HERE Technologies and Wherobots announce partnership Learn More

Comparison

RasterFlow vs. Google Earth Engine and ArcGIS

RasterFlow, the imagery engine in the AI Context Engine for Physical World Data, turns satellite and aerial imagery into features you can query. See how it compares with Google Earth Engine and ArcGIS Online at each step.

Start for free Talk to sales
Cost before you run

Priced by data processed.

Models included

Five built-in models plus your own.

Your storage, open formats

Zarr and GeoParquet on Amazon S3.

Computer vision for geospatial data

RasterFlow, Earth Engine and ArcGIS Online, capability by capability

Each row is a step from raw imagery to features you can query. Every Earth Engine and ArcGIS Online cell comes from Google and Esri documentation, linked below the table. RasterFlow is in Public Preview.

RasterFlow, Earth Engine and ArcGIS Online, capability by capability
Capability Imagery engine Wherobots RasterFlow Mosaics, model inference and vectorization as managed tasks, priced by data processed. Imagery platform Google Earth Engine Public data catalog and scripted analysis on Google Cloud. GIS platform ArcGIS Online Esri SaaS with imagery hosting, raster analysis and pretrained models.
Prepare imagery
Bring your own imagery How your own rasters get into the platform Included: Reads your own GeoTIFF and Cloud Optimized GeoTIFF (COG) imagery in place from your S3 bucket, indexed in a GDAL Raster Tile Index; no upload or ingest step Partial: Upload GeoTIFFs as image assets: up to 10 GB in the Code Editor, larger files from the command line. Or register COGs in Cloud Storage as COG-backed assets. Assets count against your storage quota Partial: Upload imagery to create hosted imagery layers, converted to Cloud Raster Format; needs a Creator, Professional or Professional Plus user type, and storage consumes credits
Built-in imagery Imagery ready to use without loading it Sentinel-2 L2A with all 12 bands at 10 m, and USDA NAIP aerial imagery at 30 cm and 60 cm Included: The Earth Engine public data catalog, including Sentinel-2 surface reflectance Included: ArcGIS Living Atlas imagery such as Landsat, Sentinel and NAIP
Mosaicking One mosaic over an area of interest and a time range Included: One task builds a seamless mosaic, reprojected onto your output grid Included: mosaic(), qualityMosaic() and median() composites you script over a filtered collection Included: Mosaic multiple images into one imagery layer in raster analysis
Cloud cover removal Cloud-free composites Included: Built into Sentinel-2 mosaics: scene filter, per-pixel mask of cloud, shadow and cirrus, median composite Partial: Cloud Score+ and the QA60 band; you write the masking step Partial: A pretrained Living Atlas cloud mask model run through Classify Pixels
Run models
Built-in models Pretrained models run by the platform Included: SAM3, Fields of the World, Tile2Net, Meta CHM v1 and ChesapeakeRSC, plus bi-temporal change detection Partial: Classical classifiers you train (CART, RandomForest, NaiveBayes and SVM); no deep learning in the Machine Learning APIs Included: Over 100 pretrained Living Atlas models, such as building footprints, land cover and trees
Bring your own deep learning model Included: PyTorch or TensorFlow models; PyTorch 2 archives load from Amazon S3 or Hugging Face Partial: TensorFlow, PyTorch or AutoML model on a Vertex AI endpoint, through ee.Model.fromVertexAi (public preview) Included: TensorFlow, PyTorch or Keras models packaged as .dlpk
Model hosting Where the model runs Included: On Wherobots GPU runtimes, inside the RasterFlow task A Vertex AI endpoint you create, deploy and pay for separately; 1.5 MB default payload per request Raster analysis in ArcGIS Online with a Professional or Professional Plus user type; GPU use not documented
Edge effect cleanup Overlapping tiles merged at seams Included: Overlapping windowed inference, merged by weighted average Partial: inputOverlapSize sets tile overlap (default 0); you discard the edge pixels in your code Included: Padding blends predictions at tile borders, up to half the tile size
Use the results
Vectorize predictions Raster output to polygons Included: vectorize_mosaic, written as GeoParquet Included: reduceToVectors Included: Raster to feature conversion; object detection writes polygon feature layers
Output formats Included: Zarr stores, multiscale Zarr for maps and GeoParquet; PMTiles through WherobotsDB Included: GeoTIFF, Cloud Optimized GeoTIFF and TFRecord images; CSV, SHP, GeoJSON, KML, KMZ and TFRecord tables Partial: Hosted imagery layers (Cloud Raster Format); feature exports to CSV, shapefile, file geodatabase, GeoJSON, KML and GeoPackage
Where results land Included: Written to your own S3-compatible storage Partial: Export tasks to Google Drive, Cloud Storage, an Earth Engine asset or BigQuery; 8 or 20 concurrent batch exports by plan Partial: Hosted layers in ArcGIS Online; writing to your own bucket not documented
Open formats and catalogs Included: Zarr, GeoParquet, Parquet and COG; Apache Iceberg tables through a Managed Catalog Partial: Reads Zarr v2 from Cloud Storage; Zarr, GeoParquet, Iceberg and STAC output not documented Partial: STAC read; Zarr output, GeoParquet export and Iceberg not documented
Data on Amazon S3 Included: Reads from and writes to Amazon S3 Not included: Not documented Partial: S3 and Azure data stores feed ArcGIS Data Pipelines; imagery from S3 not documented
Join results with vector data in SQL Included: WherobotsDB spatial SQL over the vectorized output, in the same platform Partial: Export to BigQuery, then BigQuery SQL, including ST_REGIONSTATS Partial: SQL queries on hosted feature layers; Join Features and zonal statistics tools
Cost and operations
Infrastructure Included: Serverless; no cluster or model endpoint to manage Partial: Fully managed compute; model endpoints run on Vertex AI Included: SaaS managed by Esri
Data volume per run Included: No documented limit; the Fields of the World 2024 to 2025 global release ran on RasterFlow Exports capped by maxPixels per task; 8 or 20 concurrent batch exports by plan Uploads up to 500 GB per file; dynamic imagery requests up to 4,000 x 4,000 pixels
Billing unit RasterFlow Spatial Units: pixels × bands × time periods × a complexity factor EECU-hours, plus a monthly plan from $500 (Basic) or $2,000 (Professional) ArcGIS Online credits by pixels processed, plus 1.2 credits per GB per month for tiled imagery storage
Cost known before a run Included: Calculated from area, resolution, bands and time periods before a task starts Partial: Usage is reported when a request completes; a Pricing Calculator estimate is based on projected usage Included: Estimate credits before you run a tool
Example: 150,000 km² Sentinel-2 mosaic, 12 bands Wherobots estimate for one mosaic About $13.50 About $50 to generate and export Not published
Noncommercial use Billed by RasterFlow Spatial Units Included: Free for nonprofit scientific research, education and noncommercial activities Credits apply to analysis and storage

