Near Real-time Wildfire Monitoring Methods Based On Himawari-8 Data

In recent years,wildfires occur frequently around the world,seriously endangering human life and health as well as the normal order of social production activities.Forest fires in southwest China and northeast China are also affected by forest fires all year round,which cause great harm to local people and ecology.Therefore,wildfire prevention and control has been the focus of local government work.As an important means of fire monitoring,near real-time monitoring of wildfire is the prerequisite and key link of fire fighting.With the development of remote sensing technology,wildfire monitoring based on geosynchronous satellite data has become the mainstream method in the industry.However,due to the limitations of satellite sensor technology,the relevant monitoring algorithms have low spatial resolution and cannot accurately locate the location of wildfires,and also have defects such as high missed detection and low accuracy.These problems limit the advantages of high time resolution based on geosynchronous satellite and make the near real-time wildfire monitoring algorithm fail to play its corresponding effect.In this study,himawari-8 remote sensing data were used to study large-scale fire monitoring methods in southwest and northeast China.The improved traditional wildfire monitoring method and multiple machine learning methods were used to design the monitoring model,and the effectiveness was evaluated in the validation dataset.The model with high accuracy and low missed detection rate was selected.Based on the optimal model,a fire monitoring pipeline（Pipline）integrating near-real-time data acquisition,analysis and fire point analysis is implemented.It solves the problems of low timeliness and high missed detection in wildfire monitoring based on geosynchronous satellite data,and ensures certain monitoring accuracy.The main work and related achievements of the thesis are as follows:（1）Improve the traditional wildfire monitoring algorithm based on spatial information in spatial context,improve the monitoring accuracy of traditional methods,and reduce the missed detection rate.A variety of intermediate functions were added and ablation experiments were conducted to screen the optimal scheme.The validation data set was constructed with VNP14 IMG and MCD14 ML of firepoint products.The results show that the algorithm with three steps of spectral detection,land cover and context operator is simple and the F1 score is high.The missed detection rates were 68.81%（southwest China）and 71.2%（northeast China）in 20 images in each of the two study areas.It shows that the traditional methods have performance defects.（2）Use mainstream machine learning models to improve the monitoring effect.Combined with a variety of information,three types of feature schemes are constructed,and Random Forest（RF）and Light Gradient Boosting Machine（Light GBM）are used.Multilayer Perceptron（MLP）training;Convolutional Neural Network（CNN）is designed to learn the spatial spectral features.Multiple models were verified based on 20 images in each of the two study areas.The results showed that RF based on spectral feature training achieved the missed detection rate of 14.23%（southwest China）and 15.52%（northeast China）.Accuracy: 54.13%（southwest）,54.17%（northeast）,better than other models.It has certain application value.（3）Build a near real-time wildfire monitoring pipeline.Using Himawari Standard Data（HSD）with low time delay,combined with the selected optimal RF model,a Pipline covering four provinces in southwest China was constructed.Three-hour operation verification was carried out to achieve a maximum processing time of 238.48 seconds and an average processing time of 76.86 seconds.Compared with WLF,himawari-8’s own fire spot product,the daily scale verification was carried out,and 6 wildfire events were successfully monitored,4 were consistent with WLF,and 2 additional events were monitored.In this way,the real sense of high precision,low missed detection,near realtime and even real-time fire monitoring is realized.