Change Detection

Change detection is a computer vision technique for identifying meaningful differences between observations of the same scene captured at different times. Instead of asking only what appears in an image, it asks what appeared, disappeared, moved, grew, shrank, or changed state. A system might compare two photographs, consecutive video frames, or a longer time series. Its output can be a binary change map, labeled regions, altered objects, or measurements such as the area of newly flooded land.

How Change Detection Works#

A typical workflow begins with two or more images of the same area. Bitemporal change detection compares two dates, while multitemporal change detection analyzes a sequence to distinguish isolated events from gradual trends.

Before comparison, the images must be made compatible:

1. Alignment: Image registration places corresponding scene points at the same pixel coordinates. Even a small camera shift can otherwise create false edges around every object.
2. Normalization: Brightness, color, atmospheric conditions, and sensor responses are adjusted. For satellite analysis, surface reflectance products improve comparisons by reducing atmospheric effects.
3. Representation: The system compares raw pixels, extracted visual features, detected objects, or segmentation masks. A basic method may use OpenCV absolute image differences, while learned models can recognize semantic changes despite modest lighting or viewpoint differences.
4. Decision: A threshold separates meaningful change from noise. Similarity measures such as those in the scikit-image image metrics API can quantify how strongly two observations differ.
5. Postprocessing: Small isolated detections may be removed, neighboring pixels merged into regions, and changes assigned labels such as “new building” or “missing component.”

The minimal detectable change is the smallest alteration a system can reliably separate from sensor noise, registration error, and normal variation. It depends on image resolution, object size, contrast, preprocessing quality, and the selected threshold.

Related Concepts and Key Distinctions#

Change detection overlaps with several vision tasks, but each answers a different question:

• Anomaly detection identifies observations that differ from a learned definition of normal, sometimes using only one image. Change detection requires a temporal or baseline comparison and does not imply that the change is abnormal.
• Object detection locates and classifies objects independently in each image. Comparing detections across dates can reveal added or removed objects.
• Image segmentation assigns labels at the pixel level. Change detection can compare segmentation masks to measure altered shapes or areas precisely.
• Object tracking maintains object identities across video frames and emphasizes motion and trajectories. Change detection usually focuses on changes in scene content or state, which may occur over minutes, months, or years.
• Data drift describes changes in the statistical distribution of production inputs. It concerns model reliability, whereas visual change detection concerns differences within observed scenes.

Real-World Applications#

Earth observation and disaster assessment: Analysts compare satellite images to map deforestation, urban expansion, coastline movement, wildfire damage, or flooding. Satellite image analysis may combine multispectral data with segmentation to classify each changed region. The USGS Earthshots collection demonstrates how before-and-after Landsat imagery reveals environmental transformation, while NASA provides a practical log-difference change detection workflow. Long-term programs such as NOAA Coastal Land Cover and Change convert these comparisons into information for flood planning and coastal management.

Industrial inspection: A fixed camera can compare an assembly before and after a production step. Detection may reveal a missing bolt or newly placed component, while segmentation can measure a crack or damaged surface. In automated manufacturing visual inspection, an undetected change can allow a defective product to continue downstream, while false alarms can cause unnecessary line stops and manual reviews.

Change Detection with Ultralytics YOLO#

Ultralytics YOLO can provide the semantic perception stage of a change detection pipeline. For example, a custom YOLO26 segmentation model can process registered before-and-after images through YOLO predict mode:

from ultralytics import YOLO

# Load a segmentation model trained for relevant objects or regions
model = YOLO("yolo26n-seg.pt")

# Run inference on an aligned image pair
sources = ["before.jpg", "after.jpg"]
results = model(sources)

# Save semantic predictions for comparison
output_names = ["before_predictions.jpg", "after_predictions.jpg"]

for result, output_name in zip(results, output_names, strict=True):
    result.save(filename=output_name)

This workflow produces object masks and labels for each observation. A downstream application can match corresponding classes and compare their presence, position, or mask area. For specialized categories, teams can annotate paired imagery, train custom models, and manage deployment using the Ultralytics Platform.

Practical Challenges and Best Practices#

Reliable change detection requires more than subtracting images. Camera movement, shadows, seasonal vegetation, reflections, clouds, and occlusion can resemble real change. Teams should keep viewpoints and sensor settings consistent, register images before comparison, train with representative “no-change” variation, and select thresholds using validation data.

Evaluation should measure both false positives, such as shadows labeled as structural changes, and false negatives, such as small defects being missed. The appropriate balance depends on consequences: disaster mapping may prioritize recall, while a high-volume inspection line may need tighter precision to avoid excessive alarms. Human review remains valuable for ambiguous or high-impact results.