Skip to main content
View as Markdown

Technical Specification

Planet’s Vessel Detection and Classification products leverage advanced machine learning and computer vision techniques to identify maritime activity over vast areas of interest. Derived from high-frequency PlanetScope imagery, this solution provides automated, oriented bounding box detections of vessels, enabling customers to monitor maritime domains, secure coastlines, and analyze economic activity with precision.

Planet supports three Vessel Detection offerings:

  • Standard Vessel Detection: Only provides geojson feature collections containing polygons indicating the location and time vessels were detected along with associated metadata (for example, estimated length/width/heading, confidence score and other relevant information). This is a good fit for users with existing access to maritime data/imagery.
  • Standard Vessel Detection with Open Water Monitoring: Includes everything for Standard Vessel Detection but also includes open water imagery over the area of interest.
  • Enhanced Vessel Detection with Open Water Monitoring: Additional metadata is included for vessel classification and ship-to-ship (STS) event detection.
Standard Vessel DetectionStandard Vessel Detection with Open Water MonitoringEnhanced Vessel Detection with Open Water Monitoring
Geojson Feature Collection of Vessel Detections✅✅✅
Estimated Width, Length, Heading✅✅✅
Confidence Score✅✅✅
Imagery❌✅✅
Vessel Classification❌❌✅
Ship-to-ship (STS) event detection❌❌✅

Methodology​

Our vessel detection model provides reliable, high-quality results that you can trust to drive critical decisions. The process uses a sophisticated deep learning approach.

Validated Approach​

  • Segmentation: The model first performs a segmentation step on the PlanetScope imagery. It analyzes the image pixel by pixel to create a probability mask that highlights all areas likely to contain a vessel.
  • Polygonization and feature extraction: The model then processes this mask to generate a precise, object-oriented bounding box around each probable vessel. During this step, it calculates key metadata for each detection, including the detection confidence, estimated length, width, and heading, and centroid coordinates. This two-step process is highly effective at identifying vessels longer than 25 m while minimizing false positives from other features on the ocean's surface.
  • Classification and Ship-to-Ship Events: For Enhanced Vessel Detection, we apply a secondary classification step that adds a predicted class property to the result, including Tanker, Cargo, and Military classes. We also flag vessels assessed to be in a ship-to-ship (STS) transfer event, with an associated STS event confidence score.

Robust Training Data​

A model is only as good as the data it is trained on. Planet Vessel Detection is trained on a massive, hand-labeled training dataset curated from our extensive imagery archive. This dataset features a broad geographic and temporal distribution, encompassing thousands of images from different oceans, seasons, and times of day. This diversity ensures the model is robust and performs reliably across a wide variety of real-world conditions, including different sea states, atmospheric haze, sun angles, and vessel types. It is continually updated and expanded to reduce model drift and address issues as they arise.

Example of vessel detection

Reliable Performance​

Our model is rigorously tested to ensure it meets the demanding needs of our users. Planet leveraged 80% of the above dataset to train the segmentation model, holding out 20% of the labeled dataset for validation to measure the model's performance. By comparing the model's results over this 20% holdout dataset, we can calculate the following precision, recall, and F1 measurements specifically over open water areas for vessels greater than 25 m in length.

  • Precision: 0.89 - This measures the accuracy of our detections. A high precision score means that when we identify a vessel, it is highly likely to be an actual vessel, minimizing false alarms.
  • Recall: 0.84 - This measures the completeness of our detections. A high recall score means we successfully identify a high percentage of the actual vessels present in the imagery.
  • F1-Score: 0.86 - This provides a single, balanced measure of the model's overall accuracy by combining both Precision and Recall.

Delivery Mechanisms​

Results are available via the Analytics API and Feed Viewer. For more details please see:

Full Metadata Reference​

Properties schema update

On October 2, 2026, the Vessel Detection properties schema was updated and the previous schema was retired. If your integration was built before this date, see Migrating from the previous schema and the Changelog.

Feature-level Fields​

PropertyFormatExampleDefinition
createdISO 8601 DateTime"2026-02-03T19:02:45.270966Z"The timestamp when the feature was created in the system
geometryGeoJSON Geometry{"type": "Polygon", "coordinates": [...]}The oriented bounding box of the detected vessel
idUUID"17f21835-93c4-4664-b2ef-c2f57f5809a5"The unique identifier for the feature
linksArray of Link Objects[{"href": "...", "rel": "self", "type": "..."}]Array of related resources including visual tiles, quads, process info, and observation data
typeString"Feature"The GeoJSON feature type

