High-resolution Remote Sensing Image Classification Based On Fully Convolutional Neural Network

Remote sensing image classification plays an important role in remote sensing image interpretation and is the basis for conducting various remote sensing technology applications.High-resolution remote sensing image contains rich texture and structure information,but its classification results are easily disturbed by factors such as unbalanced target category,large change of target scale and mixing of ground objects,etc.Improving the classification effects and classification accuracy of high-resolution remote sensing images is an urgent problem to be solved.Traditional remote sensing image classification methods often adopt manual features extraction,which is subjective,time-consuming and laborious,while machine learning algorithm only has a shallow model structure and limited features extraction,both of which are difficult to obtain better classification results.Deep learning can extract features from low to high level of the input image and automatically build feature mapping models.Therefore,this paper proposes a residual encoder-decoder full-convolutional neural network based on the fully convolutional neural network to address the problems such as mixing of ground objects and poor classification integrity in the classification of high-resolution remote sensing images.The main research contents of this paper are as follows:(1)A high-resolution remote sensing image classification model with residual encoder-decoder structure is proposed.The main network adopts an encoding-decoding structure,and the combination of depthwise separable convolution and standard convolution is chosen for the primary feature extraction in the feature encoding stage to reduce the number of parameters in the network,and the standard convolution is replaced by the residual convolution module in the deep feature extraction stage,while the number of convolution kernels is increased step by step to improve the extraction of advanced semantic information.PPM(Pyramid Pooling Module)is added between the last layer of the encoding stage and the first layer of the decoding stage to fuse the global feature information,which can improve the feature characterization ability and enhance the accuracy of image classification;CA(Channel Attention)is added in the decoding stage to perform channel feature recalibration,which can increase the proportion of valid features,thus reducing the cases of misclassification and omission and increasing the accuracy of remote sensing image classification.(2)Experimental analysis is conducted on public datasets to compare with other deep learning fully convolutional neural networks FCN-8s,SegNet,and U-Net,and quantitative analysis of classification results is performed using several accuracy evaluation indexes to explore the effectiveness of this method.Compared with FCN-8s,SegNet and U-NET classification methods,the method in this paper reaches higher accuracy and better classification effect.The experimental results demonstrate the feasibility and robustness of this paper’s method in remote sensing image classification tasks.(3)Classification ablation experiments are designed to verify the effects of each module of CA and PPM on the classification results respectively.The ablation experiment shows that CA module can reduce the misclassification of images and improve the classification accuracy,and PPM module can improve the integrity of classification and obtain more detailed ground object boundary.