Aerial vs Satellite vs Drone Imagery: Choosing Resolution and Capture for Vision Models

Start from target size, not from platform

The right capture choice falls out of one ratio: target size divided by GSD, which tells you how many pixels the object will span. GSD is the ground distance covered by one pixel edge, so a 0.3 m GSD image renders a 3 m vehicle as about 10 pixels long, which is why public 0.3 m benchmarks such as xView label vehicles and small structures. The same image renders a 2.5 cm hail strike as a fraction of one pixel. No labeling budget or model architecture can recover detail the sensor never recorded.

A working heuristic (a planning hypothesis, not a published standard): detection needs the target to span at least 5 to 10 pixels on its shortest side, classification of condition needs 15 to 30, and segmentation of fine boundaries needs more. Texture-based classes such as granule loss, corrosion or cracking need the texture itself to be resolved, which usually pushes requirements down by an order of magnitude compared with the object that carries it. Confirm the ratio on a pilot crop before you buy at scale.

The roof-condition literature shows the failure mode directly. Researchers building a residential roof assessment system noted that minor damage such as hail had rarely been addressed because earlier aerial datasets lacked the detail to show it [1]. That is a data-specification error, not a modeling error, and it is a common reason overhead-imagery projects stall.

What each platform actually delivers

Each platform trades coverage, resolution, revisit and viewing geometry differently, so compare them on all four rather than on resolution alone. The table below summarizes typical ranges buyers encounter; confirm exact figures against the specific sensor and product level you are offered.

Two consequences follow. Open satellite data is the cheapest route to change detection over large areas, but at 10 m a single residential roof is a handful of pixels. Drone imagery resolves components such as insulators, crossarms and connectors on distribution poles [3], but each flight covers a small area and captures vary with pilot, altitude and payload.

Satellite imagery: when wide area and time series matter more than detail

Satellite imagery is the right choice when the label is a footprint, a land-cover class or a change between dates rather than a small defect. Building damage datasets such as xBD pair pre- and post-event satellite images with building polygons and ordinal damage grades across many disaster events [4], which works because "destroyed" versus "no damage" is visible at building scale. Partial damage classes are where satellite labels get noisy, and buyers should inspect confusion between adjacent grades before trusting them.

Watch for three satellite-specific traps. Off-nadir angle changes apparent footprint and hides façades, so record collection elevation and azimuth per scene. Pansharpened products can show texture that is partly an artifact of the fusion algorithm. Product level matters: a top-of-atmosphere Level-1C tile and a surface-reflectance Level-2A tile are not interchangeable for training, so fix one and keep it consistent.

Licensing also differs sharply by source. Copernicus Sentinel data comes with free, full and open access and an attribution requirement [5], while commercial very-high-resolution imagery arrives under EULAs that can restrict derived products and model use. Our satellite imagery licensing guide covers those terms in detail.

Crewed aerial and close-range capture: the middle band

Crewed aerial surveys sit between satellite and drone, giving centimeter-class imagery over whole metro areas with consistent sensors and flight plans. Oblique captures from four directions show walls and roof slopes that nadir satellite imagery hides. This is why property, insurance and municipal programs often rely on aerial survey libraries refreshed on a fixed cycle.

Close-range aerial capture after an event closes the detail gap. A 2019 study used close-range aerial images to estimate the damaged share of roofs after a hurricane [2], a task that requires seeing missing shingles and exposed decking rather than just the roof outline. For the roof-specific version of this choice, including hail and wind classes, see roof condition and hail damage imagery.

The weak points are timing and consistency. Survey flights follow a schedule, so post-event imagery may arrive days or weeks after a storm, mixing damage with temporary repairs such as tarps. Sun angle, season and leaf-on versus leaf-off conditions shift between survey years, which creates distribution shift in any multi-year training set.

Drones: component-level detail with operational variance

Drones are the only practical way to resolve small hardware, surface texture and underside views on a single asset. Utility inspection programs fly drones close to poles and lines because component condition, not pole location, is the label [3]. The same logic applies to roofs, towers, bridges, solar arrays and construction sites, as our drone utility inspection imagery page explains.

