18–20 Nov 2026
Bodoland University, Kokrajhar, Assam
Asia/Kolkata timezone

Identifying the Shape of an Aerial Object from Its Radar Return

Not scheduled
20m
Bodoland University, Kokrajhar, Assam

Bodoland University, Kokrajhar, Assam

Debargaon, P.O. - Rangalikhata, J.D. Road, Kokrajhar(BTR), Assam, (India) Pin - 783370
Oral Presentation Track III: Computational Physics, Machine Learning & AI Applications in Physics

Speaker

Upashana Sharma (Sikkim University)

Description

Identifying the shape of an aerial object from its radar return is
a central problem in target recognition, traditionally addressed
through expert interpretation of Radar Cross Section (RCS)
signatures. This work investigates whether a convolutional neural
network (CNN) can identify an object's cross-sectional profile
from its radar scattering characteristics. A two-dimensional
scattering simulation was developed for 17 boundary profiles,
including flat, tilted, V, W, M, parabolic, circular, sinusoidal,
zigzag, stepped, catenary, and asymmetric notch geometries.
For each of 44 wavefront columns, the local surface normal is
computed numerically, while the Fresnel equations and Snell's
law determine the reflected and transmitted amplitudes and
directions. Radar aspect angles ranging from 5° to
85°and wave phases are randomised. The resulting
scattering point clouds are converted into 64 by 64 echo
images using bilinear interpolation. A four-block CNN containing
approximately 490,000 parameters was trained on 1,360 simulated
images. The model correctly classified all 340 held-out validation
images, achieving 100% validation accuracy and a validation
loss of 0.0064. The signatures remained distinguishable across
all tested aspect angles. These results demonstrate the potential
of combining physics-based wave simulation with deep learning
for automatic radar target recognition, with possible applications
in air surveillance and drone detection.

Author

Upashana Sharma (Sikkim University)

Co-author

Mr Rupak Mukherjee (Sikkim University)

Presentation materials

There are no materials yet.