Synthetic Aperture Radar (SAR) image change detection has the capability to continuously acquire land surface information under all weather and time conditions, making it widely applicable across various fields. However, detecting small area changes using SAR images remains a challenging task, primarily due to the impact of speckle noise and the imbalance between change and no-change classes. To address these challenges, this paper proposes an improved change detection method based on coarse and fine classification. Firstly, a multi-scale superpixel reconstruction method (MSRDI) is employed to generate a Difference Image (DI). This method enhances image edges by utilizing local spatial information within superpixels across multiple scales. Secondly, a weighted two-stage center-constrained fuzzy C-means clustering algorithm (WTCCFCM) is introduced to prevent incorrect class migration. This algorithm uses parallel clustering to classify pixels into change, no-change, and intermediate classes. Finally, the principal component analysis network (PCANet) is used to train and finely classify the pseudo-label samples of the first two classes. Experimental results demonstrate the effectiveness of the proposed method and its ability to significantly suppress speckle noise.