The　correlation　between　multiple　patterns　of　micromagnetic　signatures　and　the　surface　hardness　of　three　types　of　carbon　steels　(referred　as　Cr12MoV　steel,　45　steel,　and　martensitic　stainless　steel　of　3Cr13　in　Chinese　standards)　was　investigated　in　this　study.　Back-propagation　neural　network　(BP-NN)　models　were　established　for　the　quantitative　prediction　of　surface　hardness　in　carbon　steels.　The　quality　of　the　specimens　used　for　calibration　(or　more　precisely　the　data　distribution　of　the　target　properties)　greatly　affected　the　accuracy　of　prediction　models.　The　uniformly　distributed　data　were　more　suitable　for　establishing　high-accuracy　BP-NN　models.　For　the　data　which　did　not　follow　a　uniform　distribution,　a　two-step　modelling　strategy　of　classification　and　subsequent　regression　was　proposed　to　improve　the　accuracy　of　micromagnetic　and　quantitative　prediction　of　surface　hardness.　The　proposed　two-step　modelling　strategy　integrating　the　K-nearest　neighbor　classification　(KNN)　algorithm　with　BP-NN　modelling　method　was　verified　by　external　validation　experiments.