What is Geospatial Intelligence? GEOINT Defined, Examples

Geospatial intelligence (GEOINT) is the analysis of images and map data to describe physical features and the activity around them on earth. US law defines it that way in 10 U.S.C. § 467, and splits it into three parts: imagery, imagery intelligence, and geospatial information. This article argues that GEOINT's value comes from joining those parts. An image shows what is there today. A map records what someone already charted. Each one misses what the other holds. A test on the Mississippi River at New Orleans shows how much. Aerial imagery from 2021 showed 175 vessel hulls, outlines of one vessel or of several moored side by side. Only 19 of them sit within 30 meters of a pier or quay in the open map.

Key takeaways

• Geospatial intelligence is a legal term in the United States. It means analyzing images and map data, in three parts: imagery, imagery intelligence (what an image shows), and geospatial information (maps and charts).
• The discipline grew from reading reconnaissance photos (the CORONA satellites from 1960, the Cuban Missile Crisis in 1962) into one agency. The National Imagery and Mapping Agency (NIMA) formed in 1996 and was renamed the National Geospatial-Intelligence Agency (NGA) in 2003.
• Detection models now do part of the work of reading images. SAM 3, a model that finds objects in images, outlines every object that matches a short text prompt such as "ship" or "barge."
• In Wherobots, SAM 3 found 175 vessel hulls in 2021 and 156 in 2023. The imagery is 30 cm aerial photography from the National Agriculture Imagery Program (NAIP) of the US Department of Agriculture (USDA). The study area is a 28.09 km² stretch of the Mississippi at New Orleans. Only 19 and 21 of the hulls sit within 30 m of a pier or quay in Overture Maps. Most of the waterfront where vessels moor is blank in the map.
• A visual review of 60 randomly sampled hulls found a vessel or a vessel-like floating structure in 59 of them. Barges moored side by side (rafted) merge into one outline, so a hull count is a count of moored groups.

What is geospatial intelligence?

The US Code gives the working definition. Section 467 of Title 10 says that "the term 'geospatial intelligence' means the exploitation and analysis of imagery and geospatial information to describe, assess, and visually depict physical features and geographically referenced activities on or about the earth." It adds that GEOINT "consists of imagery, imagery intelligence, and geospatial information." The same section defines each part:

• Imagery is "a likeness or presentation of any natural or manmade feature or related object or activity and the positional data acquired at the same time." That includes images from satellites, aircraft, and drones. The positional data matters: an image without a location cannot be matched to anything.
• Imagery intelligence is "the technical, geographic, and intelligence information derived through the interpretation or analysis of imagery and collateral materials." This is the step where a person or a model says what is in the image, checked against other reports where they exist.
• Geospatial information "identifies the geographic location and characteristics of natural or constructed features and boundaries on or about the earth." It comes from remote sensing, mapping, and surveying.

The National Geospatial-Intelligence Agency (NGA) publishes the discipline's doctrine, its formal guidance on how GEOINT is done. NGA's GEOINT Basic Doctrine, Publication 1.0 calls itself "the highest level of doctrine for GEOINT." It is unclassified and written for people who use GEOINT or are new to producing it. A military manual, Joint Publication 2-03, covers geospatial intelligence support to joint operations.

The definition is broad. It names no sensor and no customer. A port authority counting berth use follows the same three steps as an insurer checking roof damage after a storm or a defense analyst counting vessels at a naval pier. Each one gets imagery, finds what it shows, and places the result against mapped features.

Geospatial intelligence vs earth intelligence vs GIS

Three terms overlap in search results and in product pages. They answer different questions.

	Geospatial intelligence (GEOINT)	Earth intelligence	Geographic information system (GIS)
What it is	An analytic discipline with a legal definition in 10 U.S.C. § 467	A commercial term for running models on satellite and aerial imagery to produce layers and counts	Software and data for storing, editing, and mapping geospatial data
Typical inputs	Imagery, imagery intelligence, maps, charts, other intelligence reports	Imagery and model outputs: detections, segmentation masks, classifications	Vector and raster layers
Typical output	An assessment of activity at a place and time	A layer, such as field boundaries or roads, or a count	A map, an edited dataset, a spatial query result
Who uses the term	Defense and intelligence agencies, NGA, the armed services, contractors	Imagery and analytics companies	Every sector that maps anything

Earth intelligence covers the model step of GEOINT and often ends with the layer the model produces. GEOINT goes on to match that layer to maps and to judge what it means. GIS supplies much of the map data and the tools both of them use.

