Research On Deep Learning Fire Point Detection Based On Geosynchronous Satellite Data

Forest fires have occurred frequently in recent years.Forest fire is an important part of natural ecology,with a variety of ecological benefits.Most natural fires do not require additional human intervention,but in some cases,intensely burning forest fires pose a serious threat to the safety of human life and property.Forest fires have the characteristics of fast spreading speed.Small fires in the early stage of fires are relatively easy to intervene,but in a short period of time,fires can gradually spread to the surrounding into large fires that are difficult to control.Therefore,it is necessary to detect fires as quickly and accurately as possible.Remote sensing satellite images have high resolution,can be monitored for a long time,large area,and low cost,and have become an ideal tool for detecting forest fires.Around different remote sensing satellites,researchers have designed many fire detection algorithms,and many countries have also actively developed many fire products using these remote sensing satellite fire detection algorithms.However,the current fire detection algorithm mainly uses the threshold method to identify the fire point.These methods are very sensitive to the setting of thresholds.Due to the different characteristics of fires in different seasons and landforms,the same threshold is difficult to apply to other regions.Therefore,researchers began to use deep learning algorithms with intelligent and automated features to detect fire points.At present,in the field of remote sensing,semantic segmentation models are mainly used to map fires in remote sensing images.However,these semantic segmentation models usually require that the input image has high spatial resolution and can provide more detailed semantic information.Therefore,they are usually applied to solar synchronous satellite images with high spatial resolution,which is difficult to apply to geosynchronous satellites.However,the sun synchronous satellite revisit time is long,it is difficult to achieve near real-time monitoring.The geosynchronous satellite with high time resolution provides the possibility for real-time fire monitoring.Therefore,it is necessary to design a deep learning method to detect the fire point of the geosynchronous satellite data.Due to the low spatial resolution of geosynchronous satellite images,the accuracy is still limited even if the semantic segmentation deep learning model is applied.And due to the lack of semantic information caused by low spatial resolution,the deep learning model is easy to fall into the local optimal solution.The lack of semantic information will also lead to the low robustness of the model to cloud,water,cloud edge and thin cloud.At the same time,the current deep learning methods do not make use of the high temporal resolution of geosynchronous satellite data,and do not extract features in the time dimension.Therefore,it is necessary to improve the original deep learning model and design a deep learning model more suitable for fire detection tasks.At the same time,due to the limitation of spatial resolution,the number of fire pixels in geosynchronous satellite images is very small,which leads to an extreme imbalance between the number of fire points and the number of non-fire points,which is not conducive to the training of deep learning models.Therefore,it is also necessary to design a strategy to solve such data imbalance problems.According to the characteristics of geosynchronous satellite data,this paper proposes two strategies for fire detection.One is the convolutional neural network FireCNN based on artificial features,and the other is SBT-FireNet based on spatial,band and time features.FireCNN increases the feature representation of fire points by manually adding background information.Using multi-scale convolution and residual design,it can learn the features of fire points from different scales and prevent the original features from being ’ forgotten ’ in the learning process.It can effectively extract the accurate features of fire points and mine the deep features of fire at different scales.FireCNN was tested on a dataset containing 1823 fire points and 3646 non-fire points.The experimental results show that FireCNN is fully capable of fire detection tasks,and its accuracy is 35.2 % higher than the traditional threshold method.Although FireCNN has high accuracy,it cannot be separated from the additional features of manual design,and cannot achieve complete automatic feature extraction,and it is too fitted with data labels to supplement the traditional algorithm.The SBT-FireNet uses three independent feature extraction modules to comprehensively extract the spatial,band and time features of the fire point,and does not use any artificial features at all.Through the three feature extraction modules,the environmental information,own band information and time information of the fire point are studied respectively,and their spatial,band and time features are output respectively.Finally,the features extracted by each module are integrated through the full connection layer for comprehensive identification.The accuracy of SBT-FireNet in four test areas is about 35 % higher than other deep learning classification methods.At the same time,in order to solve the problem of serious data imbalance in geostationary satellite data,this paper proposes a series of strategies to alleviate this phenomenon,including data enhancement methods that can expand the number of fire points and in order to make the created data set as reasonable as possible.The ’ rumination ’ method,using these methods to construct the data set of the training set,can retain the sample points that are easily misclassified by the network model as much as possible,so that the network can fully learn the characteristics of various types of misclassified points,thereby improving the robustness of the network to clouds,water,cloud edges,and thin clouds.