The image that helps you find an observation may not be the data you need to measure it
A satellite catalogue shows a clear, colourful picture of a coastline. You download the first linked image and begin comparing brightness across dates. The file opens and looks convincing, but it may be a browse image designed for discovery rather than a calibrated measurement asset. Successful viewing has answered a different question from successful scientific input selection.
The SpatioTemporal Asset Catalog, or STAC, describes observations and their associated assets. This guide follows the choice from a catalogue Item to the actual file used in a calculation. It distinguishes dictionary keys, semantic roles, media types, band meanings, per-Item availability and usable coverage, with original small examples that can be checked without downloading a real satellite archive.
The fictional assets below are teaching records, not product recommendations or measurements of a real coast. A STAC description supports discovery and interpretation; it does not certify every referenced file for every analysis. For actual measurements, consult the product’s official documentation, quality information, licence and processing requirements as well as the catalogue metadata.
Geoinformatics subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1: Identify the asset’s job · Chapters 1–3
Route 2: Work a file-selection example · Chapters 4–5
Route 3: Check availability and meaning · Chapters 6–8
Check the individual Item, not only the Collection definition
Route 4: Validate inputs and coverage · Chapters 9–10
Route 5: Transfer and document · Chapters 11–12
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Identify the asset’s job · Chapters 1–3
Separate the observation record from its files
Read keys and roles as different information
Distinguish media type from measurement meaning
Work a file-selection example · Chapters 4–5
Compare the size of a preview with its analytical support
Select dependencies before calculating a ratio
Check availability and meaning · Chapters 6–8
Check the individual Item, not only the Collection definition
Resolve the actual resource before claiming availability
Read band definitions, scaling and processing level
Validate inputs and coverage · Chapters 9–10
Keep Item footprints and usable pixels separate
Validate the resource after opening it
Transfer and document · Chapters 11–12
Transfer from imagery to another kind of asset catalogue
Hand over an evidence-based asset choice
CHAPTER 1 OF 12
1. Separate the observation record from its files
A STAC Item describes an observation or other spatiotemporal entity and points to associated assets. The Item is not itself the full scientific array. Its geometry, time and properties help you find and understand the record; its asset objects lead to resources that can be downloaded or streamed.
One Item can refer to several resources serving different jobs: a small preview, a larger overview, measurement bands, quality information and metadata. Counting five assets therefore does not establish five independent observations. The assets can be different representations or supporting pieces of the same observation.
Imagine one fictional acquisition called Coast-17. It has a preview image and two numeric bands. A user browsing the catalogue needs the preview to recognise the scene. A learner calculating a band ratio needs the appropriate numeric inputs. Both are using Coast-17, but their asset requirements are not interchangeable.
Write the analytical question before choosing a file. Are you illustrating a lesson, measuring a physical quantity, checking a quality flag or identifying the acquisition? Each question calls for different evidence. A catalogue becomes more useful when its files are treated as purposeful resources rather than an undifferentiated list of download buttons.
CHAPTER 2 OF 12
2. Read keys and roles as different information
The assets member is a dictionary. Each asset has a key that distinguishes it within that dictionary. The STAC standard cautions that clients should not assume universal meaning from those keys. A provider may use a familiar-looking name, but another provider can use a different convention for the same purpose.
The roles field describes semantic purpose. Common role names include thumbnail, overview, data and metadata. Roles can be multivalued and providers can use additional terms. A client should interpret recognised roles in context, not assume that a single universal role vocabulary completely determines scientific suitability.
For a fictional test, give asset key A the role data and asset key science the role thumbnail. A program that chooses the key science because it sounds authoritative selects the wrong resource for our intended calculation. A role-aware program has better information, but still needs to inspect what the data asset actually measures.
Treat descriptive titles as useful human guidance rather than immutable computational identity. A translated or revised title can change while the asset’s intended identity stays stable. This is one reason the existing STAC language guide protects identifiers and resource links separately from reader-facing descriptions.
