Natural gas leak detection and quantitative assessment at urban gas stations are crucial for accident prevention and environmental protection. This study developed a novel method for the quantitative inversion of natural gas leaks, integrating distributed polling sampling with deep learning. We employed a Distributed Gas Polling Sampling (DGPS) monitoring system to acquire high spatiotemporal resolution data. Subsequently, we constructed a dual-branch deep learning model. This model achieved precise quantitative inversion of natural gas leaks across multiple flow rate scales, with a particular enhancement in identifying minute leaks, by adaptively fusing predictions from a main branch and a dedicated small-flow-rate branch for weak signals. Test results indicated that the model achieved a coefficient of determination (R2) of 0.9157 and an average relative error of 12.18% for overall predictions. For small flow rates (<0.1 m3/h), the relative error was 17.51%. During a 24-hour continuous application at an operational natural gas station, the method demonstrated a relative error of 17.86% compared to measurements obtained by the traditional bagging method. This research provides a field-validated, high-precision online monitoring solution for the online monitoring and accurate quantification of methane emissions from natural gas facilities.