Estimation of Doppler velocity from incoherent scatter spectra using context-aware transformers

Abstract. We present a context-aware transformer model for estimating Doppler velocity from incoherent scatter radar (ISR) spectra. The model is based on the standard transformer encoder with adaptations from the Vision Transformer. Trained entirely on theoretical spectra, the AI model generalizes well for Arecibo ISR data and outperforms the traditional fitting methods significantly. To assess performance, we compare the plasma drift velocities estimated by the AI model with those from the conventional least-squares fitting (LSF) technique. The analysis is focused near 110 km altitude, where pronounced vertical velocity and electron density gradients provide an environment that most readily highlights the difference between the two estimation methods. Our results show that the LSF velocity error is 1.5 to 3.5 times that of the AI model using 5 input heights. An inference from the AI model is approximately 100 times faster than the LSF method and requires minimal hardware, making it practical for large-scale or real-time processing. Although the AI method is demonstrated for the Arecibo ISR, it can be potentially applied to other cases where the spectra can be parameterized.