How geospatial intelligence works

A GEOINT question moves through the three parts of the definition in order.

1. Collect imagery. Satellites, aircraft, and drones record the scene with a timestamp and a location. The imagery can be optical, multispectral (several color and infrared bands), or radar. Its resolution sets what can be found: 30 cm aerial imagery shows individual barges and tugs, while at 10 m a 100 m ship spans about 10 pixels.
2. Derive imagery intelligence. An analyst or a model finds objects and changes: vessels, aircraft, vehicles, new buildings, cleared land. This step used to be all photo interpretation. Today detection and segmentation models (models that outline each object) produce outlines with a confidence score. Analysts review the results.
3. Combine with geospatial information. Each detection is matched to mapped features by location. That might be the pier it sits beside, the terminal it belongs to, the road it blocks, or the boundary it crosses. This spatial join turns a shape on an image into a fact about a named place.
4. Assess and depict. The analyst counts, compares across dates, and maps the result. The definition's verbs, describe, assess, and visually depict, all happen here.

Step 3 is where most of the difficulty lies at scale. Imagery arrives as rasters (grids of pixels) in one coordinate system. Maps arrive as vectors (points, lines, and shapes) in another. The two are updated on different schedules. A system that holds raster and vector data in one engine can run steps 2 to 4 without exporting between tools.

A short history of geospatial intelligence

Imagery intelligence predates the term GEOINT by four decades. The CORONA program returned no usable imagery on its first 13 missions. Then, on 18 August 1960, Mission XIV returned "more photographic coverage of the Soviet Union in a single mission than in all previous U-2 missions combined." By its 145th and final launch on 25 May 1972, CORONA had photographed about 750 million square miles of the earth's surface. Two years after that first success, analysts at the National Photographic Interpretation Center played what NGA calls "the crucial role" in the Cuban Missile Crisis.

Mapping and imagery stayed in separate agencies until the 1990s. On 1 October 1996 the National Imagery and Mapping Agency (NIMA) absorbed the Defense Mapping Agency, the National Photographic Interpretation Center, and the Central Imagery Office. It also took in imagery units of the CIA, the Defense Intelligence Agency, and the National Reconnaissance Office. In 2003 Congress renamed NIMA the National Geospatial-Intelligence Agency. NGA's history ties the new name to "a new intelligence discipline," GEOINT. The definition in 10 U.S.C. § 467(5) dates from the same law, Pub. L. 108-136 of 24 November 2003. Pub. L. 116-283 of 1 January 2021 widened it from activities "on the earth" to activities "on or about the earth."

The detection step has moved from people to models in the last decade. Benchmark datasets, large sets of labeled images for training and testing models, made that possible. xView (Lam et al., 2018) labeled over 1 million objects in 60 classes across 1,400 km² of 0.3 m WorldView-3 satellite imagery. DOTA (Xia et al., 2017), a dataset for object detection in aerial images, collected 2,806 aerial images labeled by experts. SpaceNet (Van Etten et al., 2018) released labeled satellite imagery with a series of prize competitions for map features. Models that take a prompt came next. Segment Anything (Kirillov et al., 2023) trained an outlining model on over 1 billion masks, the pixel outlines of objects. It works on new images without retraining. SAM 3 (Carion et al., 2025) finds and outlines every object that matches a short phrase, such as "ship."

Geospatial intelligence examples

• Maritime domain awareness, knowing which vessels are at sea and where. Paolo, Kroodsma and colleagues combined satellite imagery, vessel GPS data, and deep learning models over the world's coastal waters from 2017 to 2021 (Nature, 2024). They found that 72 to 76% of the world's industrial fishing vessels are not publicly tracked. They also found that 21 to 30% of transport and energy vessel activity is missing from public tracking systems. The xView3-SAR benchmark (Paolo et al., 2022) targets the same gap with synthetic aperture radar (SAR), which images through clouds and at night.
• Port and terminal activity. Vessel counts per berth across dates show which terminals are busy. The New Orleans walkthrough below is this example on open data.
• Disaster response. NGA describes humanitarian and disaster relief work "in conjunction with such organizations as FEMA," the Federal Emergency Management Agency. Damage assessment joins post-event imagery to building footprints and roads.
• Construction and facility monitoring. Repeated imagery of one site shows new piers, buildings, and vessels. The Hainan comparison later in this article uses commercial satellite imagery of a naval port for this purpose.
• Agriculture and infrastructure. Field boundaries, roads, and buildings detected from imagery fill gaps in maps; see satellite imagery analysis and building footprints.

