Change Detection with Remote Sensing: Tracking Urban Expansion Over Time
Urban areas are among the fastest-changing landscapes on Earth. Cities grow outward into agricultural land and forests. Peri-urban zones shift from rural to built-up within a single decade. Infrastructure extends into areas that satellites imaged as open land just years before.
Remote sensing is one of the few tools capable of capturing this change systematically, repeatedly, and at scale. Change detection, the process of identifying differences in land surface conditions between two or more points in time, is a core application of satellite imagery in GIS and urban planning. This article walks through the concepts, methods, data sources, and workflows that GIS analysts use to track urban expansion using remote sensing.
Why Urban Expansion Matters
Urban expansion is not just a planning concern. It has cascading effects on hydrology, heat distribution, biodiversity, carbon stocks, and social equity. Impervious surfaces replace permeable land, increasing runoff and flood risk. Urban heat islands intensify. Green corridors are fragmented.
Tracking where cities are growing, how fast, and at the expense of what land cover types is foundational to evidence-based urban governance. Remote sensing makes it possible to do this at city, regional, and global scales with consistent, reproducible methodology.
What is Change Detection?
Change detection is the process of quantifying and locating differences between two or more observations of the same geographic area taken at different times. In the context of urban expansion, it answers questions like:
- Where has built-up area expanded between 2005 and 2024?
- Which agricultural zones have been converted to residential use?
- How has impervious surface coverage changed in a watershed?
- What is the annual rate of urban growth in a metropolitan region?
Change detection is not limited to urban analysis. It is also used for deforestation monitoring, flood extent mapping, post-fire damage assessment, and coastal erosion tracking. But urban expansion is one of its most common and consequential applications.
Key Principles Before You Start
Radiometric Consistency
For change detection to produce meaningful results, the imagery used at each time step must be radiometrically comparable. Differences caused by sensor calibration, solar angle, atmospheric conditions, or seasonal variation can produce false detections.
Always use surface reflectance products (also called Level-2 products) rather than raw digital numbers or top-of-atmosphere reflectance. Surface reflectance has been corrected for atmospheric effects and is the standard for multi-temporal analysis.
Phenological Consistency
Vegetation changes with season. If you compare a dry-season image to a wet-season image, you will detect vegetation stress and growth rather than land cover change. Where possible, acquire images from the same season or the same phenological window across years.
For urban expansion analysis, the goal is to compare imagery from the same time of year, ideally within a few weeks of each other across the date range you are studying.
Sensor Consistency
Mixing sensors with different spatial resolutions, band configurations, or radiometric sensitivities introduces noise. Landsat-to-Landsat and Sentinel-2-to-Sentinel-2 comparisons are more reliable than cross-sensor comparisons. When cross-sensor analysis is necessary, use harmonized products such as the Harmonized Landsat Sentinel-2 (HLS) dataset produced by NASA.
Data Sources for Urban Change Detection
Landsat Archive
Landsat provides the longest consistent Earth observation archive available, with global coverage going back to 1972. This makes it the only free option for studying urban expansion over multi-decadal periods.
Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 and 9 OLI all provide 30-meter multispectral imagery. USGS Collection 2 provides surface reflectance products that are cross-calibrated across missions, making long-term time series analysis significantly more reliable than earlier collections.
For urban change detection, the key bands are the shortwave infrared (SWIR) bands, near-infrared (NIR), and visible red. These bands are sensitive to built-up surfaces, bare soil, and vegetation, which are the dominant land cover classes in urban transition zones.
Sentinel-2
For analysis from 2015 onward, Sentinel-2 provides 10-meter multispectral imagery with a 5-day revisit cycle. Its higher spatial resolution makes it better suited for detecting fine-grained urban change at the neighborhood or parcel level.
Sentinel-2 Level-2A products (surface reflectance) are available globally from the Copernicus Data Space Ecosystem and from platforms like Google Earth Engine and Microsoft Planetary Computer.
MODIS and VIIRS
MODIS and VIIRS are better suited to regional and continental-scale urban monitoring where high spatial resolution is less important than temporal frequency and geographic coverage. MODIS 500-meter land cover products provide annual global land cover classifications that can be used to track broad-scale urban expansion trends.
Global Urban Datasets
Several analysis-ready global datasets exist specifically for urban change analysis:
- Global Human Settlement Layer (GHSL): Produced by the European Commission Joint Research Centre. Provides built-up surface maps at 10-meter resolution for 1975, 1990, 2000, 2015, and 2020. Also includes population grids and settlement classification.
- World Settlement Footprint (WSF): Produced by DLR (German Aerospace Center). Global binary built-up / non-built-up classification at 10-meter resolution using Sentinel-1 and Sentinel-2.
