HiPredict
A U-Net-based machine learning framework that reconstructs failed spectral channels of planetary probes using multi-channel imagery and metadata.
Timeline (last 90 days)
The inputs to HiPredict include the IR10 image at the same location, the blue-green channel (BG12) image, nearby red channel observations, and acquisition metadata.
HiPredict adopts a U-Net architecture to learn the nonlinear spatiotemporal-spectral relationship between the inputs and the target channel.
HiPredict training used 81,717 training samples corresponding to 13,625 observations.
HiPredict was evaluated on 23,348 test samples drawn from 3,892 observations before the failure occurred.
The normalized mean absolute error of HiPredict is 1.1%, significantly better than SYN4's 3.1%.
The reconstructed images are closer to the real images on geological features such as ice-bearing craters, strata exposed along ridge tops, and chloride deposits.
The paper releases HiPredict as an open-source software tool.
The HiPredict method relies on the presence of other channels at the same location and nearby observations, and is not applicable to scenarios where redundant information is completely lost.
All Facts (8)
The HiPredict method relies on the presence of other channels at the same location and nearby observations, and is not applicable to scenarios where redundant information is completely lost.
50%UnverifiedHiPredict adopts a U-Net architecture to learn the nonlinear spatiotemporal-spectral relationship between the inputs and the target channel.
50%UnverifiedHiPredict training used 81,717 training samples corresponding to 13,625 observations.
50%UnverifiedHiPredict was evaluated on 23,348 test samples drawn from 3,892 observations before the failure occurred.
50%UnverifiedThe normalized mean absolute error of HiPredict is 1.1%, significantly better than SYN4's 3.1%.
50%UnverifiedThe reconstructed images are closer to the real images on geological features such as ice-bearing craters, strata exposed along ridge tops, and chloride deposits.
50%UnverifiedThe inputs to HiPredict include the IR10 image at the same location, the blue-green channel (BG12) image, nearby red channel observations, and acquisition metadata.
50%UnverifiedThe paper releases HiPredict as an open-source software tool.
50%