Modal　parameters　are　structural　inherent　characteristics　that　can　be　applied　for　revealing　performance　of　railway　bridges.　Free　vibration　signals　generated　by　a　passage　of　train　are　commonly　utilized　to　estimate　the　modal　parameters　of　railway　bridges　due　to　their　higher　signal-to-noise　ratios　compared　to　random　vibrations　caused　by　ambient　loads.　However,　since　free　vibration　signals　rapidly　decay　over　time,　the　available　free-vibration　data　is　typically　short-time.　When　using　the　fast　Fourier　transform-based　spectral　estimation　method　for　modal　identification　from　short-time　vibration　data,　a　phenomenon　known　as　spectral　leakage　occurs,　leading　to　miss-identification　of　some　structural　modes.　In　this　study,　the　classical　frequency　domain　decomposition　(FDD)　is　improved　for　modal　identification　of　railway　bridges,　in　which　the　higher　resolution　auto-power　spectral　density　(PSD)　and　cross-PSD　functions　are　calculated　through　the　autoregressive　(AR)　model-based　method.　The　AR　model-based　method　improves　both　the　smoothness　and　resolution　of　the　PSD　functions　compared　to　the　fast　Fourier　transform　technique.　These　AR　model-based　PSD　functions　are　then　employed　in　the　FDD　process　to　facilitate　frequency　and　mode　shape　identification　while　avoiding　spurious　noise　modes.　The　proposed　eigenvalue　fitting　technique　is　subsequently　utilized　to　estimate　damping　ratios.　Numerical　simulation　data　as　well　as　vibration　data　from　an　actual　bridge　are　analyzed　to　validate　the　proposed　method,　with　a　comparison　made　to　the　Welch’s　PSD-based　method.　The　results　demonstrate　that　the　modified　FDD　approach　enables　more　effective　identification　of　structural　modes,　even　in　the　presence　of　closely-spaced　modes.　©　The　Author(s)　2024.