Your Change-Detection Map Shows Change. Did the Ground Actually Change?

Your Change-Detection Map Shows Change. Did the Ground Actually Change?

When a client gives you two satellite images — one taken in the spring and the other in late autumn — you carry out a satellite image change detection analysis and the resulting map flashes with indications such as the edges of fields, the banks of rivers, the outskirts of a town, and a number of large areas where something obvious must have taken place. This appears to be a significant result.

To get down to business: a change map indicates the areas in which two images differ. It does not show where the ground has changed. Although the images overlap, they are not made up of the same pixels, and it is precisely in this gap that land-cover monitoring, disaster assessment, urban expansion studies and environmental reports make their mistakes.

In this article I'll look at the main reasons why two images can be different when there has been no change to the ground, explain how to select a pair of images such that these reasons are kept to a minimum, and describe how to check the results before including them in a report. The kind of checking I'm referring to is the one that underlies any serious remote sensing output.

How Satellite Change Detection Works

In simple terms, the software matches two images of the same area and compares the pixel values; you can subtract one date from the other, divide them, or classify each date individually before comparing the two images. Radar (SAR) procedures perform a similar operation but use backscatter rather than colour. The result is a layer showing the differences, typically displayed as areas of change and areas of no change.

What do we mean by false change in the context of change detection?

A pixel value results from the combination of the surface, the atmosphere above it, the sun's angle, the sensor, and the processing chain; if any one of these factors changes between two dates then the value will alter even if the ground itself has not changed. This kind of change is a false one and in the final output it appears completely identical to a real change. For this reason a change map should be regarded as a list of possible changes rather than as a definite result.

A comparison based on classification includes an additional penalty. When two maps are each correct 90 per cent of the time and their errors are independent, only about 81 per cent of the pixels will be correct on both occasions. That calculation is my own, not a result obtained by measurement, but it illustrates how quickly small errors accumulate.

Seasons, Soil Moisture and Crop Cycles Look Like Change

Vegetation is by far the main cause of false changes and the most difficult to detect since the images appear completely reasonable. A field in April and the same field in October show different surfaces, and so do a floodplain during the dry season and during the wet season, or a hillside before and after snowmelt.

Timing is one of the issues involved. Since plants reach each growth stage on different dates from year to year, two images taken on the same calendar date can still show different stages of growth. Moreover, there is no consensus regarding when the growing season is said to begin. A comparison carried out by the USGS of four commonly used methods found that they agreed only weakly (with an r² value between 0.10 and 0.35) since each method picks up a different stage, ranging from snowmelt to early growth to the appearance of the first leaf. When four such methods applied to a single time series give contradictory results, a comparison based on two dates will not be able to resolve the issue.

The answers are also affected by the type of sensor and by the presence of clouds. A study carried out on mangroves in northern Australia showed that the apparent phenological changes depended on the sensor used, the location, and the degree of cloud contamination. Although Landsat and Sentinel-2 produced similar results in that study, neither of them matched MODIS.

Soil moisture functions in the same manner: wet soil appears darker in optical images and radar backscatter increases as the soil becomes wetter, which means that a dry period followed by a week of rain can be interpreted as a change in land. When you also take into account ploughing, fallow periods, irrigation and harvesting, a farming landscape can change completely between two dates even if there has been no change in land use.

What you need to do is to compare the same time of year, or ideally use a time series in order to see what normal variation is like before you label something as being abnormal.

For a personal project carried out in Multan I created a multi-year Sentinel-2 NDVI stack in Google Earth Engine rather than comparing two different dates. By using a stack you are able to see the usual yearly variation of a field or canopy and thus you can assess any change in relation to this pattern. Both the project and the script are available on my GitHub.

Resolution, Sensor Differences and Image Alignment

How great an accuracy does image alignment need to have?

More than most people anticipate. In a well-quoted 1992 study, Townshend and his colleagues employed degraded Landsat images with no actual change. In four out of the seven test regions, a misalignment of just one pixel resulted in an error equivalent to more than half of the true variation in NDVI. To reduce that error to 10% it was necessary to align the images by about 0.2 pixel in the areas with dense vegetation, while the more sparsely vegetated semi-arid areas could tolerate an alignment of 0.5 to 1.0 pixel. Dai and Khorram later stated the same requirement of one-fifth of a pixel.

Misregistration is noticeable at the edges; at any sharp boundary—for example, a road, a field edge, a building or a shoreline—a slight shift causes a bright area to be adjacent to a dark one and the software then registers this as a change. A simple test is this: if the changes in your change layer are shown as thin lines along roads and field edges, then you should suspect the alignment before you suspect anything about the land cover. The coordinate system also plays a role, and we have already explained how a mismatch can quietly cause a shift in the deliverables in our previous post on datum mismatches in drone deliverables.

