I'm currently in my third year at ENSEEIHT, and I’m doing an internship at the company Magellium as my final-year project. My work focuses on integrating physical knowledge into neural networks.

​

The application domain is atmospheric correction, which involves correcting the impact of the atmosphere on satellite images, which are L1C products (top-of-atmosphere reflectance images). To do this, we typically use Radiative Transfer Models (RTMs), which simulate how light interacts with the atmosphere by solving complex physical equations. These simulations are used to determine how to correct each pixel in the satellite image.

​

However, RTMs are computationally expensive so running them over an entire image, which may contain millions of pixels, is impractical. To address this, the standard approach is to use Look-Up Tables (LUTs). These are precomputed datasets generated from RTM simulations over a range of atmospheric and geometric conditions. When correcting a pixel, we perform interpolation over the LUT to estimate the correction based on the pixel’s properties.

​

While effective, this method sacrifices accuracy and is not well-suited for operational production chains due to its complexity and limited flexibility.

This is where emulators come in. Emulators are surrogate models that mimic the output of RTMs, enabling much faster simulations. The current state-of-the-art for emulation in atmospheric correction relies on Gaussian Process Regression (GPR) which is a machine learning method that provides both predictions and uncertainty estimates.

​

Therefore, the goal of my internship is to develop a new kind of surrogate model based on neural networks that incorporates physical constraints directly into its structure. I aim to compare this physics-informed neural network with GPR models and to assess the uncertainty of the predictions. Ultimately, this research could help justify the integration of neural networks into operational atmospheric correction workflows in the future.

​

My internship is still ongoing, so the results and conclusions will be finalized at the end of the internship !