Deep Learning Detection Method Of Burned Area Based On Landsat8 Data

Due to the influence of various factors and the limitation of conditions,forest fire is extremely difficult to extinguish once it occurs.In addition,due to the limitation of natural environment conditions,it is not easy to carry out field investigation for disaster assessment after the occurrence of forest fire.Especially for a large area of forest grassland hinterland or artificial inaccessible areas.However,estimates of burned area are the basis for effective post-disaster assessments and the extent of damage to vegetation and soil.Therefore,this paper uses satellite remote sensing images to extract the area of forest fire with the deep learning method.Traditional burned area extraction algorithms mainly rely on the changes of the spectra or temperature and other physical features before and after the occurrence of a fire to build a physical model for extraction,or build vegetation index as a reference to carry out burned area detection.Therefore,the traditional extraction algorithm can only take into account the important physical change characteristics to extract the burned area.Therefore,in the process of extraction,the time resolution of remote sensing images is required to be very high,and there may be fewer features or defects that only take into account the main features,resulting in missing and misclassification.At the same time,it takes more time to extract fire from a large area.However,deep learning convolutional neural network can learn and process multiple features of data at the same time,which is more suitable for rapid and effective extraction of fire trace regions from remote sensing images.In order to verify the effectiveness of deep learning algorithm in burned area detection,this paper takes Landsat8 multispectral image processing as the main line,and centering on the hot issue of fire trace region classification,this paper conducts the following research:(1)According to the spectral characteristics and band characteristics of Landsat8multi-spectral remote sensing images,the paper first carried out a simple pretreatment of remote sensing images,namely radiation calibration and atmospheric correction.Face to the problem that all the bands of multi-spectral remote sensing images are used for burned area detection,which results in the reduction of data redundancy accuracy,the band optimization combination experiment is designed,and the most suitable band combination for burned area detection is obtained.(2)Aiming at the effectiveness of deep learning method in burned area detection and the lack of research.In this paper,we construct a new convolutional neural network,design parameter optimization experiments for the network model,and study the important parameters affecting the network performance.Furthermore,in order to verify the generalization ability of the model,the stability of the data input in different regions of the improved convolutional network was evaluated.At last,the common machine learning method and the improved convolutional neural network classification results were compared and analyzed to get the influence of different classifiers and different parameters on the classification results.(3)In view of the problem of long time series mapping,this paper use remote sensing data of 8 years in a small area of Amazon forest to discuss the distribution of test set and training set in time and estimates the overburned area by using deep learning method.In this paper,Landsat8 images are used to analyze the influence of band combination on burned area detection,and the effectiveness of deep learning method in burned area detection is verified.The results show that the deep learning method is better than the classical machine learning method.