基于多模态大模型的城市内涝短临预报方法探索OA

Urban waterlogging nowcasting provides an essential basis for efficient disaster emergency response.Existing methods mostly rely on locally developed physics-based flood forecasting models,which face several limitations,including limited transferability of models and experience from other areas,complex processing of multi-source input data,and high computational cost.These limitations make it difficult to meet the requirements of rapid forecasting.Large models are capable of jointly processing multimodal data and rapidly generating results,offering potential for improving urban waterlogging forecasting.However,the types and organization of input information for this task remain insufficiently investigated.This study develops an urban waterlogging nowcasting method based on a multimodal large model.The method integrates inundation status,future rainfall,and terrain conditions to infer the evolution of water depth over future time periods.Using three monitoring points in Dalian as a case study,ablation experiments were conducted on different input conditions.The results show that the proposed method can accurately predict urban water depth,with a mean absolute error of 0.03 m under the optimal input condition.The incorporation of terrain conditions significantly improves the forecasting performance,reducing the error by 28.8%～69.1%.This study provides a new pathway for applying artificial intelligence to urban waterlogging nowcasting.