Wondering whether a forest, shoreline or city block really changed? Satellite image change detection compares two or more dates of the same place; reliable results need matched sensors, cloud-free pixels and a clear threshold.
This guide belongs to the wider topic of satellite imagery, so it focuses on the comparison step rather than on how sensors work.
| Item | Value |
|---|---|
| Median cloud-free land observations in 2022 (Harmonized Landsat and Sentinel-2) | 66 cloud-free observations |
| Same-day Landsat versus Sentinel-2 difference, red, NIR and SWIR bands | below 4.2% |
| Between-sensor difference for most vegetation indices | below 4.5% |
| Point at which CCDC flags a pixel as possible land cover change | three times the RMSE |
These figures come from the USGS record of the Harmonized Landsat and Sentinel-2 dataset and the Esri help page for the CCDC tool. Datasets get reprocessed and documentation gets revised, so check the source before you rely on them.
What satellite image change detection compares
At its core, change detection means looking at the same ground location on two or more dates and flagging pixels that moved more than normal variation would explain. The hard part is that pixel values shift for several reasons, and only some of them reflect real change on the ground.
The CCDC help page sorts those fluctuations into three kinds:
- Seasonal change: shifts caused by temperature, sunlight and precipitation through the year.
- Gradual change: trends linked to climate variability or long-term land management, such as bare soil slowly expanding as rainfall declines.
- Abrupt change: sudden land cover change from deforestation, urban development, natural disasters and similar events.
The CCDC algorithm identifies all three, yet its main purpose is to catch abrupt change. In the northern hemisphere, a higher density of green vegetation in summer than in winter is expected, so a summer image set against a winter one shows a gap that is not news.
Why clouds cause false alarms
Cloud contamination is one of the largest sources of error. A USGS-hosted study of Landsat forest monitoring notes that unflagged clouds may be mapped as false changes, and that the level of such false changes can be comparable to or many times more than the actual changes, even for images with small percentages of cloud cover.
The masking algorithm in that study produced omission errors of around 1% for the cloud class, according to the same paper. Most remaining errors sat around the edges of detected clouds and shadows, and many came from thin clouds that are hard to flag.
Esri’s documentation describes a different masking route. Because cloud shadow and snow look very dark in the shortwave infrared band, while clouds and snow look very bright in the green band, those two bands are used to find outliers and remove them from the analysis, as the CCDC page explains.
Making images from different sensors comparable
Mixing sensors adds a second source of false change, because two satellites can record slightly different reflectance for the same surface. The NASA Harmonized Landsat and Sentinel-2 (HLS) project was built to remove that gap, and the USGS publication record lists its processing chain:
- Atmospheric correction applied to top-of-atmosphere reflectance.
- Masking of clouds and cloud shadows.
- Normalizing view angle effects on bi-directional reflectance.
- Adjusting for sensor bandpass differences, with the Landsat imager as the reference.
- Delivering the harmonized data on a common grid.
The authors tested the result on 545 pairs of globally distributed same-day Landsat and Sentinel-2 images from 2021 to 2022, per the same record. In the final products the red, NIR and SWIR bands differed by less than 4.2% between sensors.
The blue and green bands stayed slightly higher than in the unprocessed data. The paper attributes this to the difficulty of correcting atmospheric effects in the shorter visible wavelengths, so treat those bands with extra caution in the HLS data.
More dates also help. For 2022, the first whole year with two Landsat and two Sentinel-2 satellites in orbit, HLS offers a global median of 66 cloud-free observations over land, and its record starts in April 2013, according to the USGS abstract.
Choosing a band or index to compare
You can run change detection on raw bands or on a derived index. The Esri page notes that continuous change detection can be performed on an NDVI raster, because abrupt changes in that index can indicate deforestation or other sudden vegetation loss, which is why the CCDC documentation uses it as an example.
The LandTrendr tool segments only one band, so the choice matters. Near-infrared may suit vegetation, while the Modified Normalized Difference Water Index, which uses the green and SWIR bands, may suit open water, as the LandTrendr help page suggests.
Indices also travel reasonably well between sensors. The HLS evaluation found the between-sensor vegetation index difference below 4.5% for most of the indices tested, as reported in the USGS record.
Time-series methods: CCDC and LandTrendr
Two images show that something changed; a series shows when and how. CCDC fits harmonic regression and trend models to the pixel history, and sudden deviations from those models point to abrupt change, per the Esri CCDC page.
A pixel is flagged when the gap between modeled and observed values is three times the RMSE. A single odd reading is treated as an outlier, and the pixel counts as changed only after a given number of consecutive observations, as the same page describes.
If you would rather not run the algorithm yourself, Google’s Earth Engine catalog hosts precomputed CCDC results built from 20 years of Landsat surface reflectance data. The catalog entry warns against using it at coarse zoom levels, because the downsampled results no longer represent the full-resolution data.
LandTrendr takes another view. It breaks a pixel’s history into linear segments to describe disturbance, a short period of change, and recovery, the longer return toward the original state, such as after a forest fire, according to the LandTrendr help page.
It needs only one image for each year, and at least six years of data is recommended. If a year has multiple images, preferably from a single season, you can remove cloud and shadow from each and combine them into one representative image, per the same documentation.
| Feature | CCDC | LandTrendr |
|---|---|---|
| Image needs | A stack of images over time | One image per year, at least six years |
| What it models | Seasonal and trend behavior, with break points | Linear segments between vertices |
| Best suited to | Abrupt change such as deforestation | Disturbance followed by recovery |
The comparison draws on Esri’s help pages for CCDC and LandTrendr; tool settings may change, so confirm them there before you rely on them.

Checking results before you trust them
A change map is a set of candidates, not a verdict. Before acting on one, read up on how to interpret satellite images so you can tell haze, shadow and seasonal colour shifts from genuine surface change.
A practical review routine looks like this:
- Compare images from the same season, or use a method that models seasons explicitly.
- Mask clouds, cloud shadow and snow before any comparison.
- Prefer harmonized or otherwise consistent sensor data.
- Pick one band or index that matches the phenomenon you care about.
- Inspect flagged spots against the original imagery.
The last step matters because model output is hard to read directly. The Esri page says change analysis rasters are difficult to interpret visually, and suggests temporal profile charts to explore how a pixel moved over time, as the CCDC page notes.
Frequently asked questions
Can I compare Landsat and Sentinel-2 images directly?
Show answer
Raw products can disagree slightly, so harmonized data is the safer route. The HLS authors conclude that their harmonization is robust and that the data suit quantitative terrestrial applications, according to the USGS publication record.
Why does a cloud look like land change?
Show answer
An algorithm sees only pixel values, and a bright cloud looks like a sudden jump in the surface. That is why masking is treated as a prerequisite for reliable land cover change mapping in the USGS-hosted cloud masking study.
How many years of images do I need?
Show answer
It depends on the method. LandTrendr works from annual composites, and its help page recommends a record of at least six years, while CCDC is built to use many dates, per the LandTrendr documentation.
Can I track water instead of vegetation?
Show answer
Yes, by swapping the index. The LandTrendr help page points to MNDWI for open water features, since it draws on the green and SWIR bands, as described on its documentation page.
Is there a ready-made global change dataset?
Show answer
Google’s catalog offers precomputed Landsat-based CCDC segments. Its entry states there are no current plans to add later assets, so the record ends where the catalog says it does, per the Earth Engine catalog.



