About AI Labeling
Learn about AI Labeling in Oracle Spatial Studio on Autonomous AI Database.
Topics:
What is AI Labeling
AI Labeling in Oracle Spatial Studio prepares training data for deep learning by creating and managing labels directly on geospatial raster data, including airborne photos, satellite imagery, and any other gridded data.
Geospatial rasters are sourced from GeoRaster objects in Oracle Spatial Database or from geospatial imagery stored in Oracle Cloud Infrastructure (OCI) Object Storage. They are presented to the user as geospatial raster chips for labeling. Chips are raster tiles appropriate for training deep learning models, usually small in size. AI Labeling supports both georeferenced and non-georeferenced geospatial rasters.
AI Labeling supports the complete data preparation workflow for Spatial AI. It also includes:
-
Manual labeling of geospatial raster chips to mark features such as buildings, roads, or vegetation.
-
Loading labels from supported formats (for example, GeoJSON, YOLO, or mask images).
-
Storing labels and their corresponding raster chips in Autonomous AI Database, OCI Object Storage, or locally, to continue the workflow directly in Spatial AI.
-
Auto-labeling using geometry features either from Oracle Spatial databases or from Oracle-managed geometry data, which maps physical features on the ground.
These capabilities make Spatial Studio the primary tool for preparing labeled geospatial raster datasets for use in Spatial AI to train new models or apply transfer learning on existing models with additional data.
AI Labeling in Spatial Studio is typically used in the following scenarios:
-
To automatically generate initial labels through auto-labeling, followed by manual review and edit.
-
To manually prepare training datasets by labeling geospatial raster chips or refining auto-labeled results.
-
To leverage existing label data by loading labels from supported formats.
-
To continue the AI workflow by storing labels in the database, OCI Object Storage, or local files for use in Spatial AI.
Key Concepts in AI Labeling
Oracle Spatial Studio provides a labeling environment that integrates tightly with Spatial AI for preparing geospatial raster data for deep learning.
This section introduces key concepts used in AI Labeling.
Geospatial Raster
A geospatial raster is an image composed of a grid of pixels, each tied to a specific location on the Earth’s surface. In AI Labeling, geospatial rasters are the raw input for deep learning workflows. They may be georeferenced (tied to real-world coordinates) or non-georeferenced (treated as plain images).
Also, note that:
-
Geospatial rasters stored in OCI Object Storage are accessed as collections of raster chips, each chip being a separate image used for labeling.
-
Large geospatial rasters stored as GeoRaster objects in Oracle Spatial can be split into chips when creating a project in Spatial Studio, where the user defines the chip size for labeling and deep learning processing.
Chip
A chip is a smaller portion of a geospatial raster image and the basic unit of work in AI Labeling. Chips make it possible to apply machine learning and deep learning techniques to very large geospatial rasters by dividing them into manageable pieces.
In Spatial Studio, users label chips directly. These labeled chips provide the training data for deep learning models.
Label
A label assigns semantic meaning to geospatial raster data. In AI Labeling, labels can be:
-
Feature-based: Attached to specific regions in a chip, such as buildings, roads, or vegetation.
-
Chip-based: Applied to the entire chip, used for image classification (for example, labeling a chip as urban, forest, or water).
Labels are the ground truth used to train machine learning and deep learning models. They can be created manually in Spatial Studio, imported in supported formats, or generated automatically.
Auto-labeling
Auto-labeling is the process of automatically generating labels without manual drawing. It uses known features from maps to label geospatial raster chips. In Spatial Studio, auto-labeling can be performed in two ways:
-
Auto-labeling with Oracle-managed data: Uses an Oracle-managed set of supported feature types (such as buildings or water). These features are mapped to the label categories defined in the project.
-
Auto-labeling with user-managed data: Uses user-managed data containing SDO_GEOMETRY objects (that is, features) to generate labels aligned with geospatial raster chips.
Labeling Types
AI Labeling in Spatial Studio supports different labeling types, depending on the machine learning or deep learning task:
Image Classification
A single label is applied to an entire geospatial raster chip (for example, urban, forest, water).
Object Detection
Labels are applied to bounding boxes that mark the location and type of objects within a geospatial raster chip (for example, detecting buildings or vehicles).
Semantic Segmentation
Labels are applied at the pixel level, assigning each pixel in a geospatial raster chip to a category (for example, distinguishing between vegetation, road, and water areas).
Sparse Labeling (variant of Semantic Segmentation)
Instead of labeling all pixels, only a subset of pixels is labeled across the geospatial raster chip (for example, annotating key sample points of roads or vegetation). This approach reduces annotation effort while still guiding model training.
Instance Segmentation
Similar to semantic segmentation, but separates individual objects of the same category (for example, labeling multiple distinct trees within the same chip).
Nested Labels (Object Detection & Instance Segmentation)
For the project types discussed under “Labeling Types”, it is also possible to assign additional labels or attributes per instance. For example, in instance segmentation of buildings, each building polygon can include nested information such as height, usage, or other classification labels. This extends the concept of image classification to the level of individual detected objects.