Retrieval of aerosol extinction coefficients for spaceborne high-spectral-resolution lidar based on neural networks

Abstract

The high-spectral-resolution lidar (HSRL) system enables direct retrieval of aerosol extinction coefficients without assuming a lidar ratio. However, traditional retrieval methods depend on the range differentiation of atmospheric optical thickness (AOT), which imposes high demands on the signal-to-noise ratio (SNR). Based on data acquired by the Aerosol and Carbon Dioxide Detection Lidar (ACDL) onboard the atmospheric environment monitoring satellite (DQ-1), this study proposes an end-to-end retrieval method using a convolutional neural network (CNN). By establishing a direct mapping between lidar signals and aerosol extinction coefficients, the proposed method significantly improves the stability of the retrieval results. Furthermore, to construct a high-quality aerosol extinction coefficient dataset, a screening algorithm constrained by reference lidar ratio values was designed. Experimental results demonstrate that in high-SNR regions, the CNN method achieves good agreement with the traditional retrieval method (relative error < 20%); in low-SNR regions, the CNN method still yields continuous and stable aerosol extinction coefficient profiles, whereas the traditional retrieval method exhibits significant data distortion and loss. The retrieval results were validated using three comparative approaches: comparison with reconstructed signals from the ACDL molecular channel; validation against aerosol extinction coefficients and lidar ratio products from the Micro-Pulse Lidar Network (MPLNET); and evaluation using aerosol optical depth (AOD) products from the Aerosol Robotic Network (AERONET). The results indicate that the CNN-based retrieval method demonstrates significantly improved noise resistance and reliability compared to the traditional retrieval method. This approach provides, to our knowledge, a novel technical pathway for high-precision aerosol data retrieval using HSRL.