Wildfire Detection Using Streaming Satellite Imagery

Abstract

We propose a novel wildfire detection algorithm for multispectral satellite images. By observing that wildfire pixels are sparse outliers residing in a spatially correlated background, we isolate them using robust principal component analysis. A novel cloud masking approach based on T-point thresholding is also proposed to reduce false alarms. Compared to existing methods, our proposed method adapts to the spatial and temporal heterogeneity of satellite images, does not require training on labeled images, and is computationally efficient for online monitoring. We present an application of our proposed algorithm to the GOES-R imagery in monitoring recent California wildfires.