An Advanced Deep Learning Framework for Mapping Forest Fire Smoke and Air Pollution Impacts on Human Health Using Satellite Imagery

Abstract

The smoke produced during forest fires is a large-scale plume that contains high levels of hazardous air pollutants (HAPs) that can affect a wide geographical area and cause environmental and public-health problems. Traditional monitoring systems, however, only monitor at a small number of spatial locations, and they are not always sufficient for determining population level smoke exposure. In this study, an advanced deep learning approach is suggested to extract information about forest-fire smoke, air pollution, and impacts on human health from multispectral satellite data and observations from the atmosphere. This framework combines the spatial feature extraction by Convolutional Neural Networks (CNNs) with the segmentation of smoke-plumes by U-Net and the spatiotemporal analysis of pollution by Long Short-Term Memory (LSTM) networks. Satellite-based aerosol optical depth, fire radiative power, land surface temperature, vegetation indices, PM2.5 and air quality observations are combined to define exposure intensity. The experimental results show that the mapping of smoke affected areas is accurate with 96.75%, precise with 96.18%, recalled with 95.62%, F1 score with 95.90% and AUC with 0.982, and IoU with 93.84%. The novelty of the approach is related to the integration of fire detection, segmentation, estimation of pollution and mapping of health risk in a given architecture. The framework delivers scalable identification of pollution hotspots and communities with high exposures. The results show its potential in the support of early health advice, environmental surveillance, data-driven disaster-response planning.