Geospatial intelligence software

GEOINT software covers the same three parts as the definition. Imagery management stores and serves large archives of scenes. Detection and segmentation models produce imagery intelligence; open models such as SAM 3 can run on a team's own imagery. Spatial databases and GIS tools hold the geospatial information and run the joins. The practical question for a team is how many of those steps run in one system. Each export between tools adds a format conversion, a coordinate system check, and another copy to keep current.

In Wherobots, RasterFlow (Public Preview) stitches images into mosaics, runs vision models on them, and saves the results as shapes (vectors) on the same platform. WherobotsDB then matches those shapes to map data in spatial SQL. The Wherobots MCP server connects AI assistants to Wherobots, so an AI agent can build the same GEOINT pipelines. The walkthrough below uses RasterFlow and WherobotsDB.

Geospatial intelligence in Wherobots: vessels on the Mississippi at New Orleans

The experiment tests the article's claim. If vessels detected in imagery mostly sit at piers and quays the open map already holds, the imagery adds little the map lacks. If most sit where the map has no feature, the information comes from combining imagery and map.

Sources

Source	Format	Size	What the job read
USDA NAIP, Louisiana 2021 survey	30 cm aerial imagery, red, green, and blue bands plus near infrared	10 scenes flown 29 November 2021 and 30 January 2022	Red, green, and blue bands inside the study area
USDA NAIP, Louisiana 2023 survey	30 cm aerial imagery	10 scenes flown 1, 3 and 4 May 2023	Red, green, and blue bands inside the study area
Overture Maps base theme, water	Polygons	One "Mississippi River" polygon, 13.25 km² inside New Orleans	Study area: that polygon buffered 200 m, 28.09 km²
Overture base theme, infrastructure	Lines and polygons	108 piers (9 named) and 2 quays in the study area	Piers and quays near each hull
Overture Maps places	Points	27 places named as a wharf, terminal or dock within 500 m of the study area	Terminal names for each hull

The study area follows the river inside the Overture New Orleans boundary. The 200 m buffer reaches the wharves on both banks, where vessels moor.

Method

The RasterFlow job ran SAM 3 on each survey's 30 cm NAIP mosaic with the text prompts "ship" and "barge". It saved every detection as a polygon with a label and a score. Detection took 1,773 seconds for the 2021 survey and 1,230 seconds for 2023. SAM 3 returned 3,775 polygons for 2021 and 4,074 for 2023. About three quarters of them (74.7% and 75.8%) were slivers under 1 m². Only 22.9% and 19.4% scored 0.5 or higher. The SQL below keeps detections with a score of at least 0.5 and an area of at least 100 m². It merges overlapping outlines into hulls and matches each hull to Overture piers, quays, and terminal names. Distances are measured edge to edge in meters on the Universal Transverse Mercator (UTM) zone 15N map grid. So a hull moored along a long pier counts as touching it. The thresholds are this article's choice, set after reviewing score and area distributions. They are not documented model defaults.

The three input tables come from two steps. First, Overture gives the study area: the "Mississippi River" water polygon inside the New Orleans boundary, buffered 200 m.

-- Study area: the Mississippi River inside New Orleans, plus 200 m
CREATE TABLE org_catalog.geoint_walkthrough.aoi_nola AS
WITH nola AS (
  SELECT geometry FROM wherobots_open_data.overture_maps_foundation.divisions_division_area
  WHERE country = 'US' AND region = 'US-LA' AND subtype = 'locality' AND names.primary = 'New Orleans'
),
river AS (
  SELECT ST_Intersection(ST_Union_Aggr(w.geometry), FIRST(n.geometry)) AS geom
  FROM wherobots_open_data.overture_maps_foundation.base_water w JOIN nola n ON ST_Intersects(w.geometry, n.geometry)
  WHERE w.names.primary = 'Mississippi River' AND ST_GeometryType(w.geometry) IN ('ST_Polygon', 'ST_MultiPolygon')
)
SELECT ST_MakeValid(ST_Buffer(geom, 200, true)) AS aoi FROM river;

