CROSS-LOCATION NEURAL NETWORK TRANSFER LEARNING FOR RAIN ATTENUATION ESTIMATION AT MILLIMETER-WAVELENGTHS

K. Karachristos1,*, M. Biscarini¹, L. Luini², and D. Comite¹

¹Dipartimento di Ing. Informazioni Elettronica, Sapienza University Rome, Italy, ²Dipartimento di Ing. Elettronica, Politecnico di Milano, Italy, konstantinos.karachristos@uniroma1.it, marianna.biscarini@uniroma1.it, lorenzo.luini@polimi.it, davide.comite@uniroma1.it, *Presenting Autor

Accurate atmospheric attenuation estimation during precipitation events represents a critical challenge for millimeter-wave satellite communications. Classical radiometers utilize the radiative transfer equation to estimate attenuation by combining measured antenna noise temperatures with estimated mean radiating temperature (MRT). This approach is reliable in clear-air and cloudy conditions, where MRT can be accurately estimated from surface meteorological data. However, during precipitation, when scattering effects become significant, especially at millimeter-wavelength, the estimation from surface meteorological data is no longer accurate. Sun Tracking (ST) radiometers circumvent the MRT estimation issue by using the Sun as a stable reference and exploiting the difference between toward-Sun and off-Sun measurements, thus enabling accurate attenuation retrievals in all weather conditions [1]. However, ST installations remain geographically limited and ST measurements are limited at daylight time and at pointing directions determined by the Sun ecliptic. Recently, machine learning approaches have emerged to train neural networks using ST-derived MRT values, enabling classical radiometers to operate during precipitation with ML-predicted MRT instead of unreliable surface-based estimates [2]. However, existing implementations are limited by location-specific training data requirements.

We propose a novel cross-location transfer learning framework enabling artificial neural networks trained on precipitation data from Rome, NY to accurately estimate attenuation during rain events at Milan, Italy. The Rome, NY dataset was collected at the Air Force Research Laboratory using RPG LPW-U72-82 radiometer with ST capability operating at 23.8, 31.4, 72.5, and 82.5 GHz frequencies [1], while Milan data originated from a ST radiometric measurement campaign using in Milan (same radiometer model) [2]. The approach evaluates both direct cross-location transfer and domain adaptation techniques that enhance model generalization by adjusting training procedures before cross-location validation.

Results demonstrate successful cross-location transfer learning during precipitation events. Direct transfer achieves root mean square error (RMSE) values ranging between 0.206 dB (23.84 GHz) and 1.093 dB (82.50 GHz), while domain adaptation further improves performance to 0.929 dB at 82.50 GHz, significantly outperforming classical radiometer approaches (5.4 dB RMSE). These outcomes support the combination of ground-based radiometry and machine learning to experimentally evaluate the atmospheric effects on millimeter-wave Earth-space links without the need of space-borne signals.

References

1. F. S. Marzano, V. Mattioli, L. Milani, K. M. Magde and G. A. Brost, “Sun-Tracking Microwave Radiometry: All-Weather Estimation of Atmospheric Path Attenuation at Ka -, V -, and W -Band,” in IEEE Transactions on Antennas and Propagation, vol. 64, no. 11, pp. 4815-4827, Nov. 2016, doi: 10.1109/TAP.2016.2606568.

2. T. Tunçkol, M. Biscarini and L. Luini, “Radiometric Estimation of Tropospheric Attenuation: A Mixed Physically Based/Machine Learning Approach,” in IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-13, 2024, Art no. 5301113, doi: 10.1109/TGRS.2024.3393506