In response to the difficulty of obtaining paired datasets in existing underwater image enhancement methods，an unsupervised underwater image enhancement method based on data-driven optimization and atmospheric light adjustment is proposed. Firstly，the underwater image degradation module is designed，and the latent components of the original image are extracted by parallel network，and the re-degraded image is generated by using the improved Koschmieder model. Then，a data-driven optimization module is designed to compare the original image with the re-degraded image and extract the perceptual transmission of different regions. It adopts a double-branch gating unit and a channel attention mechanism to selectively adjust the input features and realize cross-channel feature fusion. Finally，an atmospheric light optimization module is used to process the global information in frequency domain to offset the influence of atmospheric light on underwater images. The experimental results show that the evaluation indexes of UCIQE，UIQM，NIQE and CIEDE of the proposed method on LSUI dataset，UIEB dataset and RUIE dataset are improved by an average of 4.4%，2.7%，1.6% and 27.5% respectively compared with the suboptimal solution，and it also has obvious advantages in visual quality.