A Deep Learning-Based Fire Monitoring Algorithm Using Himawari-8 Satellite Data

Remote sensing image data has long been used for variable inversion and environmental monitoring. In this study, we constructed a fire point model using diverse spectral parameters of Himawari-8 images with fire datasets. The model's loss was used to determine the distinction between fire points and non-fire points for various parameters, assisting in the parameter selection for the fire monitoring model. Himawari-8 image data in the visible, near-infrared and infrared bands were inversed to obtain reflectance and brightness temperature. A deep learning fire recognition model is built using selected combination parameters, context-variable parameters, and indices in conjunction with the fire database. The model's accuracy is 97.8%, with a loss and false detection rate for specific fire database of 2.8% and 1.62%, respectively, the latter of which is lower than that of MOD14 product 8.08%. It reduces the number of false fire warnings and improves fire monitoring accuracy.