Sources, checked October 2026: RasterFlow overview, datasets, models, change detection, bring your own rasters, storage integration and billing; Earth Engine image upload, COG-backed assets, compositing and mosaicking, Sentinel-2 surface reflectance, machine learning, classification, Vertex AI models, ee.Model.fromVertexAi, reduceToVectors, exporting, Zarr v2 loading, usage monitoring, quotas, pricing and noncommercial use. ArcGIS Online credits, imagery layers, hosted layer exports, data stores, SQL and upload limits; Esri pretrained models, ArcGIS Image, the Detect Objects Using Deep Learning tool reference, the raster analytics API and the cloud removal sample. The Fields of the World release and the 150,000 km² mosaic figures come from the RasterFlow Public Preview announcement, and the mosaic figures are Wherobots estimates. “Not documented” means the vendor’s documentation does not describe the capability.

Customer proof

Getting data, algorithms, and compute in one place with Spark/Sedona notebooks is a huge boost. Powerful like Earth Engine, but with the control developers need to get jobs done.

John Powell Sr. Geospatial Data Engineer, Addresscloud

Which one fits your imagery work

Production imagery pipelines

RasterFlow fits teams that run the same mosaic, model and vectorize steps over any area, from one county to a global release, and need each run priced before it starts. Results land in open formats in their own storage, ready to join with parcels, roads or assets in WherobotsDB.

Open the Model Hub

Your own models at scale

RasterFlow runs your own PyTorch or TensorFlow model over any mosaic on Wherobots GPU runtimes, with overlapping inference to clean up tile seams. There is no endpoint to deploy.

Bring your own model

Teams on AWS

RasterFlow reads from and writes to Amazon S3, so imagery and results stay in the account where the rest of the data lives.

Use your own storage

Teams standardized on ArcGIS

ArcGIS Online fits teams already working in ArcGIS that want pretrained Living Atlas models inside their GIS, with results as hosted layers for their maps and apps.

Exploration over a public catalog

Earth Engine fits analysts exploring its public data catalog in the Code Editor, training classic classifiers such as random forest, and noncommercial research, which Earth Engine offers free of charge.

RasterFlow and Google Earth Engine FAQ

Is RasterFlow an alternative to Google Earth Engine?

Yes, for turning imagery into features. RasterFlow builds the mosaic, removes clouds, runs a computer vision model and vectorizes the output as managed tasks, and writes the results to your own storage. See satellite imagery analysis.

How does RasterFlow pricing compare with Earth Engine pricing?

RasterFlow bills by data processed, so the cost of a task is known before it runs. Earth Engine bills EECU-hours plus a monthly plan and reports usage when a request completes. Wherobots estimates a 150,000 km² Sentinel-2 mosaic with all 12 bands at about $13.50 in RasterFlow and about $50 to generate and export in Earth Engine.

Can Earth Engine run deep learning models?

Earth Engine’s Machine Learning APIs do not support deep learning. Earth Engine sends image tiles to a TensorFlow, PyTorch or AutoML model you deploy on a Vertex AI endpoint, which Vertex AI bills separately. RasterFlow runs five built-in models and your own PyTorch or TensorFlow model on Wherobots GPU runtimes.

How does RasterFlow remove clouds from Sentinel-2 imagery?

RasterFlow skips scenes with 75% or more cloud cover. It masks cloud, cloud shadow and thin cirrus pixel by pixel, using the Sentinel-2 Scene Classification Layer. Then it takes the median of the clear observations for each pixel. See Sentinel-2.

Can I join RasterFlow results with my vector data?

Yes. vectorize_mosaic writes polygons as GeoParquet, and WherobotsDB joins them with parcels, roads or assets in spatial SQL. See WherobotsDB and change detection.

How does RasterFlow compare with ArcGIS for deep learning on imagery?

Both run pretrained and your own deep learning models over imagery and blend predictions at tile borders. ArcGIS Online keeps results as hosted layers in ArcGIS and bills credits by pixels processed. RasterFlow writes Zarr and GeoParquet to your own storage and bills by data processed, calculated before each task runs.