Research On Fire Detection And Exploration Of Automated Cloud Detection Based On Himawari-8 Remote Sensing

In recent decades,long-term climate warming trends such as rising average temperatures and small temperature differences between day and night have caused frequent forest fires around the world.Forest fires are sudden,destructive,and uncontrollable natural disasters that cause huge losses to ecosystems and humans.Accurate,real-time access to fire information is critical to mitigating the impact of such disasters.Remote sensing satellites have become an ideal tool for monitoring such disasters due to their low cost,strong real-time and wide coverage.For nearly four decades,scientists have developed many algorithms and fire products that are well-validated around the world for forest fire monitoring and detection based on different remote sensing satellite sensors.However,the current fire detection algorithm has two disadvantages,firstly,do not consider the spatio-temporal characteristics of satellite data,secondly,the inaccuracy of the cloud recognition algorithm in the fire point detection algorithm.The research topic of this paper is to solve two deficiencies in the fire point detection algorithm,and propose different improvement methods respectively,in order to obtain a more accurate fire point detection algorithm.Firstly,for the first deficiency in the fire detection algorithm,the two main remote sensing data resources currently used for detection have the obvious shortcomings:Earth Observation(EO)satellites have high temporal resolution but low spatial resolution.The polar orbiting satellite have low temporal resolution but high spatial resolution.Therefore,the existing forest fire detection algorithms that are based on a single one of these two systems have only exploited temporal or spatial information independently.There are no approaches yet that have effectively merged spatial and temporal characteristics to detect forest fires.This paper presenting a spatiotemporal contextual model(STCM)that fully exploits geostationary data's spatial and temporal dimensions based on the data from Himawari-8 Satellite.The algorithm is divided into the following parts:We used an improved robust fitting algorithm to model each pixel's diurnal temperature cycles(DTC)in the middle and long infrared bands.For each pixel,a Kalman filter was used to blend the DTC to estimate the true background brightness temperature.Subsequently,we utilized the Otsu method to identify the fire after using an MVC(maximum value month composite of NDVI)threshold to test which areas have enough fuel to support such events.Finally,we used a continuous timeslot test to correct the fire detection results and named it STCM2.The proposed algorithm was applied to four fire cases in East Asia and Australia in 2016.A comparison of detection results between MODIS Terra and Aqua active fire products(MOD14 and MYD14)demonstrated that the proposed algorithm from this paper effectively analyzed the spatiotemporal information contained in multi-temporal remotely sensed data.In addition,new forest fire detection algorithm have higher detection accuracy than traditional contextual and temporal algorithms.Secondly,for the second deficiency in the fire detection algorithm,the accuracy of the current cloud recognition algorithm has a great influence on the accuracy of fire detection,but the cloud recognition algorithm in fire detection is based on the fixed threshold method of artificial experience.This requires a lot of prior knowledge,and considering the application of different time and different regions,the thresholds are often conservative.These cloud recognition algorithms can not meet the requirements of high precision,high automation and high robustness of fire detection algorithms.Therefore,this paper explores the effects of four machine learning algorithms(SVM,kNN,Kmeans,BPNN)and artificial threshold-based dynamic threshold method and fixed threshold method in cloud recognition.The algorithm is divided into two parts.Firstly,t algorithm with the highest cloud recognition accuracy among the six algorithms is selected.Finally,this algorithm is brought into the previously proposed STCM2 fire detection algorithm to compare and analyze the influence of the improvement of cloud recognition algorithm on the accuracy of fire detection.The results show that qualitative visual interpretation and quantitative accuracy evaluation show that the four machine learning algorithms(SVM,kNN,Kmeans,BPNN)are more artificial than artificial,as the traditional MODIS cloud product(MOD35/MYD35)is used as the standard verification data.The dynamic threshold method and the fixed threshold method of the threshold are better,the average overall accuracy is higher than 16.74%,and the false detection rate is lower than 50.21%.This is because the machine learning algorithm can learn the characteristics of the cloud autonomously.However,the fixed threshold method and the dynamic threshold method have a large space-time range and the threshold selection is relatively conservative,so the cloud recognition accuracy is low.At the same time,BPNN performs best in four machine learning algorithms due to its powerful autonomous learning characteristics.Compared with the worst performing fixed threshold method,the overall classification accuracy is higher than 23%,and the false detection rate is lower than 74.5%.The development of cloud recognition algorithms should be considered in terms of multi-feature learning and deep learning.Finally,the most accurate BPNN cloud recognition algorithm among the six algorithms is brought into the previous fire detection algorithm STCM2.Compared with the previous detection results,the omission error is reduced by 4.30%,and the commission error is decreased by 14.7%.It shows that the higher precision cloud recognition algorithm can greatly reduce the misclassified pixels in the fire point recognition algorithm,but the number of missing fire point pixels is reduced less.These reveal the insufficiency of the anemone sensor in the detection of fire points.It is possible to combine the advantages of various sensors to develop a more accurate fire point detection algorithm in the future.