Second, the RasterFlow SAM 3 recipe runs on that area once per survey and the detections are saved as a table. This is the call for the 2021 survey; the 2023 survey uses a window from 1 January 2023 to 1 January 2024 and writes sam3_nola_2023.

from datetime import datetime
import pyspark.sql.functions as f
from rasterflow_remote import RasterflowClient
from rasterflow_remote.data_models import GeometryModelRecipes
from sedona.spark import SedonaContext

sedona = SedonaContext.create(SedonaContext.builder().getOrCreate())
aoi_wkt = sedona.sql("SELECT ST_AsText(aoi) FROM org_catalog.geoint_walkthrough.aoi_nola").first()[0]
out = RasterflowClient().predict_mosaic_geometries_recipe(
    aoi=aoi_wkt, start=datetime(2021, 6, 1), end=datetime(2022, 6, 1), target_crs=3857,
    model_recipe=GeometryModelRecipes.SAM3_TEXT_GEOMETRY,
    text_prompt=["ship", "barge"], confidence_threshold=0.3)
(sedona.read.format("geoparquet").load(out.uri).withColumn("survey", f.lit("2021"))
    .writeTo("org_catalog.geoint_walkthrough.sam3_nola_2021").createOrReplace())

Replace org_catalog.geoint_walkthrough with a catalog and database in your own organization. With those three tables in place, the SQL below reproduces the results.

-- Detections worth keeping: score >= 0.5 and at least 100 square meters
CREATE OR REPLACE TEMP VIEW kept AS
SELECT survey, label, bbox_score, geometry FROM org_catalog.geoint_walkthrough.sam3_nola_2021
WHERE bbox_score >= 0.5 AND ST_AreaSpheroid(geometry) >= 100
UNION ALL
SELECT survey, label, bbox_score, geometry FROM org_catalog.geoint_walkthrough.sam3_nola_2023
WHERE bbox_score >= 0.5 AND ST_AreaSpheroid(geometry) >= 100;

-- Hulls: overlapping ship and barge outlines merged per survey
CREATE OR REPLACE TEMP VIEW hulls AS
SELECT survey, explode(ST_Dump(g)) AS hull
FROM (SELECT survey, ST_Union_Aggr(geometry) AS g FROM kept GROUP BY survey);

-- Overture piers and quays in the study area
CREATE OR REPLACE TEMP VIEW piers AS
SELECT i.class, i.names.primary AS name, i.geometry
FROM wherobots_open_data.overture_maps_foundation.base_infrastructure i,
     org_catalog.geoint_walkthrough.aoi_nola a
WHERE i.bbox.xmin > -90.16 AND i.bbox.xmax < -89.88 AND i.bbox.ymin > 29.85 AND i.bbox.ymax < 29.98
  AND i.class IN ('pier', 'quay') AND ST_Intersects(i.geometry, a.aoi);

-- Overture places named as a wharf, terminal or dock within 500 m of the study area
CREATE OR REPLACE TEMP VIEW terminals AS
SELECT p.id, p.names.primary AS name, p.geometry
FROM wherobots_open_data.overture_maps_foundation.places_place p,
     org_catalog.geoint_walkthrough.aoi_nola a
WHERE p.bbox.xmin > -90.16 AND p.bbox.xmax < -89.88 AND p.bbox.ymin > 29.85 AND p.bbox.ymax < 29.98
  AND ST_Intersects(p.geometry, ST_Buffer(a.aoi, 500, true))
  AND lower(p.names.primary) RLIKE '\\b(wharf|terminal|dock)\\b'
  AND NOT lower(p.names.primary) RLIKE 'cafe|parking';

