Model inversion by parameter fit using NN emulating the forward model: evaluation of indirect measurements

Area of Science:

• Computational Science
• Data Science
• Remote Sensing

Background:

• Inverse models are crucial for deriving parameters from measurements across science and technology.
• Neural networks (NNs) have enabled the operational use of complex inverse models through emulation.
• Improving the accuracy and efficiency of inverse modeling remains a key challenge.

Purpose of the Study:

• To demonstrate how neural networks (NNs) can significantly improve the accuracy of inverse model parameter retrieval.
• To leverage the Jacobian matrix, a byproduct of NN calculations, for faster and more precise inversions.
• To extend the methodology to handle non-diagonal covariance matrices for enhanced parameter estimation.

Main Methods:

• Emulation of inverse models using neural networks (NNs).
• Minimization of the sum of squared errors during the NN training process.
• Utilizing the Jacobian matrix, efficiently computed during NN forward passes, to refine parameter estimates.
• Incorporating non-diagonal measurement covariance matrices into the inversion process.

Main Results:

• NN-based inverse models achieve improved accuracy in parameter retrieval compared to traditional methods.
• The Jacobian-derived optimization significantly accelerates the inversion process.
• The method successfully accounts for complex, non-diagonal measurement covariances, leading to more robust parameter estimations.
• Demonstrated practical application and accuracy improvements in a remote sensing example.

Conclusions:

• Neural networks offer a powerful and efficient approach to enhance inverse model accuracy.
• The Jacobian byproduct of NN computations provides a computationally advantageous route to improved parameter estimation.
• This methodology is versatile, applicable to various fields including remote sensing, and capable of handling sophisticated error structures.