A Machine Learning-Assisted Approach for the Reduction of Altitude Estimation Error in 3D Surveillance Radar Systems

• Published in: Integrated Communications, Navigation, and Surveillance Conference pp. 1 - 5
• Main Authors: Felix, Pablo Parodi, Carnero, Enrique Barbasan, Zhuk, Igor
• Format: Conference Proceeding
• Language: English
• Published: IEEE 08.04.2025
• Subjects: Air traffic control; Aircraft; Altitude error; Atmospheric measurements; Atmospheric modeling; Data models; Machine learning; Primary radar surveillance; Radar; Sensitivity analysis; Surveillance; Three-dimensional displays; Transponders
• ISSN: 2155-4951
• DOI: 10.1109/ICNS65417.2025.10976918

Altitude measurements of the aircraft in the airspace typically relies on Mode C or S transponders in Secondary Surveillance Radar systems (SSR). However, these methods present significant limitations in various operational contexts. In particular, there are situations where aircraft are not equipped with transponders, such as small aircraft, drones, or non-cooperative aircraft, including non-civil or intruder aircraft, entering airspace without providing identity information. This poses a critical challenge for Air Traffic Management (ATM), which requires the ability to monitor and track all aircraft. Additionally, secondary radar systems are not without drawbacks, such as the high costs associated with implementation, potential communication failures or denial, and limited coverage in certain regions. Three-dimensional (3D) Surveillance Radar (SR) systems can provide target altitude information. While errors caused by atmospheric conditions are intended to be corrected by Atmospheric Propagation Models, these models are based on educated assumptions and are therefore subject to inaccuracies due to, among other factors, nonlinear effects and the lack of feedback information. In this work, a model based on Machine Learning (ML) techniques is introduced, which aims to reduce the error present in the target's altitude coordinate estimation. This will result in a model which better represents the real propagation of the radar signal through the atmosphere than its traditional counterparts. Under these premises, various ML regression models were developed and trained using a training set of input data that includes both SR plot information and the real barometric altitude from the SSR. The model utilizes over 70 variables related to distinct aspects such as flight dynamics, location or real and imaginary parts of phase monopulse ratio, among others. While this variable set was reduced through a sensitivity analysis for dimensionality reduction, additional inputs were generated by preprocessing modules specifically developed for the problem, resulting in a tailored solution. A set of experiments was conducted, where several ML Ensemble model configurations are tested and compared based on the resulting reduction in altitude RMS error. With the selected ensemble model configuration, additional experiments were carried out to find the optimal number of previous detections used at each inference, considering both performance metrics and practical operative advantages. The result of the work is a model in which altitude RMS of Error Reduction over 70% is achieved for SR plots in some approaches while other approaches worsen the error in some scenarios. The reasons behind it are further analyzed and discussed, and a way ahead is presented.