Dataset for \HDF-Net: A Hybrid Deep Learning Model for Retrieving Ozone Profiles in the Arctic\

The continuous, high-temporal-resolution vertical distribution of ozone (O\u2083) in the Arctic is crucial for understanding the atmospheric chemistry and dynamics of this climate-sensitive region. However, obtaining such detailed observations remains a significant challenge. This study introduces a Hybrid Deep-learning Fusion Network (HDF-Net) to retrieve high-resolution ozone profiles by synergistically fusing ground-based Multi-Axis Differential Optical Absorption Spectroscopy (MAX-DOAS) observations with meteorological reanalysis data. The HDF-Net employs a multi-branch architecture where a Convolutional Neural Network (CNN) extracts spatial features from ERA5 meteorological profiles, while parallel Bidirectional Long Short-Term Memory (BiLSTM) networks capture the sequential information from both MAX-DOAS slant column densities (SCDs) and differential slant column densities (DSCDs). These deep features are fused with solar geometry information to jointly retrieve a 25-layer ozone profile with minute-level temporal resolution, extending from the near-surface to approximately 40 km. The HDF-Net 's performance was validated using multiple independent datasets. Comparison with ozonesonde measurements reveals excellent agreement, with the correlation (r) peaking at 0.92 at 150 hPa. Crucially, retrieved near-surface ozone shows a stronger correlation with co-located in-situ measurements (r = 0.58) and a lower RMSE (6.43 ppbv) than EAC4 data. Furthermore, HDF-Net successfully captures transient dynamic events missed by the 3-hourly reanalysis product. These results establish that HDF-Net can reliably retrieve ozone profiles, providing a powerful new tool for investigating the spatiotemporal variability of ozone in the data-sparse Arctic region.

北极地区臭氧（O₃）的连续高时间分辨率垂直分布，对于理解这一气候敏感区域的大气化学与动力学机制至关重要。然而，获取此类精细化观测仍面临重大挑战。本研究提出一种混合深度学习融合网络（Hybrid Deep-learning Fusion Network，HDF-Net），通过协同融合地基多轴差分光学吸收光谱（Multi-Axis Differential Optical Absorption Spectroscopy，MAX-DOAS）观测数据与气象再分析资料，反演高分辨率臭氧廓线。HDF-Net采用多分支架构：卷积神经网络（Convolutional Neural Network，CNN）从ERA5气象廓线中提取空间特征，并行双向长短期记忆（Bidirectional Long Short-Term Memory，BiLSTM）网络则分别从MAX-DOAS斜柱浓度（SCDs）与差分斜柱浓度（DSCDs）中捕捉序列信息。将上述深度学习特征与太阳几何信息融合后，可共同反演得到从近地面延伸至约40 km高度、共计25层的分钟级时间分辨率臭氧廓线。研究通过多套独立数据集验证了HDF-Net的性能：与臭氧探空仪观测结果对比显示二者吻合度极佳，在150 hPa处相关系数（r）峰值可达0.92。尤为关键的是，反演得到的近地面臭氧与同位置原位观测数据相关性更强（r=0.58），且均方根误差（RMSE）为6.43 ppbv，表现优于EAC4再分析数据。此外，HDF-Net成功捕捉到了3小时间隔再分析产品遗漏的瞬态动力事件。上述结果表明，HDF-Net可可靠反演臭氧廓线，为探究数据稀缺的北极地区臭氧时空变异性提供了强有力的新工具。