Abstract: The range direction inhomogeneous evaporation duct is an abnormal atmospheric structure, which has a high probability of occurrence at sea and has a strong electromagnetic capture ability for low-altitude radar at sea. However, due to the inhomogeneous variation of the range direction profile parameters in the inversion process of the modified refractive index profile of the low-altitude evaporation duct at sea, there is a large inversion complexity and error in the actual marine environment. To solve the above challenges, in this paper, a one-dimensional residual dilated causal convolutional autoencoder network (1D-RDCAE) is proposed to realize low-dimensional non-uniform evaporation duct profile modeling. On this basis, a multi-scale convolutional attention residual network framework (MSCA-ResNet) is constructed to realize horizontal inhomogeneous evaporation duct profile inversion. To verify the effectiveness of the model, we first verify the effectiveness of the dimensionality reduction model on the simulated sea clutter power data set. The experimental results demonstrate that the causal convolutional autoencoder based on one-dimensional residual expansion is closer to the original data after dimension reduction and reconstruction than principal component analysis (PCA), stack autoencoder and one-dimensional convolutional autoencoder, and the convergence speed is faster in the model training process. Secondly, to verify the effectiveness of the inversion model, we tested on the simulated sea clutter and the measured sea clutter data. Using the multi-scale convolution residual network inversion method proposed in this paper, based on the simulated sea clutter and the measured sea clutter data, the evaporation duct height inversion accuracy is 96.98% and 91.25%, respectively, which is better than the current typical inversion method. The deep learning inversion method of inhomogeneous evaporation duct based on sea clutter has the characteristics of high model inversion efficiency, low model complexity and small inversion error, which provides a new technology for real-time high-precision cognition of marine anomalous propagation environment.