Reliability and Conflict-Aware Closed-Loop Multimodal Edge Framework for Landslide Early Warning: A Simulation-Based Algorithmic Evaluation

This paper presents a reliability- and conflict-aware multimodal edge fusion framework for landslide early warning under sensor degradation. The framework treats modality reliability and cross-modal disagreement as separate factors and adjusts sensor contributions using availability, range validity, temporal stability, and drift. A frozen chronological evaluation protocol is developed, with warning thresholds selected under a validation-stage false-alarm constraint, using a real landslide inventory from three regions of Bangladesh. On the locked test set, the proposed method achieves 66.7% episode recall while producing fewer false alarms than rainfall-only and conventional fusion. Leave-one-episode-out evaluation across 16 reconstructed episodes further indicates higher recall and lower false-alarm rates. The study provides a reproducible framework for assessing robust multimodal edge warning under controlled sensor degradation, while distinguishing algorithmic robustness from operational forecasting performance.