Remote Sensing meets Deep Learning: Exploiting Spatio-Temporal-Spectral Satellite Images for Early Wildfire Detection

Wildfires are getting more severe and destructive. Due to their fast-spreading nature, wildfires are often detected when already beyond control and consequently cause billion-scale effects in a very short time. Governments are looking for remote sensing methods for early wildfire detection, avoiding billion-dollar losses of damaged properties. The aim of this study was to develop an autonomous and intelligent system built on top of imagery data streams, which is available from around-the-clock satellites, to monitor and prevent fire hazards from becoming disasters. However, satellite data pose unique challenges for image processing techniques, including temporal dependencies across time steps, the complexity of spectral channels, and adversarial conditions such as cloud and illumination. In this paper, we propose a novel wildfire detection method that utilises satellite images in an advanced deep learning architecture for locating wildfires at pixel level. The detection outputs are further visualised in an interactive dashboard that allows wildfire mitigation specialists to deeply analyse regions of interest in the world-map. Our system is built and tested on the Geostationary Operational Environmental Satellites (GOES-16) streaming data source. Empirical evaluations show the superior performance of our approach over the baselines with 94% F1-score and $1.5$ times faster detections as well as its robustness against different types of wildfires and adversarial conditions.