- Impervious Surface datasets: NLCD (National Land Cover Database) for the United States provides impervious surface percentage at 30-meter resolution for multiple epochs.
These datasets are useful as baselines, validation layers, or direct inputs for change analysis rather than starting from raw imagery.
Methods for Change Detection
1. Image Differencing
Image differencing is the simplest and most widely used change detection method. It involves subtracting the pixel values of one image from another taken at a later date. Pixels where little has changed produce values near zero. Pixels where significant change has occurred produce high positive or negative values.
A common application in urban analysis is NDVI differencing. NDVI (Normalized Difference Vegetation Index) responds strongly to vegetation removal, which often precedes urban development. Areas of significant NDVI decrease between two dates are candidates for built-up conversion.
The limitation of simple differencing is sensitivity to noise. Thresholding is required to separate real change from radiometric variation, and selecting thresholds is often subjective or iterative.
2. Band Ratio and Index-Based Methods
Beyond NDVI, several spectral indices are informative for urban change detection.
NDBI (Normalized Difference Built-up Index):
NDBI uses SWIR and NIR bands to highlight built-up surfaces. It tends to produce high values over urban areas and low values over vegetation.
Formula: NDBI = (SWIR – NIR) / (SWIR + NIR)
Comparing NDBI across time steps highlights areas transitioning from non-urban to urban land cover.
UI (Urban Index):
The Urban Index is another SWIR-NIR based index specifically designed to separate built-up surfaces from bare soil and vegetation.
BCI (Biophysical Composition Index):
BCI decomposes land cover into three biophysical components: high-albedo surfaces (urban), vegetation, and low-albedo surfaces (water and dark impervious). It is more robust than NDBI in complex urban environments where bare soil can be confused with built-up land.
3. Post-Classification Comparison
Post-classification comparison is the most interpretable and widely applied method for urban change detection. The workflow involves:
- Classifying imagery independently for each time step (e.g., 2000 and 2020)
- Producing land cover maps for each date
- Overlaying the two classified maps
- Computing a change matrix that shows which classes transitioned to which
The change matrix (also called a transition matrix or cross-tabulation) reveals not just where change occurred, but the nature of that change. You can identify specifically how much agricultural land was converted to built-up area, how much open land became commercial, and how much forest was replaced by residential development.
The limitation of this approach is that classification errors at each time step are compounded. If your individual classifications are 85% accurate, the change product will be less accurate than either input alone.
4. Continuous Change Detection (CCDC)
Continuous Change Detection and Classification (CCDC), developed by researchers at Boston University, is an algorithm that models pixel-level spectral trajectories using time series data. Rather than comparing two images, it fits seasonal harmonic models to the full Landsat archive for each pixel and detects anomalies that indicate actual land cover change rather than seasonal variation.
CCDC is available in Google Earth Engine and represents the state of the art for robust, date-specific change detection from dense Landsat time series. It produces both change maps and land cover maps for any point in time within the archive.
5. Deep Learning Approaches
Machine learning and deep learning methods are increasingly applied to change detection from satellite imagery. Siamese neural networks, which process bitemporal image pairs through shared weight architectures, have shown strong performance on urban change detection benchmarks.
These methods require labeled training data and computational infrastructure but can generalize well across cities and sensors when trained appropriately. Frameworks like PyTorch and TensorFlow, combined with geospatial libraries like rasterio, torchgeo, and segmentation-models-pytorch, are commonly used in this space.
Step-by-Step Workflow: Urban Expansion Analysis with Sentinel-2
The following outlines a practical workflow for detecting urban expansion between two dates using Sentinel-2 imagery and a post-classification comparison approach.
Step 1: Define the Study Area and Time Period
Define a bounding box or polygon for your area of interest. Select two target dates separated by the time period you want to analyze. Aim for the same seasonal window in both years to minimize phenological differences.
Step 2: Acquire Surface Reflectance Imagery
Use the Copernicus Data Space Ecosystem, Google Earth Engine, or Microsoft Planetary Computer to acquire Sentinel-2 Level-2A imagery for both dates. Filter for low cloud cover (ideally below 10 percent). If a single scene is cloud-free, use it. If not, create a cloud-masked median composite from a short acquisition window around your target date.
Step 3: Compute Spectral Indices
For each date, compute the following indices:
- NDVI: (NIR – Red) / (NIR + Red) — captures vegetation
- NDBI: (SWIR1 – NIR) / (SWIR1 + NIR) — highlights built-up surfaces
- MNDWI: (Green – SWIR1) / (Green + SWIR1) — isolates water
These indices will serve as input features for classification.
