L10 - Change detection

1. Context¶

L09 introduced image translation: transforming one image into another, generated, image. This lesson compares two real observations of the same place instead. Change detection takes imagery from two or more dates and identifies where the surface has genuinely changed, producing a binary mask, a categorical map, or a continuous measure of change intensity.

This is also the point in the course where you start comparing images to each other rather than analyzing a single scene, a shift that carries forward into later, more advanced project workflows. The interpretation habits from L09, checking whether a difference reflects a real change versus something else entirely, apply directly here too.

2. Motivation¶

Many of the most consequential geospatial questions are fundamentally about change: how fast is a city expanding, how much forest was cleared this year, how far did a flood reach, how much of a coastline eroded. Answering any of these requires comparing imagery across time, not just describing a single snapshot.

Change detection is also deceptively easy to get wrong. Two images of the same location taken months apart differ not only because the surface changed, but because of sun angle, atmospheric conditions, sensor calibration, and seasonal vegetation cycles. A substantial part of this lesson is about learning to tell these apart, since a change detection workflow that cannot distinguish real change from irrelevant variation produces results no one should trust.

3. Learning objectives¶

By the end of this lesson, you should be able to:

• Decide when change detection is the right tool, and what form of output (binary, categorical, or continuous) your project actually needs

• Explain what makes two images genuinely comparable, and identify the main obstacles when they are not

• Apply traditional, interpretable baseline methods: image differencing and change vector analysis

• Explain why deep learning models, particularly siamese networks, can outperform pixel-wise comparison

• Run a deep learning change detection workflow with geoai, and interpret its full set of outputs

• Adjust detection thresholds deliberately and export results as GIS-ready files

4. Lesson roadmap¶

1. When change detection fits — typical use cases and the three output types you can choose between.

2. Change types and challenges — what kinds of change exist, and what makes two images actually comparable.

3. Traditional methods — image differencing and change vector analysis on Landsat imagery, as interpretable baselines.

4. Deep learning approaches — why spatial context helps, and how siamese networks compare image pairs.

5. The ChangeStar workflow — the main hands-on workflow: detecting building change in Las Vegas NAIP imagery.

6. Reading the results — interpreting ChangeStar’s full output: change maps, probabilities, and per-date building masks.

7. Thresholds and exports — comparing model variants, tuning sensitivity, and exporting GIS-ready files.

8. Project transfer — designing a change-detection workflow for your own project.

Pages 1–2 build the conceptual foundation, pages 3–4 cover two different approaches (traditional and deep learning), page 5 is the main hands-on workflow, and pages 6–7 extend it with interpretation, tuning, and export. Page 8 transfers everything to your project.

5. Project framing¶

While reading this lesson, keep your own project in mind and consider:

• Does your project need to know that something changed, what kind of change occurred, or how much change happened?

• What two (or more) dates would you realistically compare, and are they close enough in season to be genuinely comparable?

• Would a simple, interpretable baseline (image differencing or CVA) already answer your question, or does your project need the spatial context a deep learning model provides?

• How would you validate a change detection result without exhaustive ground-truth data?

6. Before class¶

• Review the L07 semantic segmentation lesson, since ChangeStar’s per-date outputs are segmentation masks.

• Make sure geoai is installed and importable in your environment.

• Think of one change process relevant to your own project area (growth, loss, damage, or a seasonal or disaster-related change).

7. After this lesson¶

After working through this lesson, you should have:

• A clear decision on whether change detection fits your project, and which output type it should produce

• A working understanding of both traditional and deep-learning change detection, and when each is the more appropriate choice

• A first sketch of your own change-detection plan: dates, comparability checks, method, threshold, and validation approach