Abstract: To address challenges such as the large search space, long computation time, and insufficient prior information associated with matched-field geoacoustic inversion using ship-radiated noise in shallow-water environments, a seabed-classification-prior assisted passive geoacoustic inversion method is studied. Ship-radiated noise from a fixed shipping lane is used as a passive source. The received signals of seabed nodes are converted into spectrograms according to AIS (Automatic Identification System) tracks, and a ResNet18 model is used to classify seabed sediment types. The classification result is then used as a seabed-type prior. Kullback-Leibler divergence (KLD) is adopted to evaluate the statistical distribution difference between candidate geoacoustic models and the classification reference model, so as to constrain the search space of matched-field inversion. Adaptive simplex simulated annealing is finally used to estimate sediment sound speed, density and attenuation coefficient. Processing results of Zhoushan sea-trial data show that the ResNet18 model achieves an accuracy of 98.5% on the simulated test set, and the multi-node decision result indicates clay sediment. With the classification prior and KLD constraint, the inversion time is reduced from 6547 s to 3756 s, a reduction of about 43%. The proposed method improves the efficiency and convergence stability of passive geoacoustic inversion while retaining the physical interpretability of matched-field inversion. Unlike existing methods that directly employ deep learning to predict continuous geoacoustic parameters, the proposed method utilizes deep learning to generate seabed-type priors and subsequently integrates them with matched-field inversion for continuous parameter estimation. This strategy improves inversion efficiency while preserving the physical interpretability of matched-field inversion.