The pass lasts eleven minutes. During those eleven minutes a ground station pulls down everything the spacecraft collected since the last contact, at a rate the antenna and modem can sustain but the storage behind them frequently cannot. Miss the write and you do not get a retry; the next opportunity is ninety minutes away and the data is gone. Everything downstream in an Earth observation ground segment inherits its shape from this constraint.
The Downlink Window Sets the Floor
Contact bandwidth varies enormously by mission class. An X-band downlink at several hundred megabits per second is common for small commercial optical constellations; higher-capability missions push into the gigabit range. Convert to bytes and a single ten-minute pass can deliver anywhere from tens of gigabytes to well over a terabyte of raw telemetry and instrument data.
What the ingest tier must handle is not the daily average but the sustained write during contact, multiplied by however many antennas operate simultaneously. Buffer to a fast landing area sized for several passes, then move to the archive asynchronously. Writing directly into a capacity tier during a contact window is the classic design error, and it fails quietly by dropping frames rather than loudly by throwing an error.
Every Processing Level Multiplies the Footprint
Raw Level 0 is only the beginning. Radiometric correction produces L1A or L1B, geometric correction and orthorectification produce L1C, and atmospheric correction produces L2 surface reflectance. Each level is typically larger than the one before, because calibration converts packed integers into floating point and because reprojection expands the pixel grid.
A reasonable planning assumption is that the full chain occupies three to six times the L0 volume, before analysis-ready derivatives. Deciding which intermediate levels are retained versus recomputed on demand is the largest single lever on total capacity when specifying NAS solutions for satellite imagery, and it deserves a written policy rather than a default of keeping everything.
Revisit Cadence Turns Images Into Time Series
A constellation with daily revisit does not produce images, it produces a stack. Vegetation monitoring, subsidence analysis and yield forecasting all read across time rather than space: give me this 10 km tile for every acquisition in the last three years. That is a scattered read touching hundreds of files, and it performs badly on any layout that assumed users would read one scene at a time.
Chunking and tiling strategy matters more than raw media speed. Organizing the archive so spatially co-located data across dates lands near each other, and keeping directory structures shallow, affects analyst experience more than faster disks. Platforms sized as scale-out NAS for IoT and big data cope better with these fan-out reads because metadata handling scales alongside capacity instead of concentrating on one head.
Reprocessing Is Not an Exception, It Is a Cycle
Cloud masks improve. Atmospheric correction algorithms get revised. Calibration coefficients are updated after on-orbit characterization. Each triggers a campaign that reads the entire retained L0 or L1 archive and writes a new generation of products.
This is why raw retention is non-negotiable: without L0 you cannot benefit from a better algorithm, and the historical record you sold customers stays frozen at the quality of the day it was produced. Reprocessing also imposes a concurrent read-and-write load for days or weeks. Systems specified for steady-state delivery and never tested under a reprocessing campaign reveal their limits at exactly the moment a new product version has been announced.
Delivery Formats Change the Access Profile
Cloud-optimized GeoTIFF and similar tiled, internally-overviewed formats let a client fetch only the bytes covering a requested area at a requested zoom, using range reads instead of whole-file transfers. That is excellent for users and a meaningful change for storage: the workload shifts from large sequential reads to many small random reads scattered inside very large files.
Serving that well requires low-latency random access and enough concurrency headroom for many simultaneous partial reads. It also rewards keeping recent acquisitions on a faster tier, since demand for imagery decays sharply with age for most commercial use cases while never quite reaching zero.
Tiering, Egress and the Cost of Getting Data Back
Archive economics are where many programmes get caught. Deep cloud archive classes are cheap to fill and expensive to read, with retrieval fees and egress charges that make a reprocessing campaign financially painful. An archive you cannot afford to read is not an archive, it is a deletion with extra steps.
The pattern that holds up is a hot tier for recent and frequently requested products, a nearline capacity tier for the bulk of the archive, and cold storage only for material genuinely unlikely to be reprocessed. Keeping the reprocessable corpus on enterprise NAS storage platforms under your own control makes campaign costs predictable, because the marginal cost of reading your own archive is power and time rather than a per-gigabyte invoice.
The Catalogue Is as Important as the Pixels
STAC items, footprints, cloud cover statistics, sun and view geometry, quality flags and processing provenance are what make an archive searchable. They are small, numerous, frequently updated and read constantly by every API query. Their access profile is the opposite of the imagery they describe, which is why the catalogue database usually belongs on different storage than the pixel archive.
Keeping them consistent is the harder part. A product deleted but still listed produces failed customer requests; a product written but never catalogued is invisible and effectively lost, so reconciliation jobs that walk both sides should run continuously. Teams weighing block, file and direct-attached options will find the trade-offs in SAN, NAS and DAS compared map closely onto this split between transactional metadata and bulk satellite imagery capacity.
Specifying NAS for satellite imagery comes down to designing for four pressures at once: a burst you cannot miss, a product chain that multiplies volume, a reprocessing cycle that rereads everything, and a delivery layer wanting small random reads from very large files. Size the landing tier for the worst pass, keep raw because the algorithms will improve, and make sure the archive is one you can afford to read. Everything downstream depends on those choices made years earlier at the ground station.
Top comments (0)