Deep Learning-based Multi-source Remote Sensing Image Classification Of Dunes

Desertification has now become one of the major global environmental problems faced by human society,and has attracted wide attention from various countries around the world.In the northwestern inland area of my country,desertification is mainly manifested as desertification,and sand dunes are hilly or ridge-like landforms formed by the accumulation of sand particles under the action of wind,and are the main landform types of deserts and sand lands.Sand dunes account for 96% of the desert and sand area in my country.The development of sand areas,ecological environment protection and other aspects need to master the main forms of sand dunes and its formation rules.Therefore,the use of efficient and reasonable methods to monitor the dune morphology is a necessary measure to protect the ecological environment of the oasis.It has positive significance for the study of the regional characteristics of the aeolian landform,the formation of the evolutionary environment and process,and the reasonable allocation of measures for sand prevention and control.In the study of dune morphology classification,most of them use visual interpretation to classify the dune morphology.Although this method has high accuracy,it is time-consuming and laborious.In addition,it is relatively difficult to distinguish the junctions of different sand dune forms.The characteristics of some types are very similar.The boundaries are blurred,and the differences in the features reflected on the remote sensing images are small,which also makes the classification work difficult.In response to the above problems,this article carried out the following work:First of all,based on the content of remote sensing image classification and deep learning,systematically consult the current research status and research progress at home and abroad,and analyze and summarize the relevant research work and methods,and propose multi-source high spatial resolution remote sensing images as data sources combined with deep learning Conduct a classification study of sand dunes.First,based on Sentinel-2A,GF-1 and Google Earth remote sensing image data,combined with field surveys,CNN network Res Net-50 using image-level image classification combined with Sentinel-2A multi-scale sample data to the south of Gurbantunggut Desert The dune morphology of the edge is classified,and then combined with multi-source and multiscale sample data to classify the dune morphology,and then the accuracy evaluation of the two methods is carried out to test the applicability of the method.Then,based on the Sentinel-2A data,the FCN network FCN-VGG of pixel-level image classification was used to classify the dune patterns on the southern edge of the Gurbantunggut desert.Then the object-oriented multi-scale segmentation method is used to scale the dune boundary staggered area,and the appropriate parameters and segmentation scale are selected to obtain a new batch of sample sets and combined with FCN-VGG to improve the performance of the classifier and solve the dune shape.The problem of boundary semantic segmentation.The conclusion is as follows:1)Convolutional neural network can achieve accurate extraction of dune morphology over a large area in the study of dune morphology classification.The experimental results show that,on the problem of sand dune morphology classification,multi-scale sample data sets are used to train the model,which can solve the application problem of small samples in deep learning.In addition,the use of multi-source data to fine-tune the pre-trained model shows that transfer learning is effective to overcome the image difference problem caused by different sensors.Compared with the model before fine-tuning,the overall accuracy is improved by 2.4%,and the kappa coefficient is improved.0.0288.2)In the study of dune morphology classification,the fully convolutional neural network can solve the problem that it is difficult to accurately extract the boundary at the dune junction.The experimental results show that,on the problem of dune morphology classification,FCN-VGG combined with object-oriented multi-scale segmentation can effectively improve the classification accuracy at the dune junction.Compared with the FCN without the object-oriented method,the overall classification accuracy and kappa coefficient are 2.45% and 0.0314 higher,respectively,which proves that this method can improve the accuracy of the fully convolutional neural network in extracting the dune morphological boundary information.3)The effect of object-oriented multi-scale segmentation is greatly affected by the segmentation scale,shape parameters,and compactness index.In this paper,ESP is used to quantitatively evaluate the segmentation effect,and finally the optimal between different dune morphologies at 50～200 scales is determined.Segmentation scale and optimal parameters.