CHAPTER 3 OF 12
3. Distinguish media type from measurement meaning
A media type describes a resource’s format and helps a client choose how to read it. It does not fully specify its scientific interpretation. Two TIFF resources can contain different bands, resolutions, scales and quality conventions. Two JPEGs can be illustrations of entirely different processing stages.
An image that decodes successfully has passed a format-level test. It has not necessarily passed a unit, calibration or band-selection test. A three-channel browse image can show a useful visual combination while its red display channel is not the same thing as a product’s scientifically defined red measurement band.
For a teaching example, suppose a preview stores integers 0–255 selected by a display stretch. A scientific band stores a quantity under a separate scale and offset. Both are numeric arrays, but arithmetic on the preview values does not automatically recover the underlying physical quantity. The display transformation may discard information.
Keep three labels together: file format, semantic purpose and measurement definition. The first answers how the resource is encoded, the second why it is supplied, and the third what its numbers mean. A file extension cannot stand in for all three, even when it is familiar to the software you normally use.
CHAPTER 4 OF 12
4. Compare the size of a preview with its analytical support
Our fictional Item contains a 256 by 256 pixel browse image and two 2048 by 2048 scientific bands. The browse image has 65,536 pixel positions. Each scientific band has 4,194,304. The larger array has 64 times as many positions because each dimension is eight times larger.
Those dimensions do not alone establish ground resolution, spatial accuracy or scientific quality. They do show why a small preview cannot be assumed to retain the same spatial sampling as the larger arrays. Enlarging the preview on a screen changes its display size, not the information originally retained in it.
Suppose an imaginary narrow channel occupies two scientific pixels across but less than one browse pixel after visual downsampling. A sharp-looking coastline in the preview does not establish a reliable measurement of that channel’s width. The task needs an appropriate source, its geographic transform and an uncertainty assessment.
Do not make the opposite mistake and assume the largest file is always the right one. It may be a different product, include unnecessary channels or use a processing level unsuitable for the task. Size is one diagnostic clue. Purpose, measurement meaning and actual spatial support remain the deciding evidence.
CHAPTER 5 OF 12
5. Select dependencies before calculating a ratio
Consider an invented index I=(B−A)/(B+A). The names A and B here are classroom symbols, not real spectral-band recommendations. A valid calculation requires the intended measurements, compatible spatial and temporal support, usable cells and a permitted denominator. Selecting two assets merely because they contain numbers does not establish those conditions.
At one fictional cell let A=2 and B=6, giving I=4/8=0.5. If a browse rendering changes the display values to A=50 and B=100, its analogous calculation gives 50/150≈0.3333. The difference is not a rounding problem. The display values are not the same measurement inputs.
Now let scientific A=0 and B=0 at another valid cell. Both measurements can be present while the formula is undefined because its denominator is zero. Asset selection, observation validity and mathematical domain checks are three separate gates. Passing the first does not guarantee the others.
Before processing a large archive, make a tiny dependency table: intended quantity, required assets, required metadata, quality conditions, spatial alignment and formula restrictions. This makes the hidden assumptions inspectable. It also shows which missing resource should stop a calculation instead of being replaced by an attractive preview or an unexplained default.
CHAPTER 6 OF 12
6. Check the individual Item, not only the Collection definition
A STAC Collection can describe asset definitions for its member Items. The standard explains that these definitions can describe the union of assets that may appear, rather than guaranteeing that every asset exists in every Item. A possible asset and an available per-Item asset are different facts.
Imagine three fictional Items. Item1 has A, B and quality assets; Item2 has A and quality but no B; Item3 has A and B but no required quality resource. A Collection-level description can mention all three asset kinds while none of that changes the individual availability table.
Under our stated calculation rule requiring A, B and quality evidence, only Item1 is ready. Item2 lacks a measurement dependency and Item3 lacks the specified acceptance evidence. A different task, such as illustrating approximate scene coverage, may accept a different set. Eligibility follows the task rather than the Collection’s longest asset list.
Record missing dependencies explicitly. Do not silently copy an asset URL from another Item because its filename resembles the missing one. That substitution can mix acquisitions, processing levels or locations. If an authorised filling or substitution method is used, preserve it as a derived operation with its own justification and validation.
