Learning Set Representations for LWIR In-Scene Atmospheric Compensation

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 13; pp. 1438 - 1449
• Main Authors: Westing, Nicholas, Gross, Kevin C., Borghetti, Brett J., Martin, Jacob, Meola, Joseph
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Atmospheric compensation; Atmospheric measurements; Atmospheric modeling; Atmospheric waves; Blackbody; Cloud computing; Compensation; Computer applications; Computer architecture; dimension reduction; Dimensionality reduction; Emissivity; Graphics; Hypercubes; hyperspectral imagery; Hyperspectral imaging; Image classification; Imagery; Infrared analysis; Infrared imagery; Learning set; Mobile computing; Neural networks; Permutations; Pixels; Radiance; Radiative transfer; Radiosondes; Representations; Sensors; Surface temperature; Temperature measurement; Three dimensional models; Training
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2020.2980750

Atmospheric compensation of long-wave infrared (LWIR) hyperspectral imagery is investigated in this article using set representations learned by a neural network. This approach relies on synthetic at-sensor radiance data derived from collected radiosondes and a diverse database of measured emissivity spectra sampled at a range of surface temperatures. The network loss function relies on LWIR radiative transfer equations to update model parameters. Atmospheric predictions are made on a set of diverse pixels extracted from the scene, without knowledge of blackbody pixels or pixel temperatures. The network architecture utilizes permutation-invariant layers to predict a set representation, similar to the work performed in point cloud classification. When applied to collected hyperspectral image data, this method shows comparable performance to Fast Line-of-Sight Atmospheric Analysis of Hypercubes-Infrared (FLAASH-IR), using an automated pixel selection approach. Additionally, inference time is significantly reduced compared to FLAASH-IR with predictions made on average in 0.24 s on a 128 pixel by 5000 pixel data cube using a mobile graphics card. This computational speed-up on a low-power platform results in an autonomous atmospheric compensation method effective for real-time, onboard use, while only requiring a diversity of materials in the scene.