Abstract: Objectives: Optical remote sensing imagery spans a wide range of spatial scales, and the geometric structure of deep features evolves continuously with resolution. However, existing discrete modeling samples only a finite set of isolated scales and therefore cannot capture the continuous evolution of feature manifolds across scales. To address this limitation, this study aims to establish a continuous, geometry-aware characterization of multi-scale feature manifold evolution, to quantify how scale transformations reshape the intrinsic structure of shallow and deep representations, and to provide a quantitative basis for scale-robust remote sensing interpretation. Methods: This study proposes a Conditional Flow Matching (CFM)-based method for modeling the geometric evolution of multi-scale feature manifolds. ConvNeXt-Tiny serves as the base model, from which shallow and deep features are extracted separately. Because features at different depths occupy heterogeneous representation spaces, Procrustes analysis aligns them in a unified geometric space and establishes reliable cross-scale correspondence. On this basis, CFM learns a scale-driven continuous velocity field that smoothly transports the feature manifold along the scale axis, enabling the characterization of geometric states at arbitrary intermediate scales. For quantitative evaluation, a five-dimensional metric system is constructed, comprising intrinsic dimension, differential entropy, mutual information, silhouette coefficient, and the inter-class to intra-class distance ratio. Experiments are conducted on the NWPU-RESISC45 dataset at down-sampling rates from 1× to 4×, with classification accuracy as a task-level reference. Results: The experiments reveal three principal patterns in feature manifold evolution. First, dimension collapse occurs in deep features: the intrinsic dimension decreases by 61.3%, whereas classification accuracy drops by only 5.5%, indicating that collapse primarily eliminates redundant dimensions and that the compact semantic subspace remains intact. Second, information preservation is non-monotonic: the mutual information of shallow features peaks at the 1.5× scale, an increase of 11.4%, demonstrating the regularization effect of moderate down-sampling. Third, cluster separability improves: the silhouette coefficient of deep features increases by 150%, and the inter-class to intra-class distance ratio increases by 135%. Ablation experiments confirm that this improvement depends on the synergy of Procrustes alignment and semantic constraints. Conclusions: The proposed method replaces discrete scale sampling with continuous geometric modeling and demonstrates that multi-scale feature evolution is a structured process of redundancy elimination, information regularization, and separability enhancement rather than monotonic degradation. These findings provide a continuous modeling perspective and a quantitative basis for multi-scale feature evolution research in remote sensing. They also offer practical guidance for scale selection, multi-scale feature fusion, and resolution-robust model design.