage, but this result was verified by means of aerial photo-interpretation. Many of the SAR methods for assessing damage also depend on a prior SAR image, which is frequently not available in operational situations. The NASA JPL researchers have noted that latency and revisit time are still limitations for near-real-time response.
What I would suggest is to interpret a SAR product as meaning "something changed here" and then check it.
Aerial and Drone Surveys: Fill Specific Gaps
Copernicus can make use of imagery obtained from drones and aircraft during certain operations. A drone can fly beneath the clouds when you decide, at a resolution you have selected. This is particularly useful in addressing a particular type of situation: that of a damaged bridge, a hospital site, or a block which the satellite had only viewed from a steep angle. Since it does not provide coverage of entire districts, a crewed aircraft or a satellite should be used for the broader view.
One thing to bear in mind is that the airspace above a disaster zone is often crowded with rescue flights and might be restricted, so it's important to contact the local aviation authority before you start. In our article on thermal mapping for wildfire response, we explained how drones function. Also, if the results may be in doubt later on, the way in which a drone survey is recorded makes a difference, a point that I made clear in the post on why satellite imagery is rejected as evidence.
What Should a Damage Map Say About Its Own Limits?
A map that appears to be complete but in fact hides the information about when it was made is more dangerous than one that is unfinished but honest. Therefore, the question is—what should be printed on it?
- The acquisition date and time of every image used, in UTC
- The sensor, and the viewing or incidence angle where it matters
- The release time of this version of the map
- The method: visual interpretation, automated, or both
- The areas not observed, such as cloud-covered ground
- A confidence statement, and a version number
The Copernicus products currently have 'Not Analysed' areas, which is the correct approach. For example, regarding the forest fire in Greece in Sithonia, the first map, which was based on Sentinel-2 data, estimated that 800 hectares had been burnt, whereas the subsequent monitoring map, using a different image, showed 629.7 hectares, some of the burned area being hidden by cloud. Since the two figures are derived from different images, a person who only looks at the number has no means of knowing how much weight should be given to it.
What is the way to create a workflow that continues to update?
To cut to the chase, a damage map is not a document; it's a feed, and that's the way I'd arrange it.
- Start by deciding on the questions—specifically, which roads, shelters, hospitals, and districts are important for the next decisions?
- Start with the areas that are most urgent and deal with them first, even if the product is of a coarser type.
- Choose the sensor according to what the sky permits: use radar when the clouds obstruct optical sensing, and use optical sensing for building-level accuracy, employing drones to check the most important locations.
- First release version 1, making its gaps obvious, and then carry out an update. The Tonga activation included several updated releases for precisely that reason.
- Make a record of releases: specify the image, the time, the changes made, and the things that have still not been observed.
The Takeaway
The resolution indicates what a satellite was able to see, while the timing shows whether it still has relevance. If you are going to rely on a damage map, you should find out when the image was taken, when the map was published, what the weather permitted, and which areas were not examined. A map that honestly gives you those answers is more useful than one that is more attractive but doesn't.
Our Remote Sensing team can assist you in deciding which satellite or aerial data to use and in working out how to record it if your work involves making actual decisions based on such data.