Properties​

PropertyFormatExampleDefinition
categoryString"vessel"The category of the detected object
centroid_latFloat13.8983342Latitude of the detection centroid (WGS84)
centroid_lonFloat52.7611633Longitude of the detection centroid (WGS84)
classification_confidenceFloat or null0.996Confidence in the assigned object_class_label (0 to 1). null without Enhanced Vessel Detection
created_atISO 8601 DateTime"2026-10-06T14:48:10.956313+00:00"When the detection was created by the processing pipeline
derived_heading_degFloat213The estimated heading direction of the vessel in degrees (0 to 360)
derived_length_mFloat26.83The estimated length of the vessel in meters
derived_width_mFloat7.61The estimated width of the vessel in meters
detection_confidenceFloat0.718The model's confidence that the object is a vessel (0 to 1, higher values indicate greater certainty)
geometry_accuracyString"Medium"Confidence level in the georeferencing of the source image. Values: Low, Medium, High, derived from the georeferencing techniques used for the scene
geometry_revisionInteger2Whether the geometry is final or waiting for additional metadata to improve positioning.
1 (Provisional): work in progress; the detection may still change.
2 (Refined): finalized or re-localized; no further updates expected
idString"12c5ddd7cc2aa196bce90fe2768cda97d653c25e"Hash of the source scene, model version, detection confidence, and geometry, generated each time vessel detection runs on a scene. Not the same as the feature id
object_class_labelString"Tanker"The classification label assigned to the detected object. With Enhanced Vessel Detection, values are Title Case class names (for example, "Tanker"). Without Enhanced Vessel Detection, the value is "vessel"
observedISO 8601 DateTime"2026-09-06T07:50:42.375742Z"The timestamp when the vessel was observed
processed_atISO 8601 DateTime"2026-10-06T14:48:10.956313+00:00"When the detection was last processed. An update to the georeferencing can trigger reprocessing and a new processed_at
scoreFloat0.718The model's confidence in the detection accuracy (0 to 1 scale, where higher values indicate greater certainty)
source_asset_typeString"ortho_visual"The type of asset used for detection
source_cloud_coverFloat21The cloud cover percentage in the source imagery (0 to 100 scale)
source_image_mean_gsdFloat3.0The mean ground sample distance of the source imagery in meters
source_item_idString"20260906_075042_37_250d"The unique identifier of the source imagery item
source_item_typeString"PSScene"The type of source imagery item
source_scene_extentWKT Polygon"POLYGON ((52.7326 14.0556, 52.7002 13.8978, 52.9727 13.8436, 53.0056 14.0023, 52.7326 14.0556))"The geographic extent of the source imagery as a polygon
sts_eventBooleanfalseIndicates whether the vessel is assessed to be in a ship-to-ship (STS) transfer event. Always false without Enhanced Vessel Detection
sts_event_confidenceFloat or null0.018Confidence that the vessel is in a ship-to-ship transfer event (0 to 1). null without Enhanced Vessel Detection

Migrating from the previous schema​

Previous propertyCurrent propertyAction
angleNoneRemoved. Use derived_heading_deg
area_m2NoneRemoved. Approximate with derived_length_m × derived_width_m
bunkeredsts_eventRename
diagonal_mNoneRemoved. Compute from length and width if needed
headingderived_heading_degRename. Value may now include decimals
length_mderived_length_mRename
model_idNoneRemoved
model_versionNoneRemoved
object_class_labelobject_class_labelWith Enhanced Vessel Detection, values are now Title Case. Update case-sensitive comparisons. Without Enhanced Vessel Detection, the value is "vessel"
source_cloud_coversource_cloud_coverScale changed from a 0 to 1 range to a 0 to 100 range. Update thresholds (for example, change 0.2 to 20)
source_extentsource_scene_extentRename
source_quality_categoryNoneRemoved
source_sun_azimuth_angleNoneRemoved
source_sun_elevation_angleNoneRemoved
width_mderived_width_mRename

All other properties are unchanged.

With Planet Vessel Detection, you gain a robust, persistent, and reliable view of the maritime domain, empowering you to act with speed and confidence.

Changelog​

DateChange
October 2, 2026Breaking change: properties schema updated.
Renamed length_m, width_m, heading, bunkered, and source_extent.
Added detection_confidence, classification_confidence, sts_event_confidence, centroid_lat, centroid_lon, geometry_accuracy, geometry_revision, created_at, processed_at, and id.
Removed angle, area_m2, diagonal_m, model_id, model_version, source_quality_category, source_sun_azimuth_angle, and source_sun_elevation_angle.
source_cloud_cover is now 0 to 100.
object_class_label values are now Title Case with Enhanced Vessel Detection, and "vessel" otherwise.
See Migrating from the previous schema.