Dynamic Drought-Persistence and Weather-Momentum Fusion for Confidence-Calibrated Multi-Horizon Forest Fire Risk Forecasting
Keywords:
Forest fire prediction. wildfire risk forecasting. deep learning, drought persistence. weather momentum. multi-horizon forecasting. confidence calibration. early warning.Abstract
Current weather, as well as accumulated dryness in the environment and the rate of change, are all important factors affecting forest fire risk. Typical data-driven fire-risk prediction methods are based on a single observation of meteorological parameters (at one point in time) and yield a single fire-risk prediction that may not capture the evolution of hazardous pre-fire conditions. The paper presents a novel deep learning-based proactive forest fire risk assessment framework called the Dynamic Drought-Persistence and Weather-Momentum Fusion Framework for Confidence-Calibrated Multi-Horizon Forest Fire Risk Forecasting, which is entirely software-based. The proposed framework handles temperature, relative humidity, rainfall, wind speed, air pressure, and past observations related to droughts in various temporal windows. The DPI is based on a continuous sequence of dry periods and on the deficit of accumulated rainfall and provides a measure of long-term environmental moisture stress. At the same time, a "Weather-Momentum Vector" (WMV) is generated using the temporal shifts in temperature, moisture, pressure, wind, and precipitation to determine if the conditions in the local atmosphere are conducive to or inhibiting fire. A dual temporal deep learning architecture is trained independently to learn short-term features of weather changes and long-term features of drought memory and then combines these features for estimating fire risk. The model does not provide a single static classification but instead provides the three-year time series of fire susceptibility forecasts for the 24-hour, 72-hour, and 7-day horizons, enabling the evolution of fire risk to be understood in advance of potential fire ignition. A mechanism of confidence calibration is built in to distinguish between reliable high-risk forecasts and less certain predictions and to minimize false alarms. The forecasted fire susceptibility and fire risk, the most contributing weather conditions, the level of confidence in the forecast, and the distribution of risk by location are shown in the interactive web dashboard using Flask. The methodology proposed offers a hardware-independent and scalable solution for the early detection of potential fire-conductive conditions and to aid in forest management decisions for fire prevention prior to the onset of smoke or flame.

