Satellite Image Analysis

Satellite image analysis is the extraction of information from images of Earth captured by orbiting satellites: mapping land cover, counting ships, aircraft, or buildings, measuring crop health, and detecting what changed between two dates. It is a branch of remote sensing, which also covers imagery from aircraft and drones, and today it relies on computer vision (CV) and deep learning (DL) to process volumes of imagery no team of analysts could inspect by hand. The Landsat program alone has imaged Earth continuously since 1972.

What Satellites Capture#

Satellite images differ from ordinary photos along four kinds of resolution:

• Spatial resolution: The ground distance covered by one pixel. Landsat 8 and 9 record most bands at 30 m, Sentinel-2 reaches 10 m, and commercial satellites such as WorldView-3, operated by Vantor (named Maxar Intelligence until October 2025), resolve about 30 cm, enough to see individual cars.
• Spectral resolution: The number and width of wavelength bands. Sentinel-2 records 13 bands from visible light to shortwave infrared.
• Temporal resolution: How often the same place is imaged. The two-satellite Sentinel-2 constellation revisits the equator every five days, and Planet's fleet of small satellites images nearly all of Earth's land surface daily at about 3 m.
• Radiometric resolution: How finely each pixel records brightness, often 11 to 14 bits, compared with 8 bits in a standard photo.

Two other sensor types extend what satellites can measure. Hyperspectral imaging satellites record hundreds of narrow bands, and synthetic aperture radar (SAR) satellites such as Sentinel-1 send out microwave pulses and record the echoes, so they see through clouds and at night, which makes radar a primary source for mapping floods under storm cover. Band math turns spectral data into physical measurements: the normalized difference vegetation index, NDVI = (NIR − Red) / (NIR + Red), approaches 1 over dense, healthy vegetation because leaves reflect near-infrared light strongly and absorb red.

Core Analysis Tasks#

• Semantic segmentation: Labels every pixel to map land cover, building footprints, roads, or flood extent.
• Object detection: Finds and counts discrete objects such as ships, planes, vehicles, and storage tanks. Seen from above, these objects are small, densely packed, and pointed in every direction, so detectors often predict oriented bounding boxes that rotate to fit each object.
• Change detection: Compares images of the same place from different dates to find deforestation, new construction, or disaster damage.
• Scene classification: Assigns one label, such as farmland, forest, or urban, to each image patch.

Detection datasets for overhead imagery include DOTA, whose first version annotates 15 categories with arbitrarily oriented boxes, and xView, which labels more than 1 million objects in 60 classes on 0.3 m WorldView-3 imagery. Ultralytics YOLO26 oriented bounding box models are pretrained on DOTAv1 and predict a rotated box for each object:

from ultralytics import YOLO

# Load a YOLO26 oriented bounding box (OBB) model pretrained on DOTAv1
model = YOLO("yolo26n-obb.pt")

# Detect objects in an overhead image
results = model("https://ultralytics.com/images/boats.jpg")
result = results[0]

# Save the image with a rotated box drawn around each detection
result.save(filename="boats_obb.jpg")

Rotated boxes follow each vessel's heading, which keeps tightly moored boats from sharing one oversized box. Geospatial foundation models pretrained on large archives without labels, such as NASA and IBM's Prithvi models built on Harmonized Landsat and Sentinel-2 data, are fine-tuned for tasks such as flood and burn-scar mapping with far fewer labeled examples.

From Raw Scene to Map#

A single Sentinel-2 tile spans 10,980 by 10,980 pixels in its 10 m bands, far larger than a model's input, and it arrives as a georeferenced file such as a GeoTIFF whose pixels map to coordinates on Earth. Analysis pipelines correct for the atmosphere, mask clouds, cut the scene into overlapping tiles, run the model on each tile, merge duplicate detections where tiles overlap, and convert pixel positions back to map coordinates for GIS tools. Libraries such as Rasterio read and write the georeferenced rasters, and image registration aligns images from different dates before they are compared.

Satellite vs. Aerial Image Analysis#

Both use the same models and tasks, and the difference lies in the platform. Satellites orbit hundreds of kilometers up and image the whole planet on a fixed schedule, with archives that go back decades. Aircraft and drones fly on demand over small areas and capture centimeter-level detail, which suits inspections of roofs, crops, or wind turbine blades. Overhead object detection datasets often mix both sources, and DOTA includes satellite and aerial images.

Applications#

• Agriculture: Crop type mapping, yield estimation, and NDVI-based monitoring of crop stress across entire regions.
• Forests and climate: Landsat-based global maps of tree cover loss at 30 m track deforestation year by year.
• Disaster response: SAR flood maps and before-and-after damage assessments guide relief efforts, and the International Charter Space and Major Disasters provides satellite data free of charge to responders.
• Urban planning: Building footprint extraction and road mapping keep maps current in fast-growing cities.
• Maritime monitoring: Combining SAR and optical detections with vessel tracking data, a 2024 study in Nature found that about three quarters of the world's industrial fishing vessels are not publicly tracked.