Study On Remote Sensing Finer Classification Of Freshwater Wetland Based On Deep Learning

The wetland is an ecosystem formed by the interaction of water and land on the earth.It has an irreplaceable role in maintaining ecological balance,improving regional climate,and regulating runoff.China is abundant in wetland resources,which cover about 10% of the world's wetland area.The natural wetland area of Heilongjiang Province is 5.56 million hectares,accounting for 1/8 of the national natural wetlands.With the agricultural expansion,urban development and the global climate change,which are the causes of such as changes in the hydrological conditions of the image factors,the wetland ecosystem has been impacted significantly,especially on freshwater wetlands.It is necessary to timely and dynamically monitor the wetland vegetation and its surrounding land with advanced technical means.This paper takes Honghe Nature Reserve and Wudalianchi Nature Reserve in Heilongjiang Province as the research object based on summarizing and reviewing the research progress of the wetland classification at home and abroad,and makes the fine classification of freshwater wetlands by using high-resolution remote sensing images,and applying methods such as deep learning and compound classification and others.It explores the best model methods which are suitable for the classification in highresolution remote sensing images,and analyzes the advantages and disadvantages of deep learning methods in the wetland classification.The research contents and conclusions of the thesis are as follows:(1)Research on convolutional neural network(CNN)in the classification of freshwater wetlands.A convolutional neural network structure suitable for classification of freshwater wetlands was constructed,and was compared with the shallow classification method in classification accuracy and classification results.The results showed that CNN,as a deep-structured classifier,explored the complex spatial patterns hidden in highresolution remote sensing images,and could extract richer semantic features.These spatial patterns could not be found in shallow structures.However,the processing accuracy of the CNN classifier for the boundary was not as good as that of the SVM classifier.There was a phenomenon of uncertainty along the boundary of the object in the classification,which led to excessive smoothing to some extent.On the other hand,in the CNT classifier,even if the objects with significant features had little spatial information,they may be misclassified.CNN had a 4%-6% improvement in the classification accuracy of the shallower layers in the two study sub-regions,and the identification of complex wetland vegetation was significantly better than the shallow classification method.(2)Research on composite classifiers in the classification of freshwater wetlands.The existing convolutional neural network was improved,and the multi-connected layer was built into a multi-layer perceptron(MLP)to realize the combination of CNN features and MLP in the classification model,which could improve the efficiency of the model.Then the SVM-CNN composite classifier was constructed by decision fusion method.The classification results were optimized by studying the threshold of composite classifier.The research results showed that the SVM-CNN method combined the advantages of the SVM method and the CNN method.The support vector machine method showed great advantages in the classification of features with clear boundary types.For example,the cultivated land was distributed in a block shape,and a clear boundary could be found between it and other ground objects;on the contrary,there was a large transition zone between the swamp and the aquatic vegetation,and it was difficult for the support vector machine to find strict classification boundaries between species,thus the classification accuracy of aquatic vegetation was relatively low.The CNN method simulated the working mode of the human visual cortex by using multiple convolution and pooling operations,and realized weight sharing by translation invariance,so that rich and accurate spatial features could be extracted from the image block.Therefore,the CNN classification method was superior to the SVM method in the classification accuracy of grassland and swamp.The experimental results showed that compared with the CNN classification method,SVM-CNN had a greater advantage in the classification of wetlands in lakes.(3)Research on classification of deep learning in different resolution images.By applying the deep learning method for the classification of high-resolution remote sensing image Sentinel-2A in wetland remote sensing images,which was further validated the effectiveness of the deep learning classification method in high-resolution remote sensing image classification.In addition,due to the difference in resolution between Sentinel-2A and GF-2 images,the resolution of remote sensing images suitable for deep learning methods was studied.The results showed that the Sentinel-2A image directly applied the deep learning method and the SVM method for extracting texture features.The deep learning method did not show strong advantages,mainly because the resolution of Sentinel-2A was not as good as GF-2.High resolution,and no clear contextual semantics and texture features can be obtained directly from the image.(4)Relevant theories of deep learning were sorted out,studying the best model structure and parameters which were suitable for wetland remote sensing image classification,and analyzing the overfitting problem.Three ways to alleviate the overfitting problem,Dropout,global average pooling layer and Dither were studied.The results showed that both Dropout and Dither used by CNN could improve the classification accuracy and effectively,and also could prevent over-fitting.(5)Research on image fusion method of high score 2 remote sensing image.In order to achieve the fine classification of wetland vegetation,the image fusion was performed on GF-2 remote sensing image by applying Gram-Schmidt(GS)method,NND method and HPF method respectively,and the best fusion method that was suitable for highresolution 2 image wetland vegetation was obtained,which laid the foundation for subsequent fine classification.The experimental results showed that the G-S image fusion method had the largest amount of information,and the spatial detail extraction was more detailed.