Integrating Multi-Resolution Remote Sensing Data for Daily Forest Fire Risk Forecasting in a Nigerian Savanna Ecosystem

Forest fires represent a persistent environmental and socio-economic threat across West African savanna ecosystems, where limited ground-based monitoring and persistent cloud cover constrain effective early-warning systems. Remote sensing offers a powerful alternative for wildfire monitoring; however, the trade-off between high spatial resolution and high temporal frequency remains a key challenge for operational fire risk forecasting. This study developed a pragmatic spatio-temporal data fusion framework for daily forest fire risk forecasting using multi-resolution remote sensing and meteorological data, with Kainji Lake National Park, Nigeria, as a case study. High-resolution vegetation indices derived from Sentinel-2 imagery were integrated with daily MODIS surface reflectance products and ERA5-Land meteorological reanalysis data within the Google Earth Engine platform to generate a continuous daily dataset at a harmonized spatial resolution. VIIRS active fire detections was deployed, enabling the formulation of forest fire forecasting as a temporal classification problem. Random Forest and Extreme Gradient Boosting (XGBoost) models were trained. Both models demonstrated strong predictive performance on an independent test dataset, achieving high discrimination between fire and no-fire days. Random Forest slightly outperformed XGBoost, attaining an area under the receiver operating characteristic curve of 0.997 and an F1-score of 0.957, while both models achieved perfect recall for fire events. The results highlighted that daily fire risk in Kainji Lake National Park was primarily governed by seasonal and atmospheric conditions rather than vegetation greenness alone. The proposed framework provides a scalable foundation for early-warning systems and fire management applications in Nigeria and similar regions across West Africa.