A novel rapid flood mapping model based on social media and GF-3 satellite imagery

•A novel rapid flood mapping (NRFM) model is proposed base on social media and GF-3.•The robust NRFM model overcomes the unexplainable nature of data-driven methods.•Flood inventory can be generated with high identification rates.•The performance of NRFM model was assessed for rapid flood mapping.•Pre-training method can greatly improve the model inference efficiency. Flood mapping is paramount in flood risk management as it furnishes crucial information and support to emergency responders and decision-makers for effectively managing flood-related events. Traditional data sources like remote sensing and hydrodynamic models frequently face limitations due to data availability and time-consuming computations. To address these challenges and improve the efficiency of urban flood mapping, we propose a novel Rapid Flood Mapping (NRFM) model based on social media and GF-3 satellite imagery. The model incorporates geographic information system (GIS) theory and methods, alongside edge detection techniques like AdaptiveThreshold, and parameter optimization methods such as AdamW. Various modal data were employed as inputs to enhance the model’s learning capabilities. Evaluation of the model output was conducted using multiple metrics, including accuracy, F1-score, mean absolute error, and efficiency. Testing of the model was conducted in Zhengzhou City, yielding the following results: (1) The model, which adheres to the first law of geography, the law of gravity, and outlier detection techniques to enhance mechanism explanation, yielded remarkable results, with a mean accuracy of 0.98 and a mean F1-score of 0.98. (2) The proposed model efficiently generated flooded areas by pre-training subsamples and fine-tuning overall samples to obtain predictive models. In the case of Zhengzhou, covering an area of 7567 km2, the model required only 10 min on an Intel Xeon(R) Silver 4214R CPU. The NRFM model can offer optimal parameters as references for subsequent studies and provide consistent explanatory estimates of flood inundation in near real-time, thus significantly improving the basis for macro-decision-making and rescue efforts during flooding events.