A cloud-based Geo-AI framework for automated high-resolution flood mapping with explainable machine learning

Abstract

Rapid flood mapping is critical for emergency response, but operational workflows often depend on manual training-label generation and have limited multisource integration and interpretability. Here we present a hybrid cloud-local geospatial artificial intelligence (Geo-AI) framework that combines multisource data processing in Google Earth Engine with local machine-learning benchmarking and interpretation. The framework derives high-confidence flooded and non-flooded training samples from Sentinel-1 time-series statistics using a conservative Z-score threshold of ≤ − 3.5, enabling supervised learning without manual annotation. Applied to the 2022 Pakistan floods, the workflow compared 17 algorithms. LightGBM ranked first according to the composite normalized score, whereas random forest, extra trees and gradient boosting showed comparable classification performance with different precision-recall and timing trade-offs. SHAP-based attribution identified radar backscatter, shortwave infrared reflectance, the Normalized Difference Water Index and elevation as dominant predictors. Product-level spatial inter-comparison with an external Sentinel-1 flood product showed agreement over major floodplain features and identified discrepancies. The study demonstrates an integrated and interpretable workflow for event-specific flood mapping, while further evaluation across regions and of end-to-end operational latency is required.