Satellite change detection starts with a stable polygon rulebook

Satellite image change detection can look simple on a map: draw a polygon around a new road, cleared field, or expanding settlement. The hard part is making that label mean the same thing across years, annotators, and image conditions. If the rules shift between passes, the dataset may capture annotation drift rather than land-use change.

Teams planning remote sensing annotation should settle three questions early: what counts as change, what spatial unit gets labeled, and how to record uncertainty or missing evidence. Those choices matter more than choosing a particular annotation workflow.

Choose the target before drawing boundaries

There are two common targets that are easy to conflate. A land-cover dataset assigns a class to a location at a given time, such as water, vegetation, or built-up area. A change-detection dataset describes a transition between observations, such as vegetation-to-built-up. If the project needs both, define both outputs. A single label called “changed” will not tell a downstream team what changed or in which direction.

Specify the comparison window for each record: the dates or date ranges being compared, and which observation is the baseline. Decide whether a transition counts only when it persists in a later pass, or whether one observed difference is enough. There is no universal threshold. The right rule depends on the use case and the reliability of the imagery.

Keep classes observable. If annotators cannot distinguish two land-use categories consistently at the source resolution, combine them or provide a separate uncertain state. Do not turn low confidence into a forced class merely to make the schema look complete.

Write polygon rules that survive multiple dates

Define whether each polygon represents a full land-cover area, only the changed portion, or a transition zone. Then set a minimum mapping unit: the smallest feature that should be labeled. Without one, small patches may be included by one annotator and omitted by another.

Boundary rules need examples. State whether polygons follow visible edges, parcel boundaries, or another reference, and how to handle mixed pixels, narrow features, and gradual transitions. Decide whether adjacent areas with the same class should be merged, and whether overlapping polygons are allowed. For paired imagery, describe how to treat a boundary that is clear in one year but obscured in the other.

These decisions are especially important for GIS image segmentation. A neat polygon is not necessarily a defensible label. The guideline should explain what evidence supports the edge, not just show an idealized outline.

Make missing evidence a label state

Cloud, cloud shadow, haze, seasonal differences, sensor or resolution changes, and no-data gaps can make a location unobservable. Treating those areas as “no change” introduces false negatives. Define explicit states for unusable or uncertain evidence, and decide whether affected regions are excluded, partially labeled, or flagged for review.

Metadata should preserve the context needed to interpret each comparison. At minimum, teams should decide which fields their analysis requires: image dates, source or scene identifiers, comparison-pair identifier, coordinate reference system, class or transition code, and any confidence or review status. Use controlled values and validate required fields at delivery. If acquisition conditions or processing versions matter to the project, include them in the schema rather than relying on notes outside the dataset.

Choose a production model to fit the risk

An in-house team suits organizations that need close control over interpretation, frequent guideline changes, or direct links between labeling and domain experts. The trade-off is management work: recruiting, training, calibration, review, and keeping instructions consistent across batches. That cost is not captured by an annotation rate alone.

A general-purpose annotation provider can suit a project with a stable schema and substantial volume, particularly when the buyer already has clear remote sensing rules and staff to audit output. Before scaling, test whether the team can follow the project’s polygon and metadata conventions on difficult examples, not only on clean scenes.

A geospatial-focused partner may be a better fit when segmentation, spatial structure, or specialized review is central to the work. Still, “geospatial” is not a substitute for a task-specific pilot. Ask for a small sample using the actual date pairs, class definitions, and edge cases. Compare labels against your acceptance criteria and track disagreements before increasing volume.

XYNTRIQ lists remote-sensing segmentation and satellite imagery labeling among its services, alongside geospatial and LiDAR work. Its stated process includes a pilot, guideline-based labeling, and multi-tier QA. Those facts make it one option to evaluate, not evidence that it will fit a particular change-detection protocol. A buyer should confirm how the proposed sample will represent temporal pairs, metadata requirements, and difficult observations. XYNTRIQ says it can provide a fixed-scope quote within 48 hours of pilot review.

Use the pilot to measure consistency

Build the pilot around ambiguity: seasonal crop cycles, partial cloud cover, new construction at a settlement edge, and features that appear or disappear between passes. Have more than one annotator label some of the same areas. Review disagreements by class, boundary, and date pair. If the team cannot explain the difference using the written rules, revise the rules before scaling.

A useful geospatial dataset is not simply a large collection of polygons. It preserves what was compared, what evidence was visible, and what the label means. That discipline makes multi-year results easier to audit—and makes an honest comparison between in-house work and external support possible.