Abstract: Under the "dual carbon" goals, efficient and accurate estimation of terrestrial ecosystem carbon storage is of great significance. To effectively use multi-source remote sensing data to obtain the spatial distribution of carbon storage, this study proposes a carbon storage estimation method based on an improved U-Net model, integrating GF-1 and Sentinel-1 remote sensing features for land use classification and subsequently estimating carbon storage by combining classification results with carbon density survey data. The architecture employs dual encoders to individually extract multi-scale features from GF-1 and Sentinel-1 imagery, adopts a segment-wise feature fusion strategy, and leverages attention mechanisms with skip connections to focus on high-level semantic features. Finally, a classifier outputs land use categories, and carbon storage distribution is obtained by multiplying the classification results with carbon density values derived from field surveys. The results show that the improved U-Net model fully extracts land cover feature details and semantic features, achieving an overall classification accuracy of 87.34%; the model-estimated carbon storage is 31.28 Tg, with an overall error of 0.84% compared to measured carbon storage. In the classification results, the area proportions of forest, cultivated land, garden land, residential land, water bodies, roads, and other unused land are 83.30%, 4.42%, 6.47%, 2.72%, 1.02%, 1.10%, and 0.97%, respectively. Natural forests are the main carbon storage reservoir, accounting for 91.51% of the total carbon storage; the contribution rates of cultivated land, garden land, residential land, water bodies, and other unused land are 2.76%, 5.55%, 0.03%, 0.06%, and 0.09%, respectively. Compared with the random forest (RF) and linear regression (LR) models, the proposed model reduces the carbon storage estimation error by 1.91 and 2.94 percentage points, respectively. These results demonstrate that optical images and synthetic aperture radar (SAR) images possess good complementarity, and that the U-Net model combined with the channel-spatial attention module (CSAM) realizes the fusion of shallow and deep features, thereby improving land cover classification accuracy. The model uses widely accessible, low-cost data and has good reliability, providing a reference for sustained and efficient dynamic monitoring of carbon storage.