FFT-Based Multiscale Frequency Decomposition for Atmospheric Lidar Attenuated Backscatter Profile Forecasting

Abstract

Atmospheric light detection and ranging (lidar) measurements of the attenuated backscatter coefficient (ABSC) form high-dimensional vertical profiles that vary across multiple temporal scales. Directly processing the original time-domain sequence with a single prediction structure may entangle information associated with these different scales. To address this issue, we propose FFT-FDNet, a frequency-domain decomposition network based on fast Fourier transform (FFT). FFT-FDNet explicitly decomposes the input sequence along the temporal dimension into low-, mid-, and high-frequency components. Branch-specific predictors model these components separately, and a learnable fusion layer combines their outputs to forecast future ABSC profiles. Experiments were conducted using continuous single-site lidar observations from the Tokyo station at a temporal resolution of 15 min. FFT-FDNet was compared with eight representative time-series forecasting models at forecast horizons of 2, 4, and 6 h. Across the three horizons, FFT-FDNet achieved the lowest mean MAEstd and MSEstd and the highest mean R2 among the evaluated methods. At the 2 h horizon, these metrics were 0.1221, 0.4116, and 0.7505, respectively. The ablation results showed consistent performance degradation after removing the FFT decomposition, low-frequency branch, or mid-frequency branch, whereas the high-frequency branch provided modest improvements at some forecast horizons. The frequency band sensitivity analysis supported the use of (ν1,ν2)=(0.10,0.45) among the evaluated cutoff combinations. These results suggest that FFT-based three-band decomposition is useful for the present Tokyo single-station short-term forecasting case. Further validation using data from more stations, seasons, and aerosol conditions is still needed.