Mosaics and image catalog
Can build mosaic datasets by flight, sensor or series (Create Mosaic Dataset, Add Rasters, overviews) and turn a stack of dated scenes into a multidimensional series.
Deliverable: ready catalog + coverage recordOcto for imagery, LiDAR and AI in ArcGIS
The model points. The site inspection confirms.
ArcGIS runs the whole imagery cycle by script: mosaic datasets, Spatial Analyst and Image Analyst through ArcPy, deep learning with ArcGIS Living Atlas models, Sentinel-2 time series and LiDAR point clouds. Octo can set up the environment, write and run the workflows, measure the model on a local sample and deliver lists and maps for your team to review. A model result is an indication, not proof: inspections, citations and technical reports stay with authorized people and the competent agency.
ArcGIS Pro, documentation3.7Windows only; deep learning requires an NVIDIA GPU
CCDC, minimum12slices over at least 1 year of series
Sources14listed at the end of the page
01
What Octo does with imagery and AI.
Six possible deliverables, drawn from the survey. All of them run on the client's license and machine (or portal), and all end in human review.
Deep learning detection
Can run Living Atlas or client models (Detect Objects, Classify Pixels, Detect Change) and regularize the features, in ArcGIS Pro or distributed in ArcGIS Image.
Deliverable: candidate layer with confidenceMeasuring the model before trusting it
Prepares the calibration sample and calculates hits, false positives and false negatives in your territory. If the model does not fit, proposes fine-tuning with Train Deep Learning Model.
Deliverable: accuracy report on the sampleTime series and change
Can run CCDC and change detection on Sentinel-2 or Landsat series, and work with climate and ocean data in NetCDF, GRIB and HDF through the Multidimensional toolbox.
Deliverable: change raster + date of changeLiDAR and 3D buildings
Can classify ground and buildings, generate terrain, footprints, height and 3D blocks (LAS Building Multipatch), and cross-check with the property registry.
Deliverable: 3D blocks + list of discrepanciesReport and record of each run
Can note the date, Pro and API version, model, thresholds and cadastral base of each run, and draft the technical report with method, limits and known errors.
Deliverable: report + reproducible record
02
Eight arms, one image analyzed.
The same Octo cycle, applied to imagery and AI. Each arm has a narrow role; what to do with a finding is decided by authorized people and the competent agency.
-
Discovery
Identifies what exists
ArcGIS Pro version, licensed extensions (Image Analyst, Spatial Analyst, 3D Analyst), GPU and driver, portal with or without ArcGIS Image, and which images, point clouds and registries are already available.
-
Data
Pins down reference and quality
Metric CRS, date and resolution of each flight or scene, cloud cover, LiDAR point cloud classes, date and status of the CAR database or registry. Without that, two years cannot be compared.
-
Security
Respects the client's data
Property registries and the CAR hold data about people: handling follows the agency's policy and the LGPD. Sending images or registries to an external service, including AI, only with the client's authorization.
-
Architecture
Chooses where to run
ArcGIS Pro on the GPU machine when the volume fits; ArcGIS Image or Image Server when it does not, always with a cost estimate first. Ready-made model, fine-tuning or new model, depending on what the sample shows.
-
Development
Writes the workflow
ArcPy and ArcGIS API for Python scripts with configuration separated from code and the documentation page noted for each tool used.
-
Review
Measures before trusting
Runs on a hand-checked sample and shows hits, false positives and false negatives. Checks by the result on the image, not by the absence of errors in the script.
-
Release
Delivers for review
Only with your OK, it runs the batch in the approved environment and leaves the list, table and maps right there, marked as indications, for the team to decide.
-
Documentation
Records what applies to your project
Pro and API version, model, thresholds, cadastral base and date of each run. The following month starts from this record, and a model calibrated in one municipality is not assumed ready for another.
Octo does not inspect, issue citations, impose embargoes or sign technical reports. A model result is an indication for inspection; the decision belongs to authorized people and the competent agency.
03
Each task, its path.
The criteria Octo uses to choose the tool. Tool names and parameters checked in the ArcGIS Pro documentation on October 29, 2026.
| Task | Path | Note |
|---|---|---|
| Cataloging and serving thousands of images | Mosaic dataset: Create Mosaic Dataset, Add Rasters, overviews | Standard or Advanced license; raster type per sensor |
| Indices and map algebra | arcpy.sa (Map Algebra) or raster functions | Spatial Analyst or Image Analyst |
| Extracting buildings from imagery | Detect Objects Using Deep Learning + .dlpk model; Regularize Building Footprint | Measure on a local sample first |
| Change between two dates | Detect Change Using Deep Learning, or detection per year and comparison | Change model trained with local samples |
| Change in a time series | Analyze Changes Using CCDC + Detect Change Using Change Analysis Raster | Minimum of 12 slices over 1 year |
| Statistics per polygon (property, parcel) | Zonal Statistics As Table, Tabulate Area | Metric CRS for area |
| Running at scale, without a local machine | ArcGIS Image or Image Server, via arcgis.learn and arcgis.raster.analytics | Estimate credits first (estimate=True) |
| Classifying LiDAR | Classify LAS Ground, Classify LAS Building or Classify Point Cloud Using Trained Model | The model needs a point cloud similar to the training one |
| 3D buildings | LAS Building Multipatch | Footprints from the registry or derived from the points |
| Climate and ocean (NetCDF, GRIB, HDF) | Make Multidimensional Raster Layer and multidimensional tools | Available even with the Basic license |
04
From assessment to monthly routine.
