Label What a Sensor Sees from a Handful of Examples

Hyperspectral sensors record hundreds of wavelengths per pixel, enough to tell apart materials that look identical to a camera. Labeling that data usually takes thousands of hand-marked pixels. Planck's hybrid method works from about twenty per class.

Talk to Planck
Wavelength Reflectance Vegetation Water Submerged seagrass Aircraft Ground

Material and Target Labeling

The decision
What each pixel in a scene is: vegetation, water, a submerged object, an aircraft, bare ground.
What makes it hard
Hundreds of bands per pixel and very few labeled examples. Some classes, such as seagrass under the water surface, are invisible to an ordinary camera.
Benefits
Accurate material and target maps from about 20 labeled examples per class, instead of thousands of hand-marked pixels.
90% classification accuracy on quantum hardware, from about 20 reference pixels per class.

Accuracy on hardware rose from 60% to 90%. Simulator results are reported separately.

Hardware
IBM Brisbane, quantum SVM
Classes
Including submerged seagrass and aircraft on the ground

One Method, Commercial and Defense Uses

Labeling from few examples matters wherever ground truth is expensive or dangerous to collect.

Commercial and civilDefense and security
Coastal and seagrass monitoringDetection of objects below the water surface
Crop health and land use mappingCamouflage and concealment detection
Airport and infrastructure surveysAircraft and vehicle identification

Other Solutions

Defense Space Maritime Energy and grid Logistics and field services

Have Imagery and Too Few Labels?

Planck produces a labeled map from a few reference pixels per class, with the validation behind it.

Talk to Planck
Talk to Planck Book 15 minutes with the founder