Random Scattering Modelling and Analysis by Machine Learning

Microwave scattering in an arbitrary scene is inherently complex and highly sensitive. For a given target, its spectral response is primarily influenced by factors such as geometry and material properties. However, even a slight change in the angle between the radar and the target can significantly alter this response. As a result, it becomes challenging to classify targets based on a fixed reference coordinate system. This difficulty arises because microwave scattering responses are governed by a dynamic, unknown coordinate system that is inherently problem- and data-dependent.Therefore, extracting the coordinate system directly from the data is central to scattering analysis. This concept closely aligns with the principles of autoencoder models in machine learning. Rather than analyzing scattering signals based on predefined reference responses (fixed coordinate systems), we adopt a "learning by doing" approach. Specifically, we extract a problem-dependent coordinate system—defined as kernels—directly from radar data, enabling more robust analysis of random scattering behavior.