Systems and methods for differentiable programming for hyperspectral unmixing

Patent metadata

Field	Value
Publication number	US-12379249-B2
Application number	US-202117353077-A
Country	US
Kind code	B2
Filing date	Jun 21, 2021
Priority date	Jun 19, 2020
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

Various embodiments of a system for linear unmixing of spectral images using a dispersion model are disclosed herein.

First claim

What is claimed is:

1. A system, comprising:
a spectral imaging device operable for measuring a measured spectra of one or more materials within a scanned region associated with a mixture; and
a processor in operative communication with the spectral imaging device, the processor including instructions which, when executed, cause the processor to:
receive spectrometer data from the spectral imaging device, the spectrometer data including a measured spectra b including emissivity values for one or more materials within a scanned region;
process the measured spectra through a dispersion model to identify an endmember spectra present in the measured spectra, the endmember spectra representative of a spectrum of one or more materials of within the scanned region, wherein the dispersion model simulates realistic spectral variation to identify individual materials present in the mixture,
wherein
a set of parameters of the dispersion model are optimized by iteratively minimizing a loss between the measured spectra and a fitted endmember spectra with respect to the set of parameters; and

identify an amount for each of the one or more materials to generate a set of amounts by iteratively minimizing a loss between the measured spectra and a fitted endmember spectra with respect to the set of amounts, wherein an amount for each of the one or more materials quantifies a presence of a respective material in the mixture.

2. The system of claim 1 , wherein the fitted endmember spectra is expressed as a matrix product of a dispersion model parameter matrix including the endmember spectra and the set of amounts.

3. The system of claim 1 , wherein the steps of optimizing the set of dispersion model parameters and identifying the set of amounts are repeated until the loss between the measured spectra and the fitted endmember spectra is minimized.

4. The system of claim 1 , wherein including instructions which, when executed, further cause the processor to:
perform gradient descent on a loss function to backpropagate a loss function throughout the dispersion model; and
update each dispersion model parameter of the set of dispersion model parameters based on the result of the loss function with respect to each individual dispersion model parameter.

5. The system of claim 4 , wherein the loss function is representative of a difference between the fitted endmember spectra and the measured spectra.

6. The system of claim 1 , wherein the system alternates between optimizing the set of parameters and identifying the set of amounts in an iterative fashion.

7. The system of claim 1 , wherein the dispersion model identifies a plurality of endmembers present in the scanned region.

8. A method, comprising:
receive spectrometer data from the spectral imaging device, the spectrometer data including a measured spectra b including emissivity values for one or more materials within a scanned region;
process the measured spectra through a dispersion model to identify an endmember spectra present in the measured spectra, the endmember spectra representative of a spectrum of one or more materials of within the scanned region;
optimize a set of parameters of the dispersion model by iteratively minimizing a loss between the measured spectra and a fitted endmember spectra with respect to the set of parameters; and
identify a set of amounts for each of the one or more materials by iteratively minimizing a loss between the measured spectra and a fitted endmember spectra with respect to the set of amounts, wherein an amount for each of the one or more materials quantifies a presence of a respective material in a mixture.

9. The method of claim 8 , wherein the fitted endmember spectra is represented by a matrix product of the set of dispersion model parameters and the set of amounts.

10. The method of claim 8 , wherein the difference between the measured spectra and a fitted endmember spectra is determined using regularized least squares optimization.

11. The method of claim 8 , wherein gradient descent is performed on the loss function with respect to each individual dispersion model parameter prior to updating each dispersion model parameter.

12. The method of claim 8 , wherein the dispersion model identifies a plurality of endmembers present in the scanned region.

13. The method of claim 8 , wherein the dispersion model determines the endmember spectra by analyzing optical properties of the material to determine one or more refractive indexes of the material.

14. The method of claim 13 , wherein the one or more refractive indexes are used to model the endmember spectra in the form of an emissivity curve.

15. The method of claim 13 , wherein the endmember spectra is modeled for each of a plurality of optical axes of symmetry for crystal structures and wherein a different set of dispersion model parameters is determined for each optical axis of symmetry.

16. The method of claim 8 , wherein the steps of optimizing the set of parameters and identifying the set of amounts are performed in an alternating fashion.

Assignees

• Janiczek John
• Jayasuriya Suren
• Dasarathy Gautam
• Edwards Christopher
• Christensen Philip
• Univ Arizona State
• Arizona Board Of Regents For And On Behalf Of Northern Arizona Univ

Inventors

• Janiczek John
• Jayasuriya Suren
• Dasarathy Gautam
• Edwards Christopher
• Christensen Philip

Classifications

• G01J2003/2843

  Processing for eliminating interfering spectra · CPC title

• G01J2003/2826

  Multispectral imaging, e.g. filter imaging · CPC title

• G06V10/82

  using neural networks · CPC title

• G01J3/2823 Primary

  Imaging spectrometer · CPC title