[Objective] Physically based hydrological-hydrodynamic models for urban pluvial flood simulation are computationally expensive. In contrast，deep learning-based approaches may produce spurious shallow water depths in non-inundated areas，leading to distorted inundation extents. To address this issue，this study develops MFGRU-Y，a classification and regression collaborative model that jointly predicts wet-dry status and maximum inundation depth from rainfall sequences to improve both depth accuracy and inundation-extent delineation while maintaining high simulation efficiency. [Methods] As a multi-task surrogate model，MFGRU-Y jointly predicts inundation occurrence （wet-dry status） and maximum inundation depth from single-event rainfall time series. To accommodate rainfall events with varying durations，each sample retains its effective sequence length，and variable-length sequence encoding is employed to ensure that the recurrent unit focuses on valid time steps rather than padded zeros. The architecture includes a GRU-based rainfall encoder，a shared representation layer，and two heads for wet-dry classification and depth regression，with a non-negativity constraint applied to depth outputs. A key feature of the proposed model is that classification confidence is explicitly used to guide depth regression，thereby improving inundation-boundary delineation. To address the problem of inundation extent distortion，classification confidence is used to constrain regression through a gating design that suppresses spurious shallow depths in predicted dry regions； during inference，a binary wet-dry mask is further applied to the depth field. Training uses an imbalance-aware classification loss，a wet-region-emphasized robust regression loss，and learnable loss weighting to adaptively balance the two tasks and reduce manual tuning. [Results] MFGRU-Y achieved the best overall performance among all evaluated models. For maximum inundation depth prediction，MFGRU-Y achieved an MAE of 0.0093 m，an RMSE of 0.05 m，and an R² of 0.99 on the test set，indicating close agreement with reference simulations across events. For inundation extent evaluation，MFGRU-Y reached an IoU of 0.987 3，while the wet-dry classification accuracy was 99.4%，showing that the predicted inundation footprint was highly consistent with the ground truth. In addition to these aggregate metrics，MFGRU-Y demonstrated clear practical advantages in terms of boundary interpretability. In comparative experiments，competing deep learning baselines exhibited severe false positives，including cases in which dry nodes were almost entirely misidentified as inundated，resulting in distorted inundation extents. MFGRU-Y substantially reduced such artifacts by using classification confidence to constrain depth prediction during training and by applying an explicit wet-dry mask during inference. As a result，the final depth fields showed fewer spurious shallow patches in non-inundated areas and a sharper，more coherent inundation boundary. Importantly，these gains in extent consistency and boundary clarity were achieved without degrading overall depth accuracy，supporting the effectiveness of collaborative wet-dry discrimination and depth regression for urban pluvial flood prediction. Furthermore，in terms of computational efficiency，MFGRU-Y significantly reduced the single-event simulation time from the minute scale required by the traditional mechanistic physically based model to the second scale. [Conclusions] The experimental results demonstrate that MFGRU-Y delivers both high depth accuracy and high inundation-extent fidelity and significantly improves boundary interpretability compared with alternative deep learning models that may overpredict inundation in dry areas. From an application perspective，MFGRU-Y provides more reliable inundation extents and maximum depth fields for risk assessment and emergency response，where boundary-related false inundation can otherwise mislead decision-making. Future work may focus on two aspects： model architecture refinement and input feature enrichment. On the one hand，the model's focus on deep-water zones should be strengthened. On the other hand，richer feature information，such as topography，hydraulic engineering operation schedules，and the initial water levels in river and drainage networks，should be incorporated，thereby further enhancing the reliability of engineering applications.