Google Maps Custom Satellite Embeddings: 5-Day Monitoring
Google Maps Platform opened a Private Preview yesterday for Google Maps Custom Satellite Embeddings, a service that lets organizations request AI-derived Earth data for specific regions on cycles as tight as every five days, according to Google Maps Platform. That's a sharp break from the annual dataset it builds on, but it is not a new feed of satellite photos.
The product delivers AI-derived embeddings, a machine-learning data layer, at intervals down to five days. Google's own framing leans toward near real-time satellite monitoring, and that phrase should be read against the disclosed five-day minimum, not against literal live imagery.
The new service extends AlphaEarth Foundations, the model that has produced one embedding per location per year from 2017 through 2025, according to Google Earth Engine documentation. What follows covers what the product actually delivers, what a customer still has to build on top of it, where it slots into Google's cloud stack, and what Google hasn't yet said about cost, coverage, or accuracy.
Google Maps satellite embeddings shift toward on-demand geospatial change detection
The headline change is temporal, not visual. Organizations can now define their own area of interest and time window instead of relying on a fixed annual global grid, which moves the product from mapping toward monitoring and change detection, per Google Maps Platform.
Supported intervals include quarterly, monthly, weekly, or custom non-calendar-annual windows, with updates possible as often as every five days depending on available AlphaEarth input data sources, Google Maps Platform says. Google does not say how that dependency affects cadence across different regions or time periods.
Access is narrow for now. Organizations must apply for Private Preview by September 1, 2026, while Google plans a separate track offering selected academic researchers a free sample dataset through Earth Engine, according to the company's announcement.
That's the actual news: a move from yearly global snapshots to a request-based cadence, gated behind an application process rather than opened to the public.
What an embedding is, and what customers still have to build
An AlphaEarth embedding isn't a picture. Google describes the underlying model as a "virtual satellite" that fuses optical imagery, elevation, radar signals, and LiDAR into a single unified 64-band representation of each location, according to Google Maps Platform.
The public annual dataset offers a concrete sense of what that structure looks like. Google distributes it as Cloud Optimized GeoTIFF files, each holding 64 channels of signed 8-bit values per pixel, one channel per axis of the embedding rather than a red, green, or blue color band, according to Google Earth Engine documentation. There's no confirmation that the Private Preview product uses an identical file format, but the annual dataset's structure explains why these files are built for machine comparison rather than human viewing: each pixel is a position in 64-dimensional space, not a shade of anything.
Google Research has described these geospatial foundation models as capable of surfacing patterns in imagery that aren't readily visible to the human eye, the basis for automated comparison rather than someone reviewing two photos side by side, per Google Research.
Picture a plausible workflow: an infrastructure team requests embeddings for the same highway corridor at two five-day intervals, then compares the resulting representations to flag likely changes. Google's announcement describes data summaries, not an alerting application, so customers would likely need to build and validate their own comparison or detection workflow on top of what the service provides, per Google Maps Platform.
Where this fits in Google's broader cloud workflow
Custom Satellite Embeddings don't operate in isolation. Google has been assembling a wider stack of Earth AI products, and a few of them are worth understanding, even though the announcement doesn't confirm how, or whether, they connect to the new preview.
Google has said imagery can be analyzed directly inside BigQuery, cutting manual review time from weeks down to minutes, according to Google Maps Platform. That claim, made three months ago at Cloud Next, describes the general imagery pipeline rather than Custom Satellite Embeddings specifically.
Two Earth AI Imagery models in Google Cloud's Model Garden are pretrained to recognize infrastructure features like bridges, roads, and power lines, per the same announcement. BigQuery and Model Garden read as adjacent tools that could be relevant to turning embeddings into usable detections, but the announcement discusses them alongside other Earth AI products, not as a confirmed part of the Custom Satellite Embeddings pipeline. Vantor, a spatial intelligence company, already uses comparable models in its Sentry app to help recovery teams spot washed-out roads and prioritize repairs after storms, Google notes, though that's a case built on adjacent Earth AI models, not a confirmed deployment of the new product.
A longer-term research effort called Geospatial Reasoning aims to pair similar models with Gemini 2.5, letting someone ask natural-language questions about hurricane damage or infrastructure siting, according to Google Research. That effort remains a research initiative from last year, not a shipped feature wired into Custom Satellite Embeddings.
What the five-day claim doesn't establish
The five-day figure is a minimum, not a guarantee. Google's announcement doesn't state how consistently that cadence holds across different locations, how much time passes between image capture and delivery to a customer, or which regions have enough AlphaEarth input data to support it at all, according to the company's own post.
There's a related question about continuity underneath the whole system. Google notes separately that continued production of the annual layer depends on ongoing input streams from USGS and ESA, a caveat stated specifically for the annual dataset rather than confirmed to apply to the custom product, per Google Earth Engine documentation. Whether a disruption to those upstream feeds would ripple into the five-day product is an open question the announcement doesn't address.
Pricing is a bigger blank spot. The annual dataset carries a CC-BY 4.0 license, but its Cloud Storage bucket is configured "provider pays," per the same documentation. An open license doesn't mean free access, and the announcement does not state pricing or billing terms for the new custom service at all.
Given Private Preview status, the realistic audience today is enterprise and research teams rather than casual Maps users. Prospective applicants will need clarity on:
- Geographic coverage and resolution for their specific regions of interest
- Latency between image capture and actual data delivery
- Minimum order area and data format requirements
- Accuracy benchmarks for their particular use case
None of the current materials address any of these points. The stated use cases around disaster-response road triage or construction detection trace back to Google's broader Earth AI marketing and one adjacent customer, not to a confirmed deployment of Custom Satellite Embeddings itself, per Google Maps Platform.
What comes next
Applications close September 1, 2026. The next meaningful signal will be what Google discloses about general availability, including pricing, coverage guarantees, and any accuracy figures validated against real deployments rather than marketing examples.
Those terms should clarify the service's intended scale and practical fit for enterprise monitoring, whether that means disaster response, infrastructure inspection, or the kind of construction tracking Google has floated in adjacent product announcements. Until Google says more, the honest description of Custom Satellite Embeddings is a request-based data layer with a five-day floor, not a live camera in the sky.
Comments
Be the first, drop a comment!