Different sensors, different numbers

Resolution also has an effect here. The image taken with 30 m pixels averages together surfaces that a 3 m pixel image would separate, which is why comparing the two shows an apparent change on ground that has not actually changed.

Even those sensors designed to work together are not the same. In order to reduce the differences between the two systems, NASA and the USGS developed the Harmonized Landsat and Sentinel-2 product. In the final product, reflectance on the same day still varied by a few per cent (the figure being below 4.2% in the red, near-infrared and two shortwave infrared bands), and the atmospheric correction process slightly increased the difference in the blue and green bands when compared with the top-of-atmosphere data. A few per cent may seem minor. Yet it is not minor when the change you are looking for is also only a few per cent.

Cloud Shadows, Haze and Inconsistent Acquisition Conditions

It is easy to identify and conceal a thick cloud; shadow and haze are more difficult. A cloud shadow darkens the ground and may be mistaken for water, a burn scar or previously vegetated land. Haze raises the blue bands and reduces the contrast, thereby altering the vegetation indices.

Although automatic masks are helpful, they have not been fully resolved either. A study that compared three Sentinel-2 masking methods—Fmask, ATCOR and Sen2Cor—on 20 scenes showed an overall accuracy ranging from 89 to 92% for the entire scenes. Where the different methods gave conflicting results, the accuracy dropped to a range between 45% and 62%. A separate comparison of five algorithms revealed a compromise between missing clouds and incorrectly flagging them as clouds, with no one algorithm being the best at handling both clouds and shadows.

The majority of the scene is all right, and it's the questionable pixels that have the potential to cause a false change; in that case I would advise you to examine the mask itself, apply a buffer around the clouds and shadows, and regard any changes near the masked areas as unconfirmed.

The conditions for acquisition introduce yet another factor since the angle of the sun and the viewing angle vary from pass to pass, causing the shadows cast by trees, buildings and slopes to fall in different positions. Two images taken by the same sensor a month apart can still show different shadow patterns.

Choosing Pre-Event and Post-Event Imagery for Change Detection

The way images are chosen has a greater effect on the outcome than the algorithm. When I look at pairs of candidates, this is what I would examine:

  • The season refers to the same time of year, or at least the same growth stage and moisture level.
  • If possible, use the same sensor, otherwise use a harmonized product.
  • In the case of geometry, the sun and the viewing angles are the same; for radar it is the same orbit direction and incidence angle.
  • Cloud, shadow and haze are minimal over the area of interest, not just over the entire scene.
  • The time gap can be as short as the other points permit.

The Esri procedure for flood mapping states that reliable change detection for radar is achieved by comparing images taken before and after the flood under similar viewing conditions.

The actual compromises involved are particularly significant following a disaster. When researchers were compiling a dataset of building damage based on Sentinel imagery, they selected the clearest date available before the event, then chose the post-event images by weighing the amount of cloud cover against the length of time that had passed since the event, and kept the Sentinel-1 images that were on the same orbital path. The first image taken after an event is not automatically the best; if you select a later one, you should note the reason for doing so.

How can a satellite change detection map be validated?

That's when the map becomes evidence.

Reference data and sampling

A well-known set of good practices published by Olofsson and his co-authors suggests that you should select a probability sample of locations and assign to each one reference data which is clearly more accurate than the map. You then present an error matrix showing the overall accuracy, the user's accuracy and the producer's accuracy, together with confidence intervals. Since changes are rare, a simple random sample usually ends up on areas that have not changed. A common solution in this case is to use a stratified sample which ensures that there are sufficient points in the change category.

Area numbers need uncertainty too

Counting changed pixels on the map is tempting. However, a classification error can cause that value to lie a long way from the actual area, and the same paper suggests that the area should be estimated from the reference sample together with a confidence interval, not just a single figure.

Independent observations

The reference must be independent of the images used to produce the map; field visits, drone surveys, higher-resolution imagery and land records are all suitable methods. For instance, Copernicus emergency mapping improves its semi-automatic flood assessments through manual photo-interpretation. When making a decision of importance, carrying out a specific drone survey or a topographic survey of the doubtful areas provides evidence that is easier to defend than one obtained from a satellite map checking itself. This is the same evidence discipline that we outlined in our post on how to prove an accuracy claim.

Before including a change map in a report I would like to have answers to four questions: have the dates been matched for the season? Has the alignment been checked against the edges? Have the masked areas been reviewed? And has a sample been verified against something independent?

The Actual Takeaway

A change map consists of a list of possible candidates, not of definite findings. Most of the work is carried out earlier on, during the matching process, by aligning the season, sensor, geometry and registration so that the images differ only in the areas of the ground that have actually changed. The final step involves checking a sample against an independent source and then giving a measure of the uncertainty, since a map that has never been verified is difficult to justify if a client or a regulator asks about it.

The work of our remote sensing team involves carrying out change detection which is capable of meeting that requirement, at every stage from image selection through to validation.

Back to blog

Leave a comment