CHAPTER 7 OF 12
7. Resolve the actual resource before claiming availability
An asset’s href identifies a resource. STAC allows relative as well as absolute references. A relative reference needs the correct base context to resolve to the intended resource. Copying a small asset fragment into a new document can change that context even when its text remains identical.
For a neutral classroom example, suppose a document’s asset reference points to a neighbouring data directory. Moving the document without preserving the resource relationship can make the same reference resolve elsewhere or fail. This is a reference-resolution problem, not evidence that the observed coastline changed.
Successful catalogue retrieval also does not guarantee successful asset retrieval. The Item JSON can be public while an asset has separate access conditions, an expired temporary URL, a service interruption or a changed location. Respect the provider’s authorised access process; do not work around authentication or permission controls.
Distinguish discovered, reachable, readable and scientifically suitable states. A resource can satisfy one and fail the next. A reliable processing report keeps those transitions visible rather than reducing them to one downloaded flag. That distinction helps the next operator repair the actual problem instead of repeatedly searching for another similarly named image.
CHAPTER 8 OF 12
8. Read band definitions, scaling and processing level
A data role is not a universal promise of calibrated reflectance, temperature, elevation or any other physical quantity. Identify the product definition and relevant band metadata. Extensions and provider documentation can supply further information, but their presence and meaning must be checked rather than inferred from an attractive catalogue title.
As a fictional scaling exercise, let stored integer D represent physical value V=0.01D−1. A stored value of 250 corresponds to 1.5, not 250. If a second product uses another scale or offset, directly subtracting stored integers can mix incompatible representations even when both assets are labelled data.
Keep quality and fill conventions alongside the scaling rule. Applying a physical conversion to a fill code does not turn it into a valid observation. A valid zero should not be discarded merely because another file uses zero for missingness. The connected raster-mask guide explains how value and eligibility must remain paired.
Processing level matters as well. Two resources from the same observation can have different corrections or intended uses. Do not describe them as interchangeable without the product authority’s evidence. This article supplies selection reasoning, not a universal recipe for converting arbitrary remote-sensing products into equivalent scientific measurements.
CHAPTER 9 OF 12
9. Keep Item footprints and usable pixels separate
An Item’s geometry supports geographic discovery. It does not automatically prove that every pixel within the described region is usable for a particular analysis. Clouds, missing coverage, sensor issues and task-specific exclusions can leave a smaller eligible support in the actual assets.
Suppose an imaginary Item intersects a 100-square-unit study region. After applying the documented task rule, only 65 square units have the required valid inputs. The discovery match remains correct, but a claim of complete usable measurement coverage would be wrong. Report the eligible support rather than substituting the Item footprint for it.
For a two-date comparison, the common usable area can be smaller again. If date1 covers 65 units and date2 covers 70, their overlap is not determined by those two totals alone. They could cover largely the same or substantially different places. You need the spatial relationship of their masks.
A catalogue search is therefore an early filter, not the final evidence for complete analytical coverage. Keep the subsequent quality and support checks in the workflow. The distinction prevents a successful geographic search from being promoted into a stronger scientific statement than its metadata actually supports.
CHAPTER 10 OF 12
10. Validate the resource after opening it
Compare the opened asset with its declared role, media type and relevant scientific metadata. Inspect dimensions, bands, geographic reference, temporal context, datatype, scale, offset, nodata and quality dependencies as appropriate. A readable file can still be the wrong resource for the calculation.
Use deliberately recognisable test cases. Open a known preview and confirm it is rejected by the scientific-input rule. Open a valid measurement resource and confirm that expected bands and units are recognised. Include an Item missing one required asset so a silent fallback cannot pass unnoticed.
Verify correspondence across resources: the A and B arrays should belong to the intended Item or otherwise justified combination. Similar filenames do not establish temporal or spatial alignment. A calculation can execute perfectly while comparing one date’s A with another date’s B.
