Developing an Original Deep Learning Architecture for Multi-Hazard Early Warning Systems in Northern Nigerian Environments

Northern Nigeria faces concurrent and interacting hazard risks including Sahel drought, flash flooding, locust swarms, and conflict displacement, and existing early warning systems treat each hazard independently without leveraging the multi-hazard correlations that could enable earlier and more accurate compound-event alerts. This study developed an original deep learning architecture for multi-hazard early warning in northern Nigerian environments, introducing a novel multi-task learning framework trained on heterogeneous sensor and satellite data sources. A systematic review of 62 early warning system publications from 2019 to 2024 was conducted, revealing that no existing architecture explicitly modelled inter-hazard correlations for compound events in the Sahel climatic zone. The original Multi-Hazard Correlation Network (MHCNet) architecture was designed, implementing a shared representation backbone processing time-series inputs from five data streams: NOAA NDVI satellite imagery (drought indicator), radar precipitation composites (flood indicator), MODIS land surface temperature, soil moisture estimates, and conflict event reports from ACLED. Task-specific output heads generated independent 7-day hazard probability forecasts for each category, trained jointly using a novel inter-hazard correlation loss function that penalised predictions that violated historically observed hazard co-occurrence patterns. MHCNet was trained on 15 years of multi-hazard event data across Borno, Yobe, Sokoto, and Kebbi States. F1-score for compound drought-conflict events improved by 22.4 percentage points over independent single-hazard models. The study constitutes an original contribution to multi-hazard deep learning methodology and recommends NEMA integrate MHCNet into its northern zone emergency preparedness operations.

Keywords: multi-hazard early warning, deep learning, Northern Nigeria, MHCNet, compound events