Session
Advanced Technologies 1
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
Hyperspectral small satellites now collect far more data than they can return to the ground. A single acquisition spans tens to hundreds of spectral bands, yet the downlink window on one pass carries only a fraction of it, so operators are forced to drop scenes, bin bands coarsely, or store data that never reaches the ground. Much of this waste is avoidable. When the task is known before the spacecraft images, whether to map water, monitor vegetation, or pick out a mineral signature, only a handful of bands actually carry the answer; the rest is paid for in downlink and then discarded. The natural lever is to choose which bands to keep from the task itself, and to make that choice onboard. This is becoming practical just as the hardware allows it. A new class of reprogrammable imagers, the Simera Sense HyperScape100 among them, can be told which bands to record and reconfigured in orbit. Band selection no longer has to wait until after capture; it can move ahead of it, so the data a mission does not need is never collected in the first place. What these instruments still lack is the software that decides, for a given task, which bands to ask for. DynaBand is that missing layer: a small, context-aware selector that reads the operator's task and returns the bands that matter for it, from one model that serves every task without retraining and runs fast enough for a spacecraft. Task-driven, onboard band selection is the efficient way to spend a satellite's scarce downlink and compute, and a single conditioned model, rather than a fixed subset or a heavy per-task network, is the practical means to deliver it. This paper presents the design of DynaBand, the failure modes encountered in developing it, and a path toward running the selector ahead of capture on a reprogrammable imager.
Document Type
Event
DynaBand: Context-Aware Dynamic Band Selection for Onboard Hyperspectral Processing in Small Satellites
Salt Palace Convention Center, Salt Lake City, UT
Hyperspectral small satellites now collect far more data than they can return to the ground. A single acquisition spans tens to hundreds of spectral bands, yet the downlink window on one pass carries only a fraction of it, so operators are forced to drop scenes, bin bands coarsely, or store data that never reaches the ground. Much of this waste is avoidable. When the task is known before the spacecraft images, whether to map water, monitor vegetation, or pick out a mineral signature, only a handful of bands actually carry the answer; the rest is paid for in downlink and then discarded. The natural lever is to choose which bands to keep from the task itself, and to make that choice onboard. This is becoming practical just as the hardware allows it. A new class of reprogrammable imagers, the Simera Sense HyperScape100 among them, can be told which bands to record and reconfigured in orbit. Band selection no longer has to wait until after capture; it can move ahead of it, so the data a mission does not need is never collected in the first place. What these instruments still lack is the software that decides, for a given task, which bands to ask for. DynaBand is that missing layer: a small, context-aware selector that reads the operator's task and returns the bands that matter for it, from one model that serves every task without retraining and runs fast enough for a spacecraft. Task-driven, onboard band selection is the efficient way to spend a satellite's scarce downlink and compute, and a single conditioned model, rather than a fixed subset or a heavy per-task network, is the practical means to deliver it. This paper presents the design of DynaBand, the failure modes encountered in developing it, and a path toward running the selector ahead of capture on a reprogrammable imager.
