<p>The current unsupervised domain adaptation (UDA) in image classification faces two main challenges: First, there is a significant distribution discrepancy between the source and target domains, and the shared feature space may distort the inherent structure of the data. Second, the lack of label information in the target domain may cause the model to be biased toward high-frequency classes during prediction. To address these issues, this paper proposes the Latent Manifold Preservation and Target Prediction Balance (LMP-TPB) framework to innovatively tackle the above challenges. We develop a manifold preservation module based on the latent domain, enhancing the local geometric consistency of both source and target data. This module effectively preserves the intrinsic topological structure of the data, thereby improving the robustness of the feature representations. Furthermore, we implemented a target prediction balance strategy using an entropy regularization technique. This strategy actively optimizes the prediction distribution in the target domain, achieving a balance across class outputs and enhancing the recognition of low-frequency class features. Extensive experimental results show that our method achieves competitive results, outperforming several SOTA methods.</p>