Three steps. Each one only makes sense after the previous one has shown results in your territory, measured on a sample.
A proposal, not a promise
Proof on a subset
A neighborhood, a watershed or a set of properties. Octo sets up the environment, runs the workflow and measures the model against what your team checked.
- Documented deep learning environment
- Hits, false positives and false negatives on the sample
- Your decision: continue, adjust the model or stop
Entire municipality or state
With the model approved, the workflow runs across the whole territory, in Pro with a GPU or distributed in ArcGIS Image, with the cost estimated first.
- Fine-tuning with local samples, if the proof calls for it
- Inspection list or table per property, with a map per item
- Technical report with method and limits
Routine with field feedback
The result of each inspection comes back as a labeled sample, and the model is measured and adjusted again each cycle.
- Monthly run with a record of version and thresholds
- Accuracy tracked over time
- 3D registry updated with each new flight
05
Confirmed by source.
Each point below has a source published by the vendor, the project or the responsible agency. The sources are in the next section.
4
points with a published source, each with its reference alongside.
- ConfirmedArcGIS Pro 3.7 and GPU
- The Pro 3.7 requirements list only Windows 11 and Windows Server. Deep learning requires an NVIDIA GPU with CUDA compute capability 5.0 at minimum (7.5 recommended) and driver 551.61 or newer. The Esri repository says Pro 3.7 and Enterprise 12.1 are the last versions to support Maxwell, Pascal and Volta GPUs.Source: ArcGIS Pro 3.7 system requirements; Esri/deep-learning-frameworks on GitHub
- ConfirmedPretrained models were not built for Brazil
- Building Footprint Extraction – USA (Mask R-CNN, average precision 0.79) was built for the United States; Land Cover Classification (Sentinel-2) (U-Net, average precision 0.82, Corine 2018 classes), for the United States, Europe and India. In ArcGIS Online, they require the Professional or Professional Plus user type.Source: model pages on doc.arcgis.com
- ConfirmedTime series
- Analyze Changes Using CCDC requires a multidimensional raster with at least 12 slices over at least 1 year and generates a change analysis raster in CRF; Detect Change Using Change Analysis Raster extracts the date of the most recent, earliest or largest change, with filters by year and magnitude.Source: Image Analyst tool reference
- ConfirmedDistributed processing
arcgis.learn.detect_objectsruns in ArcGIS Image (ArcGIS Online) and Image Server (Enterprise); withestimate=Trueit returns the credits required, in ArcGIS Online only.Source: ArcGIS API for Python reference
06
What Octo does and what stays with you.
Octo prepares everything, but does not sign. The document of record stays with whoever is accountable for it.
Octo prepares
Gathers, calculates, organizes and produces the documents: spreadsheets, reports, maps and drafts.
Deliverable: material ready for reviewYour team reviews
The licensed professional reviews what Octo delivered before signing. If something doesn't add up, it goes back to Octo with the correction.
Step: review by whoever signsWho signs
Octo does not sign. Designs, technical reports, ART/RRT, accountability reports and any official document are still signed by the licensed professional at the company or agency: here, the professional of record for the map or survey.
Signature: professional of record for the map or survey
07
Sources.
Esri and SICAR documentation used in the survey, accessed October 29, 2026.
- ArcGIS Pro 3.7: system requirementsOperating systems and GPU
- Esri/deep-learning-frameworksInstaller and conda path
- Detect Objects Using Deep LearningImage Analyst
- Detect Change Using Deep LearningImage Analyst
- Train Deep Learning ModelModel types
- Analyze Changes Using CCDCTime series
- Detect Change Using Change Analysis RasterDate and number of changes
- Create Mosaic DatasetMosaic datasets
- LAS Building Multipatch3D buildings from LiDAR
- Classify Point Cloud Using Trained ModelLiDAR with deep learning
- ArcGIS API for Python: arcgis.learnDistributed training and detection
- Building Footprint Extraction – USA modelLiving Atlas
- Land Cover Classification (Sentinel-2) modelLiving Atlas
- SICAR: public searchCAR database for download
Your territory, measured before scaling.
The assessment starts with a subset your team already knows: a neighborhood, a watershed, a set of properties. You see the model get it wrong and right there, with numbers, and decide with them in hand.