Mesoscale ocean eddies shape ocean circulation and transport heat, carbon, and momentum, but their structure is often smoothed in conventional satellite altimetry maps. Here, we use DPI(deep prior interpolation) to improve global estimates of sea level anomaly by merging nadir altimeter and SWOT KaRIn observations. Without prescribing covariance structures or using labeled training data, the method learns a high-resolution spatiotemporal field directly from satellite observations through a self-supervised deep prior. A leave-one-out Top-k ensemble further limits noise fitting and improves reconstruction stability. In observing system simulation experiments, our method achieves lower errors and finer effective resolution than DUACS, and substantially improves eddy-detection F1 scores. Experiments with real observations also outperform optimal interpolation and demonstrate the feasibility of producing global 2-km SLA maps. These results suggest that deep prior interpolation can provide a new paradigm for high-resolution satellite oceanography and more reliable observations of mesoscale eddy dynamics.