Satellite image datasets: spatial resolution, time series, labels and pricing

The sensor and observation define the product

A satellite dataset is more than a collection of overhead-looking pictures. Optical imagery records reflected light in selected bands; radar and thermal observations have different physical meanings and processing requirements. Natural-color previews can be useful for discovery without preserving the bands or calibration a scientific model needs. Specify the actual input representation, not only the file extension or a marketing resolution. If the task uses a derived map, retain the distinction between source observations and inferred labels.

Ground sampling distance is only one part of effective detail. Resampling to smaller pixels does not create new measured information. Cloud, haze, viewing angle, illumination, season and processing can alter what is visible. A before/after comparison must define how closely the acquisitions need to match and which differences are expected. If a buyer requires a certain date, explain whether it is a preferred window, a mandatory acquisition or a selection from an existing archive. Those have very different feasibility and cost implications.

BigEarthNet as a benchmark, not current worldwide coverage

The BigEarthNet Sentinel-1 description documents 590,326 paired Sentinel-1/Sentinel-2 patches. Its source period and geographic coverage are bounded, and its labels are tied to a particular land-cover reference. This is a useful multimodal benchmark, not evidence of current global coverage or access to every commercial imaging product. A patch is a prepared unit derived from an acquisition, not an independently commissioned satellite observation.

Before adopting a benchmark, compare its sensor, bands, processing, label taxonomy and geographic distribution with the task. Land-cover classification is different from building detection, crop monitoring, road extraction or fine-resolution change analysis. A strong result on one benchmark does not settle performance on a new region or a different season. Preserve the dataset version and label source so that a buyer can identify which parts are measured, derived or potentially outdated.

Why geographic and temporal matching matter

Potential users include geospatial analytics developers, environmental researchers, infrastructure-monitoring teams, agricultural applications and mapping-product groups. Each needs different geographic and temporal evidence. A classifier for broad land cover may tolerate coarse labels; a building-change evaluation may need accurately registered observations and an independently checked event timeline. Describe the target unit of change and what counts as a correct prediction. A model should not receive a label that was derived from the same imagery in a way that leaks the answer unnoticed.

The geospatial schema must make coordinates reproducible

Define coordinate reference system, georeferencing, extent, pixel spacing, band order, units, missing-data values and processing level. A file can open successfully while being interpreted in the wrong coordinate frame or with the wrong band scaling. If patches are resampled, cropped or composited, preserve the transformation and source-scene links. Labels must be transformed consistently; an offset of a few pixels can be a substantial error for small objects even when a map looks plausible at overview scale.

For time series, record capture time, composite window and observation quality separately. A monthly composite is not one photograph from a specific instant. For cloud handling, specify whether the deliverable includes masks, filtered observations, reconstructed values or a stated tolerance. Keep missing observations explicit rather than silently filling gaps with a different date. If a paired product contains radar and optical data from nearby dates, retain both timestamps and the pairing criterion.

Sampling and evaluation without spatial leakage

Nearby patches can share the same objects, land-use pattern and atmospheric conditions. Random file splits can therefore overstate geographic generalization. Depending on the task, hold out source scenes, regions, acquisition periods or sensor settings. Keep overlapping patches in one group. For temporal forecasting, ensure the evaluation does not use future information through labels, composites or preprocessing. The split design should answer the actual deployment question rather than merely produce convenient counts.

Review hard conditions deliberately: small targets, dense built areas, water boundaries, seasonal vegetation, cloud edges, shadows and mixed pixels. A sample should include the classes and geographic groups that matter, not only clear scenes. For change detection, inspect registration and the reference event. Construction, vegetation cycles, shadows and sensor differences can all create apparent change. Define whether ambiguous cases are excluded, marked uncertain or retained for a specific robustness task.

Pricing and a scope-normalized estimate

No representative public transaction sample in this research supports an average satellite AI-dataset license price. Quotes may be based on archive area, new acquisition, minimum order, sensor product, processing, observation frequency and rights. Preparation and labels are separate components. A price per square kilometer is not directly comparable with a price per patch unless the same area, dates, overlap and derived processing are specified. A repeated time series can multiply observations without increasing geographic coverage.

For a hypothetical review budget, a project selects 2,400 geographic units with three dates each. That is 7,200 observation records, but only 2,400 location groups. At an assumed twenty seconds for a technical quality check per observation, the work is forty hours. At a chosen loaded rate of $35 per hour, that modeled component is $1,400. It excludes imagery access or tasking, rights, scientific processing, expert labels, registration correction and delivery. These are planning assumptions, not observed rates or a Fiund quote.

A quote comparison should expose coverage losses. If one supplier offers a nominal target area but cannot meet the cloud/date rule, compare accepted usable coverage rather than billed area alone. If another supplies preprocessed patches, ask whether their source scenes and transformations are included. State what counts as a replacement observation and whether a new acquisition can actually meet the schedule. A low nominal unit rate is not a bargain when the product lacks the bands or temporal pairing needed for the task.

Public and commercial access routes: normalize the actual product

Copernicus Data Space terms describe Sentinel access under its free, full and open framework and referenced legal notice. That does not make every third-party contributing product identical in use conditions. For a commercial alternative, Planet’s purchase guidance distinguishes archive and tasking routes, with product usage terms governing selected imagery. These are real source routes with different products, not a promise of a particular acquisition or unrestricted redistribution.

Imagine a buyer comparing a multispectral public time series with a high-detail commercial image from one date. The first may support seasonal vegetation analysis; the second may resolve objects absent from the coarser series. Neither is a substitute merely because both cover the same area. Normalize native detail, bands, processing, date windows, cloud tolerance and permitted outputs before comparing price. Keep the original acquisition and derived patch counts visible.

A pilot can choose several location groups and inspect whether each requested before/after relationship is supported by source observations and label evidence. Separate a true changed object from differences in season, shadow, viewing angle or registration. Record whether an uncertain case needs another acquisition, a corrected label or exclusion. This provides a practical cost driver: the amount of review or acquisition needed to obtain accepted paired evidence, rather than the nominal number of files delivered.

A location-level acceptance example

Consider a building-change task with two dates over the same location. The second image can appear sharper while using a different viewing angle, season and registration. A model might respond to the changed shadow or shifted roof rather than an actual structural event. The acceptance record should identify source alignment, visibility, reference-date confidence and whether the change label has independent support. If those fields are absent, a buyer may need another observation or a narrower task rather than a more confident caption.

Quote comparison should therefore include accepted geographic units after the pairing policy is applied. A supplier offering many unmatched scenes and one offering fewer reviewed pairs are selling different deliverables. Report unmatched or uncertain locations as such. Do not convert them into accepted change examples merely to meet a target row count. This distinction links the scientific question directly to usable yield and makes an estimate more informative than an unsupported average price per satellite image.

Providers, deals and a Fiund sourcing brief

Potential routes include public Earth-observation programs under their terms, commercial imagery licensors, authorized archive holders and specialized processing partners. They may provide observations, derived analytics or both; compare the actual deliverable. Publicly accessible previews and maps should not be treated as automatically downloadable training inventory. Source permission, technical product definition and preparation are separate review questions.

No verified category-specific AI-training license deal amount is established here. Satellite-company contracts, spacecraft purchases and analytics-service agreements are not automatically comparable with a dataset license. Send Fiund the task, geography, acquisition windows, sensor/band requirements, native resolution, processing level, label evidence and intended uses. Fiund can assess potential source fit, missing information and the scope of representative review. Availability, rights, processing, acceptance, delivery and prices are confirmed for the selected project; no date-specific or worldwide coverage is guaranteed by this guide.

Questions buyers ask