-- One row per hull: area, within 30 m of a pier or quay, nearest terminal name within 500 m.
-- Distances are edge to edge in meters on UTM zone 15N (EPSG:32615).
WITH h AS (SELECT survey, hull, ST_AreaSpheroid(hull) AS m2, md5(ST_AsText(hull)) AS hid,
                  ST_Transform(hull, 'EPSG:4326', 'EPSG:32615') AS hull_m FROM hulls),
p AS (SELECT ST_Transform(geometry, 'EPSG:4326', 'EPSG:32615') AS geom_m FROM piers),
t AS (SELECT name, ST_Transform(geometry, 'EPSG:4326', 'EPSG:32615') AS geom_m FROM terminals),
pier AS (
  SELECT /*+ BROADCAST(p) */ h.hid, MAX(CASE WHEN p.geom_m IS NOT NULL THEN 1 ELSE 0 END) AS at_pier
  FROM h LEFT JOIN p ON ST_DWithin(h.hull_m, p.geom_m, 30) GROUP BY h.hid
),
near AS (
  SELECT hid, name FROM (
    SELECT /*+ BROADCAST(t) */ h.hid, t.name,
           ROW_NUMBER() OVER (PARTITION BY h.hid ORDER BY ST_Distance(h.hull_m, t.geom_m)) AS rn
    FROM h JOIN t ON ST_DWithin(h.hull_m, t.geom_m, 500)) WHERE rn = 1
)
SELECT h.survey, h.hid, h.hull, h.m2, pier.at_pier, near.name AS terminal
FROM h JOIN pier ON h.hid = pier.hid LEFT JOIN near ON h.hid = near.hid;

Results

	2021 survey	2023 survey
Detections kept (ship / barge)	288 (180 / 108)	317 (161 / 156)
Hulls after merging overlaps	175	156
Median hull area	712 m²	836 m²
90th percentile hull area	6,324 m²	5,883 m²
Hulls within 30 m of an Overture pier or quay	19	21
Hulls within 500 m of a named terminal	38	34

The verdict supports the claim. Of 175 hulls in 2021, 19 sit within 30 m of a pier or quay in Overture; in 2023 the share is 21 of 156. Most hulls along the working waterfront sit where the infrastructure layer has no pier or quay. Overture places does hold the terminal names, as points with no outline. That is enough to label 38 hulls in 2021 and 34 in 2023 with the nearest named wharf, terminal, or dock within 500 m.

The table lists every name the join used, with hull counts per survey. The names are as published in Overture.

Nearest named place in Overture (name as published)	2021 hulls	2023 hulls
Algiers Ferry Terminal	4	7
1St Street Wharf, Port Of New Orleans	2	4
Perry Street Wharf	2	4
Nashville Ave. Wharf	4	1
Poland Avenue Wharf	3	2
Turn Services Crew Boat Dock @ Associated Terminals Chalmette Slip	4	1
Magellan (Hess) Dock	3	2
New Orleans Port Authority Napoleon Terminal	3	1
Port of New Orleans Container Terminal	2	2
Westway Terminal Company	2	1
Robin Street Wharf	1	2
Saint Bernard Port & Harbor Terminal	2	1
Cargil Grain Westwego Dock	1	1
Delta Terminal Services	2	0
Louisiana International Terminal St. Bernard	0	2
Toulouse Street Wharf	1	1
Jackson Avenue Ferry Terminal	0	1
Julia Street Cruise Terminal	1	0
Kinder Morgan Delta Terminal	1	0
Piety Wharf	0	1
Total labeled	38	34

Each survey is one set of flight days, so a difference of one to three hulls at a terminal is a snapshot. It is not a trend. A real activity series needs repeated imagery of the same terminals.

Spot-check of the detections

We reviewed 60 randomly sampled hulls, 30 per survey, in 140 m crops from the same NAIP mosaics SAM 3 ran on. A vessel or a vessel-like floating structure appeared in 59 of them. The one unclear crop shows a small shape at the bank, and two of the 59 may be a docking barge and a floating drydock. In one crop SAM 3 outlined a towboat and missed the four barges it was pushing. Two kinds of outline error appear. Barges rafted side by side merge into one hull, so a hull is a moored group, and a hull count is lower than a barge count. Some outlines also spill onto a pier or a crane next to the vessel. The review measures precision, the share of hulls that mark a vessel. It does not measure recall, the share of vessels SAM 3 missed. That needs an independent count. Automatic Identification System (AIS) records, the radio position reports ships broadcast, cover only part of this fleet. The International Maritime Organization (IMO) requires AIS on all passenger ships. It also requires AIS on ships of 300 gross tonnage and upwards on international voyages, and on cargo ships of 500 gross tonnage and upwards on other voyages. National rules for river towboats and barges differ, so how much of a river fleet transmits AIS needs its own check.