Record failures by mechanism. Wrong asset selection needs semantic repair; inaccessible href needs authorised retrieval repair; scale mismatch needs interpretation repair; invalid coverage needs eligibility or inference treatment. Replacing every failure with another download attempt loses the information needed to make the system dependable.
CHAPTER 11 OF 12
11. Transfer from imagery to another kind of asset catalogue
STAC can describe more than a familiar optical satellite scene. The same asset-selection reasoning can apply to other spatiotemporal resources, but the measurement definitions and appropriate checks change. A preview of a point cloud, for example, is not a substitute for the points and attributes needed by a measurement task.
Try a changed-context exercise with a fictional elevation product: one asset is a shaded-relief illustration, another is an elevation array and a third contains quality flags. The shaded picture can make terrain easy to recognise while its brightness depends on rendering choices. Measuring height from those display intensities would confuse representation with the underlying quantity.
Now change the task to classroom illustration. The shaded picture may be the appropriate choice, provided its source, date and limitations are conveyed. Rejecting all previews as useless would be just as mistaken as using them for every measurement. Suitability follows the question, not a rigid ranking of file types.
The invariant is dependency-aware selection: identify the resource’s job, inspect its actual content, preserve its meaning and verify what the next operation requires. The non-transferable parts include particular band names, units, quality flags and transformation formulas. Relearn those from the new product’s authoritative documentation.
CHAPTER 12 OF 12
12. Hand over an evidence-based asset choice
A useful handoff names the Item and Collection, selected asset keys, interpreted roles, actual resource references, product definition and processing conditions. If the choice was based on a provider-specific convention, state that convention rather than presenting it as a universal STAC rule.
For our classroom ratio, the receipt should say which resources supplied A and B, why their values are compatible, which quality rule was applied and which cells were excluded for the formula’s denominator. It should also identify resources that were used only for browsing. Another learner can then reconstruct both the selection and its limits.
Where the metadata is insufficient, say what is missing. A data role with no established measurement definition is not an invitation to invent units. An absent quality resource should not become a claim of perfect quality. A missing per-Item asset should remain distinguishable from a failed download or an unsuitable resource.
The practical endpoint is not a folder containing the largest possible number of files. It is a justified chain from question to selected resources to a calculation that those resources can actually support. A browse image can remain an excellent guide to discovery without being asked to provide measurements it was never designed to preserve.
A practical selection review can be performed without opening a large array. Give a colleague the question and a fictional asset inventory, with keys deliberately chosen to sound misleading. Ask them to state which additional facts are required before any calculation. A good answer identifies measurement definitions, scaling, quality dependencies and spatial compatibility. It does not simply select the largest resource or the key containing the word science.
Next remove one required measurement asset from an Item while leaving its preview and Collection definition unchanged. The analytical readiness check should change even though the discovery interface still looks complete. Restore the measurement but remove a required quality dependency. The check should identify a different missing input rather than reuse the same explanation. These tests make failure reporting useful: the next operator learns what must be supplied or justified.
Then provide two individually valid measurement resources from different fictional dates. A parser can read both and each can pass its own range checks. The combined operation still needs a temporal compatibility decision. If the task is a same-acquisition ratio, mixing dates violates its stated dependency. If the task explicitly compares dates, the combination may instead be intended. The correct response follows the question rather than a universal rule that all assets must always share one timestamp.
For a final handoff test, ask another learner to reproduce the chosen eligible-cell count from the recorded asset identities and quality rule. They should not need the first analyst’s memory or a screenshot of the catalogue. If their count differs, compare dependency definitions and masks before debating the final average. This is how asset discovery becomes a traceable analytical decision rather than a sequence of clicks that happened to produce a number.
Independent worked case and evidence checks
Work this independent fictional inventory. ItemA supplies asset p, a 256×256 preview with thumbnail role; assets a and b, 2048×2048 measurement arrays with data roles; and asset q, a quality resource described by the provider. ItemB supplies p, a and q but lacks b. ItemC supplies p, a and b but lacks q.
