Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data-Driven Invertible Neural Surrogates of Atmospheric Transmission

James Koch, Brenda Forland, Bruce Bernacki, Timothy Doster, Tegan Emerson

TL;DR

Data-Driven Invertible Neural Surrogates of Atmospheric Transmission (DINSAT) addresses atmospheric correction and cross-domain spectral translation by learning a differentiable, invertible surrogate for atmospheric transmission. The method uses Neural Ordinary Differential Equations: the transmission operator is defined by $\frac{dL}{dx} = f(L; \theta)$ and solved with a differentiable ODESolve to map the upwelling solar radiation $L_0$, surface reflectance $\rho$, and background $C$ to the at-sensor radiance $L_4 = C + T\left(\rho T\left(L_0\right)\right)$. The paper demonstrates both supervised (linear and nonlinear) and unsupervised training on HyMap data, achieving low MSE for surface reflectance in the linear case (e.g., ~1.5%) and revealing trade-offs for nonlinear and unsupervised models. DINSAT's differentiable, physics-informed surrogates enable invertible corrections, modality translation, and extension to multi-media or time-varying transmissions, offering a lightweight alternative to full radiative transfer codes.

Abstract

We present a framework for inferring an atmospheric transmission profile from a spectral scene. This framework leverages a lightweight, physics-based simulator that is automatically tuned - by virtue of autodifferentiation and differentiable programming - to construct a surrogate atmospheric profile to model the observed data. We demonstrate utility of the methodology by (i) performing atmospheric correction, (ii) recasting spectral data between various modalities (e.g. radiance and reflectance at the surface and at the sensor), and (iii) inferring atmospheric transmission profiles, such as absorbing bands and their relative magnitudes.

Data-Driven Invertible Neural Surrogates of Atmospheric Transmission

TL;DR

Data-Driven Invertible Neural Surrogates of Atmospheric Transmission (DINSAT) addresses atmospheric correction and cross-domain spectral translation by learning a differentiable, invertible surrogate for atmospheric transmission. The method uses Neural Ordinary Differential Equations: the transmission operator is defined by and solved with a differentiable ODESolve to map the upwelling solar radiation , surface reflectance , and background to the at-sensor radiance . The paper demonstrates both supervised (linear and nonlinear) and unsupervised training on HyMap data, achieving low MSE for surface reflectance in the linear case (e.g., ~1.5%) and revealing trade-offs for nonlinear and unsupervised models. DINSAT's differentiable, physics-informed surrogates enable invertible corrections, modality translation, and extension to multi-media or time-varying transmissions, offering a lightweight alternative to full radiative transfer codes.

Abstract

We present a framework for inferring an atmospheric transmission profile from a spectral scene. This framework leverages a lightweight, physics-based simulator that is automatically tuned - by virtue of autodifferentiation and differentiable programming - to construct a surrogate atmospheric profile to model the observed data. We demonstrate utility of the methodology by (i) performing atmospheric correction, (ii) recasting spectral data between various modalities (e.g. radiance and reflectance at the surface and at the sensor), and (iii) inferring atmospheric transmission profiles, such as absorbing bands and their relative magnitudes.
Paper Structure (8 sections, 7 equations, 3 figures)

This paper contains 8 sections, 7 equations, 3 figures.

Table of Contents

  1. Introduction
  2. DINSAT Methodology
  3. Neural Ordinary Differential Equations
  4. Atmospheric Correction
  5. Demonstration
  6. Supervised Models
  7. Unsupervised Model
  8. Conclusion

Figures (3)

  • Figure 1: Solar radiation, denoted $L_0$, is both absorbed and scattered in the atmosphere, the extent of which is tightly coupled with atmospheric properties. A sensor records radiation intensity as a combination of lumped upwelling effects ($C$) and direct path radiation ($L_3$).
  • Figure 2: An example in-scene target is shown in (a) (reproduced from Snyder et al. snyder2008development). The laboratory-measured spectral signature for the target is shown in (b) alongside estimates provided by off-the-shelf atmospheric correction routines. In (c), shown are variations of our model-derived corrected spectra compared against the same reference spectra.
  • Figure 3: Summary of DINSAT results. Left: estimated in- and outgoing radiation at a surface-based target. Middle: Estimated reflectance spectra. Right: estimates of at-sensor radiance given a DINSAT model and known (library) spectra.