Satellite Image Change Detection: Comparing Different Times

Satellite image change detection is a technique for finding change by comparing images of the same place taken at different times. This post covers the difference between before-and-after comparison and time series decomposition, and how to separate false changes from real ones.
In short
- Before-and-after comparison uses two images; time series decomposition separates trend and seasonality.
- Shadows, tides and seasonal differences are easily mistaken for false changes.
- Reading development, flooding and collapse each calls for a different change detection strategy.
Put a satellite photo from this time last year next to one taken today, and something clearly looks different. A place that was forest has turned brown, or the edge of a river has widened. But is it really because of a landslide or flooding, or does it just look that way because the sun angle or the season differed on the days the images were taken?
Change detection is, literally, the work of comparing images from two or more times to find what has changed. The principle looks simple, but in practice telling real change from false change is much harder. At a disaster site, one false detection can lead to a wrong evacuation order or a rescue dispatch to the wrong place, so it is important to understand and use this concept accurately.

What is the difference between before-and-after comparison and time series decomposition?
Before-and-after comparison overlays exactly two times, one image from before the disaster and one from after, to find the pixels that changed. It is useful when you need to look quickly at before and after a specific event, such as an earthquake or a landslide. Time series decomposition, on the other hand, stacks dozens or hundreds of images over a long period and mathematically separates the trend (change that steadily increases or decreases), seasonality (a pattern that repeats every year) and the remaining irregular change. It lets you tell whether a field is showing the same crop change it repeats every year, or whether this year is unusually strange. Before-and-after comparison is fast and intuitive, but it has difficulty filtering out cases where conditions on that particular day happened to be unusual. Time series decomposition needs a lot of data, but it is strong at seeing patterns such as long-term land change or the spread of wildfire-prone areas. For the early response to a flood, where response speed matters, before-and-after comparison fits better; for cases that need cumulative trends, such as monitoring landslide risk areas, time series decomposition fits better.
Why do season and shooting conditions create false changes?
Even the same place has a completely different vegetation color and density if you shoot it in summer and in winter. In winter the leaves fall and the soil is exposed, so it can look as if the forest has disappeared. And if the soil moisture is high after rain, reflectance drops, and it is picked up as change even though nothing really happened. The difference in sun elevation with shooting time also changes the shadow length, making the boundaries of buildings or mountains look different. With SAR (Synthetic Aperture Radar, a method that sends radio waves and builds an image from the reflected signal) images, the same terrain is also captured differently when the viewing angle or the orbit direction changes. To reduce such false changes, the basic approach is to choose and compare images from as similar a season, as similar a time of day and as similar an orbit condition as possible. So when you look at change detection results, you need a habit of first checking the shooting dates and conditions of the two images, and this is a key step for reducing false detections in areas with large seasonal vegetation change, such as when reading wildfire-damaged areas.
How do you read change caused by development?
Change caused by urban development usually has a clear geometric pattern. When straight roads and rectangular building layouts appear where there were fields or mountains, change detection algorithms often capture not only the reflectance difference but also the regularity of the shapes. However, a construction site keeps changing its appearance stage by stage, so a before-and-after comparison at a single time may leave you unsure whether construction is under way or already finished. In that case, it is safer to stack several times in sequence and check the direction in which the change is progressing. The same principle is used to monitor illegal mountain development and unpermitted fill work. If the vegetation index in a certain area suddenly drops and stays that way, you judge that it is likely permanent development rather than temporary logging. Checking the persistence of change in this way is used to catch early signs of disaster, such as mountain damage and illegal landfill.
How do flooding and collapse look different in change detection?
Flooding covers the ground surface widely as water rises, so it appears as a dark, smooth color in optical images and as black, with almost no reflection, in SAR images. In change detection, if a stretch that was normally dry ground suddenly shows these features, it is judged to be flooding. Building collapse and landslides, on the other hand, appear as local, irregular patterns. The straight outline of a building that was intact disappears, or the vegetation pattern on a hillside turns into an irregular pattern of exposed soil and rock. Flooding is a relatively simple change over a wide area, so automatic detection is easier, but collapse is often misrecognized because of shadows and piles of debris, so a person must zoom into the images and check. In disaster situation rooms, a two-track procedure is often used that separates the two: flooding is scanned over a wide area quickly with automatic algorithms, and suspected collapse areas are checked further with high-resolution images.
How do you check whether a change is false or real?
You cannot trust every pixel the algorithm marks as change. This is exactly the part you only learn by doing: you must cross-check suspected areas against images from other sources, ground photos, news reports and field reports. For example, if a certain field suddenly looks darker in the change detection result, you can tell whether it is flooding or simply soil turned over after harvest only by comparing it with farming patterns in other areas at the same time. Cloud shadows and temporary sensor errors can also be caught as change by mistake, so the habit of also checking images from a day before or after greatly reduces false detections. This kind of cross-validation cannot be skipped before a judgment that leads to real action, such as an evacuation order, and at disaster sites a system is recommended in which automatic detection results are used only as first-stage screening and the final judgment is made by a person.
Step by step on screen
The figure below is a diagram that summarizes the order of the steps. Menu positions are the same in the English interface.