The specified task needs both measurements and the provider’s quality rule. ItemA has all required dependencies. ItemB cannot supply the ratio because b is absent. ItemC is not ready under this particular rule because the required quality evidence is missing. The presence of p in every Item makes the catalogue visually complete while the analytical dependency set is incomplete.
At three eligible ItemA cells, the physical A values are 1,2 and 3 and B values are 3,6 and 9. Each ratio (B−A)/(B+A) equals 0.5. Suppose an arbitrary browse stretch instead produces display pairs (20,40), (40,80) and (80,120). Their analogous ratios are 1/3,1/3 and 1/5. The mean is approximately0.2889, not0.5.
This is not a statement about a real sensor’s display stretch. It is a constructed demonstration that an image can preserve recognisable appearance while not preserving the arithmetic meaning required by an index. A calculation cannot be justified solely because its input values lie in an array that software can read.
Next suppose the scientific arrays are valid over 70 of 100 equal-area study cells, and q excludes10 of those70 for the task. There are60 eligible cells, not100 and not70. If their index sum is24, the eligible-cell mean is24/60=0.4. Replacing the40 excluded cells with zero gives24/100=0.24, a different and unjustified interpretation unless a separate model supports it.
For transfer, change the question to locating the acquisition in a classroom catalogue. The thumbnail can be useful even for ItemB, whose missing scientific band prevents the ratio. The same Item is usable for one question and not ready for another. Explain the dependency difference rather than calling the entire Item good or bad.
Finally remove the roles field from a fictional asset while retaining a well-documented provider definition. Do not invent a universal rejection rule: determine what the applicable STAC schema requires and what the task can establish from other evidence. Roles improve semantic discovery, but scientific suitability ultimately needs a complete, justified interpretation of the actual resource.
Keep the discarded candidates in the exercise record. ItemB was not rejected because its thumbnail looked different; it lacked a required measurement. ItemC lacked required quality evidence. Recording those precise reasons lets a later analyst reconsider the Items when another task has different dependencies, without confusing the earlier decision with a permanent judgement about the entire catalogue.
Worked comparison
| Evidence | What it helps establish | What it does not guarantee |
|---|---|---|
| Asset dictionary key | Identity within the asset map | Universal scientific meaning |
| Roles | Intended semantic purpose | Calibrated units or task suitability |
| Media type | How to decode the resource | Correct band or processing level |
| Collection asset definition | Possible member-Item assets | Every asset present in every Item |
| Opened file and quality checks | Actual task dependencies and support | Unobserved conditions outside that support |
Questions and practical checks
Is a data role enough to start measuring?
No. Establish the product meaning, units, processing level, spatial support and quality requirements for the task.
Can I select assets by filename?
Only under a documented provider convention. Keys and names are not a universal STAC vocabulary.
Does every Item have the assets described by its Collection?
Not necessarily. Check the individual Item’s actual asset map and availability.
Is a thumbnail always scientifically useless?
No. It can be excellent for discovery or illustration. It should not be assumed equivalent to the measurement assets needed by another task.
Does an intersecting Item footprint prove full usable coverage?
No. Inspect the actual assets and quality masks for the study region and required calculation.
Continue with the connected subjects
Keep discovery and data interpretation connected
STAC language editions: preserving asset and catalogue identity
Raster masks: usable cells and multiband dependencies
Choose appropriate representations
Vector tiles: display representations versus source-object evidence
Categorical resampling: when numeric image operations lose meaning
Sources and scope
OGC STAC1.1: Items, Assets, roles and Collection asset definitions
USGS: STAC discovery for Landsat resources
Rasterio: valid-data masks and their interpretation
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 30 October 2026.
Connect the representation checks
Each guide below addresses a different dependency. Use the one that matches the evidence you still need rather than treating any single successful check as complete validation.
CRS negotiation: selection coordinates versus returned positions. Follow the reference chain from input search bounds to returned positions and the next recipient.
GeoJSON null geometry: present records without map locations. Reconcile source records, supplied geometry and the map display before claiming completeness.
GeoPackage SRS consistency: references, headers and coordinate meaning. Distinguish internal reference agreement from independently justified coordinates in the saved package.