Research On Deep Learning-based Classification Method Of Multisource Remote Sensing Data Fusion

In the context of remote sensing big data,fusing multi-source remote sensing data for terrain classification has become a hot issue in remote sensing intelligent information processing,and deep learning,as one of the most popular breakthrough technologies in big data processing,has been proven a powerful tool for making breakthroughs in many areas.Although some studies have been carried out with deep learning techniques to fuse multi-source remote sensing data for terrain classification,it is still in its infancy,therefore,it is of great practical significance to carry out research on multi-source remote sensing data fusion classification based on deep learning.The main research contents and conclusions of this paper are as follows:(1)The terrain classification methods of remote sensing data and the related theories of deep learning and multi-source remote sensing data fusion are summarized.Furthermore,a terrain classification framework consists of two-stream convolutional neural network for feature extraction and fusion of multi-source remote sensing data is constructed.In the feature extraction stage,the two-stream network is constructed to extract the feature information of hyperspectral imagery and LiDAR DSM data respectively.In the feature fusion and classification stage,the two branch networks are separately trained and the respective classification layer is removed to concatenate the feature map at the end.Finally,a new feature fusion and classification network is designed and utilized to complete the deep fusion of heterogeneous features and classification.(2)Aiming at the performance bottleneck problem of standard convolutional neural network architecture for remote sensing data feature extraction,a terrain classification method with two-stream densely connected convolutional network(DenseNet)for multisource remote sensing data fusion is studied.Firstly,the 3-D and 2-D DenseNet models are constructed to extract the spectral-space-elevation features of hyperspectral imagery and LiDAR DSM data.Then,the feature fusion and classification network composed of two fully connected layers are used to complete feature fusion and terrain classification.The experimental results show that the two-stream DenseNet model can achieve feature reuse and fusion at different levels,and obtain higher overall classification accuracy than the existing methods on the Houston and Trento standard datasets.(3)Considering at the problem of computational redundancy and low fine-grained segmentation with convolutional neural network architecture,a terrain classification model with two-stream fully convolutional DenseNet(FC-DenseNet)for multi-source remote sensing data fusion is proposed.The model adopts an encoder-decoder structure,the encoder uses 3-D and 2-D FC-DenseNet to downsample the hyperspectral image and LiDAR DSM data,and the decoder fuses the multi-level feature maps from the encoder through skip connection and upsamples with transposed convolution for full resolution semantic segmentation.The experimental results show that the overall classification accuracy of the two-stream FC-DenseNet method on the two datasets is improved compared with the proposed two-stream DenseNet method,and a better classification effect is achieved.