Determination of sample fire points by combined medium and high-resolution datasets and its implementation in Himawari-8 to detect fire: Deep learning algorithm

The coverage of large area and high-frequency observation capability of the Himawari-8 satellite can play an important role in fire monitoring, however, the inadequate use of information sources and the lack of information mining capability of available methods cause their accuracy of fire point monitoring very limited. However, deep learning algorithms can efficiently use multiple types of information sources and perform deep information mining, which has great potential in fire monitoring. On the other hand, the reliance on many representative and high-quality sample data constrains the application of these algorithms in Himawari-8 fire detection. To solve this problem, this paper proposes a method for fire point sample determination with combined medium and high-resolution remote sensing data, using high-confidence fire samples from global MOD14/MYD14 fire products validated by Landsat-8 of high-resolution remote sensing data to provide sample data for Himawari-8 satellite fire detection. A method for converting the MODIS data to the Himawari-8 AHI correlation band is investigated. This method provides many high-quality sampled datasets for extracting Himawari-8 AHI fire point information using deep learning algorithms. Thus, they will be utilized to construct an algorithm called Deep Neural Network Fire Detection Algorithm (DPFDA). To make full use of the valid information of the Himawari-8 AHI, the selection methods of various feature parameters for deep learning algorithms were investigated, using 28 types of information sources such as apparent reflectance, vegetation index, mid-infrared-thermal infrared band combination, background information, and zenith angle. Fire detection experiments were conducted in forest and grassland fire-prone areas, showing that when compared with the original Himawari-8 AHI product, the pseudo-fire point detection rate was significantly reduced from 38.26% to 10.33%, and the recognition accuracy was improved from 58.67% to 65.34%.