Change Detection from Space: Three Approaches That Actually Work

The Core Idea

Satellite change detection is conceptually simple: compare images from two dates and identify what's different. In practice, it's riddled with pitfalls — atmospheric differences, seasonal variation, misregistration, and the fundamental question of whether an observed change is real or an artifact.

After working with change detection across diverse applications — deforestation tracking in Southeast Asia, post-earthquake damage assessment, coastal erosion monitoring — I've found that method selection matters more than algorithmic sophistication. Picking the right approach for your specific question is 80% of the work.

Approach 1: Image Differencing

The most straightforward method. Subtract one image from another, and non-zero values indicate change.

Change = Image_T2 - Image_T1

For optical data, this might mean differencing NDVI values, specific band reflectances, or even simple brightness. For SAR, you'd difference calibrated backscatter (σ⁰) values.

When it works well

• Flood extent mapping: Pre-flood SAR minus post-flood SAR clearly delineates inundated areas as strongly negative difference values (backscatter drops when land becomes water)
• Burn scar detection: Post-fire optical imagery shows dramatically lower NIR reflectance, making the difference signal strong and unambiguous
• Large-scale deforestation: Clear-cut logging creates a stark contrast in both optical and SAR data

When it fails

• Gradual changes: If the change happens slowly over months, the difference between any two adjacent dates may be too small to reliably detect
• Seasonal confusion: A deciduous forest will look dramatically different between summer and winter without anything actually "changing" in the meaningful sense. Always compare the same season across years
• Atmospheric differences: Two optical scenes acquired under different atmospheric conditions will produce a non-zero difference everywhere, even if nothing on the ground changed

Practical tip

Normalizing the difference helps. Instead of simple subtraction, use:

Normalized Difference = (T2 - T1) / (T2 + T1)

This ratio partially compensates for absolute brightness differences and produces values between -1 and +1, making threshold selection more consistent across scenes.

Approach 2: NDVI Time-Series

Rather than comparing just two dates, track NDVI continuously over months or years. This approach excels for vegetation-related questions because NDVI has well-understood seasonal patterns that serve as a baseline.

The key insight

Healthy vegetation follows a predictable annual NDVI cycle. In temperate regions, NDVI rises in spring, peaks in summer, and declines in autumn. Deviations from this expected pattern indicate something happened — drought stress, disease, harvest, or land use change.

What to look for

• Sudden drops: An abrupt NDVI decrease outside the normal seasonal pattern suggests disturbance — logging, fire, flooding, or crop failure
• Failed green-up: If spring NDVI doesn't rise to historical levels, the vegetation may be damaged or the land use may have changed
• Permanent baseline shift: A step change that doesn't recover indicates land cover conversion (e.g., forest to farmland)

Limitations

NDVI time-series requires cloud-free optical imagery at regular intervals. In tropical regions with persistent cloud cover, gaps of weeks or months are common, making continuous monitoring difficult. This is where SAR becomes the better choice.

Sentinel-2's 5-day revisit cycle helps, but cloud contamination still reduces the actual usable frequency to perhaps once every 2-4 weeks in many areas.

Approach 3: SAR Coherence

This is the least intuitive approach but potentially the most powerful for certain applications. It exploits the phase information in SAR data that most users never look at.

What coherence measures

When the SAR satellite revisits the same area, the phase of the returned signal should be similar if nothing on the ground has changed. Coherence quantifies this phase consistency on a scale from 0 (completely different) to 1 (identical).

High coherence = stable surface. Low coherence = something changed between acquisitions.

What causes coherence loss

• Physical displacement: Ground subsidence, building collapse, or landslide
• Surface reworking: Plowing, construction activity, or flooding
• Vegetation growth: Growing vegetation changes the scattering geometry, reducing coherence even in the absence of disturbance

That last point is important. In vegetated areas, coherence tends to be naturally low because plants grow and move between acquisitions. Coherence-based change detection works best over urban areas, bare rock, and other stable surfaces.

Where coherence excels

Post-earthquake damage assessment is a textbook application. Undamaged buildings maintain high coherence; collapsed or tilted structures lose it. This can be mapped within hours of a SAR acquisition — far faster than optical damage assessment, which requires clear weather and often manual interpretation.

Choosing Your Method

Question	Best Approach	Data Source
Where did flooding occur?	Image differencing (SAR)	Sentinel-1 VV
Is this forest degrading over time?	NDVI time-series	Sentinel-2
Was this area burned?	Image differencing (optical)	Sentinel-2 NBR
Which buildings were damaged?	SAR coherence	Sentinel-1 SLC
Is cropland being converted?	NDVI time-series	Sentinel-2
Where is illegal mining happening?	Image differencing + coherence	Both

Most real-world projects benefit from combining approaches. Use SAR change detection for rapid initial assessment (it works regardless of weather), then refine with optical analysis when clear imagery becomes available.

Try It

Off-Nadir Delta supports multi-temporal satellite imagery comparison. Load Sentinel-1 or Sentinel-2 data from two different dates over the same area and visually compare them. For a hands-on start, pick an area where you know something changed — a recent construction site, a harvested field, or a fire scar — and build your interpretation skills from there.