Northeast India's multi-hazard warning moves past rainfall thresholds: an open platform ties landslides, flash floods and cut-off villages into one alert

An open-source platform puts landslide, flash flood and road-cut risk into a single warning chain, with alerts resolved at village level rather than district level. Landslide warning has relied mainly on time-invariant rainfall thresholds and field observation, while flash flood warnings usually stop at district scale and cannot separate individual villages in narrow Himalayan valleys. Against single-model systems and manual monitoring, SAMBHV reports detection F1 of 0.87, an 8% false alarm rate and 30-second alert latency, where the single-model system reports 0.72, 22% and 120 seconds and manual monitoring 0.55, 35% and 1800 seconds.

Alert level is set by a three-way vote of physics simulation, machine-learning prediction and ground verification, escalating only when at least two agree. A single model's high-confidence prediction does not trigger an alert unless the physics gate and ground verification support it. In the ablation, adding XGBoost spatial regression, the graph neural network road-blockage model, the ensemble decision gate and the ConvLSTM nowcaster in turn to an ARI-rules baseline raises F1 from 0.62 to 0.87 and lowers false alarms from 28% to 8%.

The platform keeps running without connectivity and issues CAP 1.2 alerts over six redundant channels, including a LoRa mesh for network dead zones. Earlier open-source disaster platforms were mostly single-hazard coordination tools without predictive modelling, multi-channel alerting or offline edge operation. The platform reports more than 72 hours of offline edge operation, alerts in 11 Indian languages, and a Sikkim prototype that recomputes alternative safe routes with pgRouting when primary roads are blocked.

The demonstrated scope is the Northeast India pilot, including the Sikkim prototype route-recomputation demo, 311 test cases and model evaluation on synthetic data; the platform is released under Apache 2.0 or compatible licenses and delivered as a 14-service Docker Compose topology.

The XGBoost spatial regressor and graph neural network use synthetic training data because official landslide inventories are unavailable; the 2/3 consensus threshold is set by domain expert judgement and not empirically optimized on labeled historical events; IndicBART translation runs in simulated mode and still needs human evaluation across the 10 target languages; sensor deployment and external API authorization keys are not fully in place.