A Study On Fire Monitoring Method Based On Himawari-8 Multi-channel Data And Spatio-Temporal Characteristics

Wildfires have significant impacts on the atmosphere,terrestrial ecosystems,and society.Timely monitoring of wildfire locations is critical for fighting wildfires and reducing human casualties and property damage.With the advantage of high temporal resolution,geostationary satellites are increasingly being used for fire monitoring.However,limited by the level of satellite sensor technology,related fire monitoring methods（1）lack the use of multiple wavebands;（2）model training datasets are difficult to construct and model features are redundant;and（3）lack the fusion of temporal and spatial feature information.This leads to the inability of these methods to precisely locate the location of wildfire occurrence,and the defects of high leakage detection,low accuracy and poor generalization ability.Meanwhile,the geostationary satellite-based fire monitoring methods do not take full advantage of the multi-channel and high temporal resolution of geostationary satellites,which makes the near real-time wildfire monitoring methods fail to play a corresponding effect.In this thesis,a fire monitoring methodology study was conducted using Himawari-8 data as the primary data source for wildfires occurring in southeastern Australia in 2019-2020.The study was conducted in the following areas.Firstly,a fire labeled dataset was constructed using the stable VNP14IMG fire product,and fire detection was performed using a Random Forest（RF）model based on Himawari-8 multichannel data.Brightness temperature,spatial features and auxiliary data were used as inputs to this framework for model training.The impact of these features on the model accuracy is evaluated using recursive feature elimination methods and redundant features are excluded.Daytime RF models and nighttime RF models are constructed separately and their applicability is analyzed.The performance of the models is extensively evaluated by comparing with the wildfire products released by the Japan Aerospace Exploration Agency（JAXA）.Secondly,potential fire pixels were extracted by preprocessing and dynamic threshold tests.Temporal information analysis is performed by comparing the pixel brightness temperature observations with the mean and standard deviation of the previous 20 days of brightness temperature observations,and further analyzing the pixels spatial characteristics if there are anomalies in the observations.The fire pixel is finally determined by comparing the interclass variance between the pixel and the background pixels.It is also tested in the sample area and compared with the single temporal method as well as the RF method.The main findings and conclusions are as follows:（1）The RF model used in this study had a higher recall（95.62%）compared to the Himawari-8 L2WLF product.In particular,the model has a 19.44%higher recall for fire pixels with low fire counts.Adding waveband computation features,spatial features,and auxiliary data can improve the richness of the RF model input information,which in turn improves the accuracy of fire detection.The feature selection method effectively eliminates the redundant information caused by a large number of similar input features,thus improving the accuracy of the RF model.The most important factors affecting daytime and nighttime fire detection are Dif_07-14-T_Dif07-14features and Tbb07/Tbb12 features,respectively.（2）The fire detection accuracy of the daytime RF model is higher than that of the nighttime RF model.This is mainly caused by the fact that the difference between the daytime fire features and non-fire features is larger than that of the nighttime case.The omission errors of the RF model occur mainly in the detection of small fires,which are mainly located at the edges of the Himawari-8 pixels.In terms of spatial distribution,the omission and commission errors are mainly concentrated on the neighboring pixels of the fire cluster.（3）The average recall rate of the time-space method is slightly lower（3.95%）compared to the temporal method,which has a significant increase in precision（7.13%）.It indicates that the addition of spatial information can effectively exclude the spurious fire pixels from the temporal method to extract fire pixels.Among the temporal,RF and time-space methods,the RF method has the highest recall rate,mainly because the method uses the multi-channel information of Himawari-8 data,while the fire pixel labels constructed by the higher spatial resolution VNPIMG14 make the trained model have better detection ability for small fires as well.The time-space method has the highest accuracy and F1-score,indicating that the simultaneous use of both temporal and spatial features of the fire pixels can effectively improve the accuracy of the model while maintaining a good recall of the model.（4）In the fire monitoring method based on the time-space approach,the low spatial resolution of the sensor and the degree of fire burning lead to missed fire detection.In addition,the influence of neighboring pixels can lead to false detections from the model.