A new video-compression technique squeezes hyperspectral footage down by nearly 90 percent without losing the detail that object-tracking algorithms need.
Researchers adapted an existing RGB video compression model built on implicit neural representation, a technique that encodes video as a trained neural network rather than a grid of pixels, so it works on hyperspectral footage, which captures dozens of wavelength bands instead of just three color channels. Tested against traditional hyperspectral compression methods applied frame-by-frame, the new approach delivered a Bjontegaard Delta PSNR gain of +4.99 dB and cut bitrate by 88.88 percent at equivalent quality. The team also checked whether that compression helped or hurt a real downstream task: tracking objects across frames using the HOT2026 hyperspectral tracking dataset. In low-data conditions, footage compressed with the new method beat video compressed with older PCA-based and JPEG2000 approaches, improving tracking accuracy (AUC) by up to 23.42 percent and distance precision by up to 35.56 percent.
Hyperspectral cameras, which record far more of the light spectrum than ordinary color sensors, are increasingly used in agriculture, defense, and remote sensing, but the resulting files are enormous and existing compression tools were mostly borrowed from single-image or RGB-video pipelines. A method that compresses harder while actually improving downstream analysis, rather than just tolerating the quality loss, addresses a real bottleneck for anyone trying to store, transmit, or analyze this data at scale.
This is one paper on one dataset, and implicit neural representation methods are notoriously slow to encode even when they decode fast, so the headline compression numbers say nothing yet about whether this could run on a drone or satellite in real time.