- Prepare two images
Download images of the same area from two times from a public satellite imagery site (for example, Sentinel Hub or USGS EarthExplorer) - Load the layers
In your GIS program, register the two images with Layer > Add Layer > Add Raster Layer - Check shooting conditions
In the metadata window, compare the shooting date, time and orbit direction, and note the differences in conditions - Run a band operation
Open Raster > Raster Calculator and make a difference image with an expression such as (image 2 - image 1) - Visualize the result
In the layer's Properties > Symbology, emphasize the color contrast of the difference image to see the changed areas - Cross-validate
Compare the areas marked as change with high-resolution maps, news and field photos to check for false detections
Menu names may differ slightly between program versions.
Prompts you can copy
Get a checklist for checking the possibility of false changes when reporting change detection results
You are a GIS analysis assistant who reviews satellite image change detection results. I am going to write a report on the result of detecting [disaster type, e.g. flooding / landslide / wildfire] in [study area]. The shooting dates of the two images are [date 1] and [date 2], and the seasonal difference is [yes / no]. Answer in the format below. 1) Three possible causes of false change under these conditions 2) The image features to check in order to tell each cause apart 3) One line of caution that must be stated in the report Do not guess and assert facts you could not confirm; mark them "needs field verification".
Replace the bracketed parts for your situation. Check the facts in any answer before you use it.
Try it yourself
- Download summer and winter images of the same area from a public satellite imagery site and compare the color difference yourself.
- Look at before-and-after disaster comparison photos in the news and see whether there are differences in shadows or shooting angle.
- With a free GIS program, make a difference image of two images and check whether the changed areas match the actual terrain.
Common questions
Q. Do I have to use SAR images for change detection?
A. No. Optical images are also widely used. SAR is advantageous in cloudy areas or when night observation is needed, and the choice depends on the situation.
Q. Can I confirm disaster damage from before-and-after comparison alone?
A. It is not recommended. Because false changes are possible, it is safer to make the final judgment after a field check or cross-validation with additional images.
Q. About how many images does time series decomposition need?
A. There is no fixed standard; it depends on the purpose and the interval of the analysis. To separate seasonality, data that covers at least one year or more of cycle is generally recommended.
What comes next
Satellite image change detection looks like the simple task of overlaying two photos, but behind it are many traps created by season and shooting conditions. A later post in this series looks at the principle of how AI automatically classifies and labels the changes detected in this way, that is, how that learning process works.
This is an English adaptation of a post in the Korean GeoAI basics series. The original is here, and more learning materials are on GeoAI for Disaster Safety.
This post was drafted with the help of AI tools and reviewed by the author. Tool features and data portals change over time, so check the official documentation when you practice.
Comments
Post a Comment