Drone data has its own failure modes. Altitude and camera vary by flight, so GSD can differ by a factor of three or more across a corpus unless you normalize or stratify by it. Inspection imagery is oblique and cluttered with sky, so object scale and background are unlike nadir imagery and pretrained overhead backbones transfer poorly. In the US, commercial operations generally fall under FAA Part 107, and operator logs are a useful provenance record.

Metadata is the other asset. EXIF and XMP fields (GPS position, relative altitude, gimbal pitch, focal length, sensor size) let you compute per-image GSD and filter by geometry. Keep them through ingestion and strip only what your privacy review requires, as discussed in EXIF metadata in image training data.

Worked GSD calculation and request specification

The quickest way to settle the platform question is to compute required GSD from the target and write it into the request. For a drone, GSD in cm per pixel equals (sensor width in mm times flight height in m times 100) divided by (focal length in mm times image width in pixels). A 13.2 mm sensor, 8.8 mm lens and 5,472-pixel width at 30 m height gives roughly 0.82 cm per pixel.

Illustrative example: invented to show structure; it does not describe an available dataset.

Then turn the table into a request.

Illustrative example: invented to show structure; it does not describe an available dataset.

target_classes: [hail_bruise, granule_loss, wind_crease, no_damage]
max_gsd_cm: 0.5
view: [nadir, oblique_45]
platform_acceptable: [drone, close_range_aerial]
capture_window_days_after_event: 0-14
per_image_metadata: [gps, relative_altitude_m, gimbal_pitch, focal_length_mm, sensor_model, capture_utc]
labels: polygon per damage instance plus roof-facet ID
ground_truth: adjuster or inspector finding linked by claim or work order
exclusions: [tarped_roofs, post_repair_captures]

Our guide to writing a data request for suppliers shows how to expand this into a complete brief, and image resolution and compression requirements covers file format, bit depth and JPEG quality limits that can erase the detail you paid for.

Ground truth and revisit decide whether imagery is trainable

Resolution makes a target visible, but labels and timing make it learnable. For damage and condition models, the strongest labels come from operational records captured near the imagery date, such as inspection findings, adjuster notes or work orders, joined to images by asset ID, parcel or claim number. Imagery without a dated ground-truth join forces you to label from pixels alone, which inherits annotator guesswork on ambiguous classes.

Revisit frequency sets what temporal questions you can ask. A 5-day Sentinel-2 cadence supports seasonal and change-detection models at field scale. Annual aerial programs support year-over-year roof aging but not storm attribution, and on-demand drone flights support event-tied labels on specific assets. Write the required interval between event and capture into the specification.

Buyers sourcing labeled, close-range imagery tied to operational records can describe the target, GSD and label source on the SourceX buyer page. SourceX sources operational datasets from US companies on request; data is not held in stock, and a request does not guarantee a match. For broader geospatial needs, see geospatial and location data and the image data hub.

Decision checklist before you buy overhead imagery

A short checklist prevents the most expensive mistake, which is buying imagery that cannot show the target.

• Compute target size divided by GSD and confirm the pixel count on sample crops.
• Fix product level, view angle range and spectral bands; record them per image.
• Stratify or normalize GSD across drone flights and survey years.
• Require a dated ground-truth join (inspection, claim or work order ID).
• Check event-to-capture interval and exclude repaired or tarped assets.
• Read the license for derived-model and redistribution limits [5].
• Verify that compression has not destroyed fine texture.

For sample-size planning once the platform is fixed, see how many images you need, and use the vendor evaluation scorecard to compare suppliers. The AI data hub links the rest of the buyer guides.

Sourcing close-range overhead imagery for your vision model

SourceX sources operational datasets from US companies on request, and every dataset is rights-reviewed and delivered under a license that defines records, uses, term and delivery. Personal details are removed or replaced before delivery, and each release is approved by the supplying company. Describe the target, resolution and labels you need at sourcex.si/buyers.