Accurately identifying the spatial distribution and sources of heavy metals is fundamental for effectively controlling agricultural pollution. This study focused on representative polluted farmland areas on the Loess Plateau. For 109 surface-soil samples, we uesd the comprehensive pollution index method, a geographic detector model, the CatBoost machine learning model to systematically analyse the spatial accumulation characteristics of heavy metals and the driving mechanisms of associated pollution in soil. The results indicated that Cd pollution was severe in the study area, with an exceedance rate of 77.98%, and exhibited an increasing trend from the central region towards the southern and northern areas. As and Cu showed mild pollution, with exceedance rates of 3.67% and 2.75%, respectively. Geographic detector analysis identified gross domestic product, population density, surface temperature and, net primary productivity as key driving factors, with factor interactions demonstrating bivariate enhancement effects. Through the integration of the correlation analysis of heavy metals with cross-validation using enterprise spatial distribution patterns, it was further confirmed that As, Cu, Hg, and Zn primarily originated from non-ferrous metal smelting and mineral mining activities. Meanwhile, Cd exhibited a composite source, it was derived from industrial emissions and agricultural chemical use. Meanwhile, high-Pb content areas were concentrated around smelting enterprises and along major transportation routes. The Cat-Boost model, which was constructed based on factors with high explanatory power, accurately predicted the spatial accumulation of heavy metals. The results showed that all six heavy metals showed a varying degree of accumulation. By integrating multi-source data and the integrated methodological approach, this study effectively improves the accuracy of heavy-metal-source analysis and spatial prediction at a small regional scale, providing a scientific basis for the classification, control and, source-oriented prevention of heavy-metal pollution in farmland soils.