Artificial intelligence and multi-sensor earth observation for flood forecasting and mapping: a systematic review of operational evidence

Climate change, rapid urbanisation, and expanding floodplain exposure are increasing demand for spatially explicit flood information. Conventional hydrological models remain essential, but their operational application can be constrained by sparse observations, calibration requirements, and processing constraints in poorly gauged catchments. This systematic review evaluates how artificial intelligence (AI), multi-sensor Earth observation, and cloud computing support flood forecasting and flood-extent mapping. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework, 22 studies integrating synthetic aperture radar (SAR), optical imagery, satellite precipitation, altimetry, and terrain data with machine-learning and deep-learning methods were assessed. Because flood-extent mapping, flood-susceptibility classification, rainfall estimation, and rainfall–runoff forecasting involve different targets and metrics, evidence was synthesised by task. Multi-sensor integration was most valuable when complementary observations represented hydrometeorological forcing, surface-water conditions, and terrain controls. Ensemble and deep-learning models frequently performed strongly within individual studies; however, differences in datasets, flood types, spatial scales, reference labels, and validation designs precluded a universal ranking of model performance. Cloud platforms supported scalable preprocessing and product delivery but did not independently determine predictive accuracy. Operational transferability remains constrained by geographically concentrated evidence, inconsistent validation, limited interpretability, insufficient uncertainty quantification, and unequal access to data and computational resources. AI-enabled flood analysis can strengthen early-warning and emergency-planning systems when combined with hydrological knowledge, uncertainty communication, local expertise, and human oversight. Future research should prioritise common benchmarks, independent transfer testing, explainable and hybrid AI–physics models, and prospective evaluation of operational workflows.