A second example: commercial imagery of a naval port

The same workflow runs on commercial satellite imagery. The Wherobots post Yulin Naval Base in Satellite Imagery: AI-Assisted Vessel Counts applies SAM 3 to Vantor multispectral satellite images of the Yulin naval port on Hainan, taken on five dates between May 25 and June 8, 2026. It stores the vessel detections as rows by area and date. The setting changes the inputs: satellite imagery with coarser pixels than NAIP and several dates within two weeks. The steps stay the same: imagery, detections from a text prompt, a join to mapped features, and a count per place and date. The New Orleans walkthrough above is the reproducible version on open data.

What changes at scale

• The join grows with the area. One survey of 28.09 km² at New Orleans produced about 4,000 polygons. A whole coastline or river system multiplies that count. Each detection must still be matched to piers, roads, and boundaries, ideally in the system that already holds the imagery and the map.
• Repeat coverage sets what can be learned. The two 30 cm NAIP surveys of this stretch were flown 15 to 17 months apart. Each shows the terminals on a few days. Daily or weekly activity needs imagery with a shorter revisit, at whatever resolution still resolves the vessels in question.
• Tracking data fills part of the gap. AIS covers large ships and passenger vessels by IMO rule. Paolo and colleagues found that 72 to 76% of industrial fishing vessels are not publicly tracked. Vessels show up in imagery whether or not they transmit.
• Open maps are updated more slowly than the waterfront changes. According to its guide, most features in Overture's base theme come from OpenStreetMap. Mooring areas and berths can change faster than they are mapped. Outlines found in imagery can be added back to the map.

Open problems

• Counting individual vessels in rafted fleets. Merged barge outlines need a model that separates each barge, or an estimate from the outline's size.
• Measuring missed vessels without a reference count. For vessels that carry no AIS, an independent count needs repeated manual review or a second sensor.
• Change across sensors. Comparing 30 cm aerial imagery with 2 m satellite imagery changes what a model can find. A count difference may come from the sensor and not the scene.

Conclusions

The New Orleans test and the legal definition agree. Geospatial intelligence depends on matching imagery to maps, and in New Orleans the open map was missing most of the places where vessels moor.

1. Treat GEOINT as three parts joined by location. The statute names imagery, imagery intelligence, and geospatial information, and the walkthrough needed all three: NAIP imagery, SAM 3 outlines, and Overture piers and terminal names.
2. Expect imagery to show activity the map leaves blank. Only 19 of 175 hulls in 2021 and 21 of 156 in 2023 sat within 30 m of a mapped pier or quay. If a question is about where vessels are, start from imagery and use the map to name places.
3. If the map has names as points, join by distance. Overture places labeled 38 of 175 hulls with the nearest wharf, terminal, or dock within 500 m, with no terminal outlines.
4. Report hulls, not vessels, for barge traffic. Rafted barges merged into one outline in the spot-check, so a hull count understates the barge count. Quote barges only after separating instances.
5. Quote precision from a visual review and leave recall open. The review found a vessel or vessel-like floating structure in 59 of 60 sampled hulls. Recall needs an independent count. AIS coverage of river barges and towboats has to be checked before AIS can serve as one.
6. Compare dates only with repeated imagery. One flight window per survey shows a snapshot, and a difference of one to three hulls at a terminal can come from the flight day alone.

Related terms

• Satellite imagery analysis: extracting information from satellite images
• Earth observation data: imagery and measurements of the earth from space and air
• Geospatial AI: machine learning and AI agents applied to spatial data
• Spatial join: the operation that ties detections to mapped features
• Raster vs vector data: the two data models a GEOINT workflow combines
• Building footprint: another layer extracted from imagery

Read more from Wherobots

• Yulin Naval Base in Satellite Imagery: AI-Assisted Vessel Counts
• Wherobots MCP Server: building GEOINT spatial pipelines with AI agents
• RasterFlow is now available in Public Preview
• How we delivered Fields of The World with RasterFlow
• Detecting objects from text prompts with RasterFlow

Run SAM 3 on your own area of interest with a Wherobots free trial at cloud.wherobots.com.