Основные понятия
A framework for constructing lightweight, physics-based neural surrogates to model atmospheric transmission and enable accurate atmospheric correction of hyperspectral imagery.
Аннотация
The content presents a novel framework called DINSAT (Data-Driven Invertible Neural Surrogates of Atmospheric Transmission) for performing atmospheric correction on hyperspectral imagery. The key highlights are:
DINSAT leverages the concept of Neural Ordinary Differential Equations (Neural ODEs) to construct tunable, physics-based models of atmospheric transmission. This allows the framework to be both invertible and physically consistent.
The authors demonstrate DINSAT in two settings: supervised learning with known ground-truth reflectance data, and unsupervised learning using only at-sensor radiance measurements. In both cases, DINSAT is able to accurately estimate surface reflectance and reconstruct at-sensor radiance.
Compared to standard atmospheric correction methods like QUAC and FLAASH, DINSAT requires less metadata (e.g. lighting, geometry, temperature) and can generalize better to out-of-distribution data.
The framework is designed to be extensible, allowing for the inclusion of more complex physical effects like adjacency, stochastic processes, and time-varying atmospheric profiles.
Overall, DINSAT provides a flexible and powerful approach for performing atmospheric correction on hyperspectral imagery, with potential applications in remote sensing, Earth observation, and other domains involving radiative transfer modeling.
Статистика
The content does not provide specific numerical data or metrics, but rather focuses on the high-level methodology and demonstration of the DINSAT framework.
Цитаты
"DINSAT is a modeling framework for inferring atmospheric transmission profiles suitable for atmospheric correction and transmission modeling tasks."
"We seek to construct minimally-parameterized models for atmospheric correction that are (i) invertible, (ii) physically consistent, (iii) robust to out-of-distribution data, and (iv) have minimal data (at-sensor radiance) and meta-data (lighting, geometry, temperature, etc.) requirements for training."
"The ODESolve(·) can be made differentiable through backpropagating through the elementary operations of the solver (e.g. differentiable programming) or via an adjoint state method [1]."