Skip to content
Shyenova

Remote Sensing

Change Detection

Every change map is really two classifications, so its errors compound. Most published ones do not admit this.

Land questions are almost always change questions. Not what is there, but what is there now that was not there in 2016, how far the settlement has spread, how much of the buffer has been built on, where the water body has retreated to, which parcels have been encroached.

Satellite archives answer these well because they are unusually complete: Landsat back to the 1980s, Sentinel-2 at five-day revisit since 2015, all free. The methodological trap is that a change map is the difference between two classifications, so a pair of maps at 90% accuracy can produce a change map at considerably less. Handling that honestly is most of the craft.

How the work runs

Four phases, and what comes out of each.

Durations are indicative and assume reasonable data availability. Where data is poor we say so in the first fortnight rather than modelling on sand.
01

1–2 weeks

Define the classes against the decision

Classification schemes are usually inherited from a standard and then found unusable, because the categories that matter to the decision do not map onto them. We define classes that are both separable in the imagery and meaningful to the person acting on the output, which frequently means fewer classes than requested, and we explain why merging two of them will produce a more trustworthy result.

Outputs

  • Classification scheme tied to the decision it supports
  • Baseline and comparison epochs selected against data availability
  • Minimum mapping unit agreed and justified
02

2–4 weeks

Build a consistent multi-date stack

Change detection is unforgiving of inconsistency. Imagery across epochs is radiometrically harmonised, co-registered, and selected for comparable season and phenology, because a February scene compared against an August scene will show enormous spurious change. Where sensors differ across a long time series, cross-calibration is done explicitly and its residual error carried forward.

Outputs

  • Co-registered, radiometrically consistent multi-date stack
  • Season-matched scene selection with the rationale documented
  • Cross-sensor calibration where the series spans platforms
03

3–5 weeks

Classify, detect change, and quantify the error

We classify each epoch, derive change, then assess accuracy on the change product itself rather than only on the individual classifications, because that is what will be used. Where change is legally or commercially consequential, high-confidence detections are separated from marginal ones so that field verification effort goes where it is worth spending.

Outputs

  • Classified maps per epoch with per-class accuracy
  • Change matrix quantifying every transition, with area statistics
  • Confidence-tiered change polygons for field verification
04

2–3 weeks

Make it repeatable

A one-off assessment ages immediately. We hand over the pipeline so the analysis can be re-run on any future epoch by your team, with the classification model, training data and validation protocol documented. Where the need is ongoing, this becomes a monitoring service that flags new change rather than a report that gets reproduced from scratch.

Outputs

  • Documented, re-runnable pipeline with training data retained
  • Change alerting on new detections above a defined threshold
  • Training for the team that will operate it

What good looks like

The ranges we design toward.

These are the improvement ranges these levers are known to produce where the underlying practice has not been examined recently. They are targets to test against your baseline, not promises.

88–94%

Classification accuracy

Typical overall accuracy for a well-specified scheme of six to ten separable classes.

30+ years

Historical depth

Available at 30m from the Landsat archive, at no data cost, for most of the world.

5 days

Monitoring cadence

Sentinel-2 revisit, cloud permitting, against the annual or multi-year cycle of ground assessment.

Deliverables

What you are left holding.

All of it handed over, documented, and with a session for whoever inherits it.
  1. 01Classified land use maps for each epoch, with per-class producer and user accuracy
  2. 02Change matrix quantifying every transition with area statistics and confidence
  3. 03Confidence-tiered change polygons prioritised for field verification
  4. 04Re-runnable pipeline, training data and documentation handed to your team
  5. 05Ongoing change alerting where monitoring rather than assessment is the requirement

What an engagement looks like

A representative shape.

Illustrative. This is the shape of work in remote sensing, not an account of a specific client. We do not publish named results, and we would be wary of a firm that publishes them without the client in the room.

16 weeks

  1. 01Situation

    An agri-processor was committing to procurement volumes months before it had any reliable view of planted acreage, relying on dealer and field-officer reporting whose historical error nobody had measured.

  2. 02What we did

    Built a Sentinel-1 and Sentinel-2 time series across the sourcing districts, trained crop classification on collected ground truth, and ran the satellite estimate in parallel with the incumbent reporting for a full season.

  3. 03Where it landed

    The satellite estimate was available several weeks earlier and diverged from field reporting by a consistent margin in two districts, which ground checks confirmed. The parallel season is what earned the commercial team's trust; the technology alone would not have.

Where this connects

Rarely a self-contained problem.

These are the adjacencies that come up most often on this kind of work. Naming them is usually cheaper than discovering them halfway through.

Questions

The ones we get asked.

How far back can you go?

To 1984 at 30m resolution using the Landsat archive, which is free and global. Detail before the mid-1980s is limited and patchy. Sentinel-2's finer 10m data begins in 2015, so a long time series usually means a coarser one, and where a study spans both, the sensor change has to be calibrated for explicitly rather than assumed away.

Can this be used as legal evidence of encroachment?

It is generally admissible as corroborating evidence and rarely sufficient on its own. Satellite change detection has a stated error rate, and a defendant's counsel will find it. We structure the output for exactly this: high-confidence detections separated from marginal ones, full provenance for every scene, and a documented methodology. It should direct where a surveyor goes, not replace them.

What does 90% accuracy actually mean here?

That roughly one pixel in ten is assigned the wrong class, but the errors are not spread evenly, which is the part that matters. Accuracy is typically high for water and dense built-up, and much lower at class boundaries and for spectrally similar pairs like fallow cropland versus bare ground. We report per-class producer and user accuracy rather than a single headline number, because the headline hides precisely the failure you are most likely to encounter.

We already have a GIS department. What would we use you for?

Usually the classification and validation methodology rather than the mapping. Many capable GIS teams are strong at spatial data management and have never had to defend a classification accuracy assessment or handle error propagation through a change product. We frequently do one epoch with the team and hand the rest over.

Start a conversation

Talk to us about change detection.

Most useful first conversations are thirty minutes and involve no proposal. Describe the problem you are looking at and we will tell you honestly whether it is one we can help with.

Or write directly to hello@shyenova.com