Step 4: Classify Land Cover for Each Date
Using training samples collected from high-resolution imagery or field knowledge, classify each image into land cover classes. Typical classes for urban expansion analysis:
- Built-up / urban
- Vegetation (dense, sparse)
- Bare soil / fallow land
- Water
- Agriculture
Random Forest classifiers work well for this task and are available in Python via scikit-learn, in Google Earth Engine, and in ArcGIS Pro through the Image Classification Wizard or the TrainRandomTreesClassifier tool.
Step 5: Validate Each Classification
Assess the accuracy of each classified map using a withheld validation sample. Compute a confusion matrix and report overall accuracy, producer’s accuracy, user’s accuracy, and the kappa coefficient. Aim for overall accuracy above 85 percent before proceeding to change analysis.
Step 6: Compute the Change Matrix
Overlay the two classified maps using a GIS union or raster calculation. Each pixel gets a combined class label indicating its land cover at time 1 and time 2. Summarize the area of each transition type. Identify the area that transitioned from non-urban classes to built-up between the two dates.
Step 7: Visualize and Interpret Results
Produce a change map highlighting:
- Areas that converted to urban (typically shown in red)
- Areas that remained urban (grey)
- Areas that remained non-urban (white or muted tones)
- Optional: areas that were urban at time 1 but show vegetation recovery at time 2
Annotate the map with growth direction vectors, proximity to existing urban cores, and the land cover types that were displaced.
Working with Urban Change Detection in ArcGIS Pro
ArcGIS Pro provides several tools directly applicable to urban change detection workflows.
The Image Classification Wizard walks through training sample collection, classifier training, and classification in a guided interface. It supports Random Forest, Support Vector Machine, and deep learning classifiers.
The Compute Change Raster tool in the Raster Functions panel computes pixel-level difference between two rasters. It can operate on raw bands, spectral indices, or classified outputs.
The Zonal Statistics and Tabulate Area tools allow you to summarize change by administrative boundary, watershed, or custom zone.
For time series analysis, the Space-Time Pattern Mining toolbox supports trend detection and hotspot analysis on temporal datasets.
The Living Atlas contains ready-to-use urban change and land cover layers including Esri’s Land Cover (derived from Sentinel-2), which provides annual global land cover classification at 10-meter resolution from 2017 onward. This can be used directly for multi-year urban change analysis without building a classification from scratch.
Common Challenges and How to Address Them
Spectral confusion between bare soil and built-up surfaces. Both classes have high SWIR reflectance and low NIR reflectance. Address this by incorporating texture features, SAR data (Sentinel-1 backscatter), or temporal information. Bare soil often shows seasonal variation while built-up surfaces remain spectrally stable over time.
Cloud contamination in tropical regions. Persistent cloud cover makes single-image analysis unreliable. Use cloud-masked composites from longer acquisition windows. Sentinel-1 SAR, being cloud-independent, is a valuable complement to optical imagery in these regions.
Mixed pixels in transition zones. At 10 or 30-meter resolution, pixels at the urban fringe often contain a mix of built-up and non-built-up land. Sub-pixel classification methods or spectral mixture analysis can improve accuracy in these areas.
Geometric misregistration. If two images are not precisely co-registered, apparent change can appear at boundaries of objects even when no real change occurred. Ensure both images use the same coordinate reference system and that geometric correction is applied consistently.
Interpreting Urban Growth Patterns
Urban expansion rarely follows a single pattern. Remote sensing analysis can distinguish between several growth modes:
Infill development occurs within existing urban boundaries. Vacant lots, low-density areas, and brownfields are converted to higher-density uses. This shows up in change analysis as new impervious surfaces appearing within the urban core.
Edge expansion is the most common mode globally. New built-up areas extend outward from the urban perimeter, typically along transportation corridors. This produces a characteristic halo of new development around the existing urban footprint.
Leapfrog development results in new urban clusters appearing at a distance from the main urban area, separated by non-urban land. This is common in peri-urban zones near highways or planned developments. It shows up in change analysis as isolated new built-up patches that are not contiguous with the existing urban footprint.
Understanding which growth pattern dominates in a study area informs infrastructure planning, environmental impact assessment, and policy intervention.
Final Thoughts
Change detection for urban expansion is one of the most impactful applications of remote sensing in applied GIS. The combination of freely available Landsat and Sentinel-2 archives, cloud-based processing platforms, and robust classification and analysis tools has made this type of analysis accessible to analysts at every level.
The critical skills are not the algorithms themselves. They are the judgment to select appropriate imagery, the rigor to apply consistent preprocessing, the knowledge to choose a method suited to the scale and context of the analysis, and the ability to interpret results in terms that are meaningful to planners, policymakers, and communities.
Urban expansion is ongoing. The satellite record is accumulating. The analysts who understand how to read that record will play an important role in shaping how cities grow.