Land Classification Of Coastal Zone Based On Sentinel-1A

Based on the research of the “Guangdong Province Coastline Dynamic Monitoring Research Based on Multi-Polarized Spaceborne Synthetic Aperture Radar”,we use Seg Net doing some research relative to land classification method of coastal zone in the Pearl River Delta(the key area is Nansha District).Seg Net network is an end-to-end semantic segmentation model,which uses the structure of Encoder-Decoder.It achieves accuracy of 73.2% on remote sensing image dataset,which is better than Deconv Net and other network models[1].The experiment used the data source as Sentinel-1A satellite imagery,which is provided by ESA freely and is suitable for a wide range of surface monitoring.Therefore,the experiment combined Seg Net and U-Net network and Sentinel-1A remote sensing image data to complete the research on the land classification in the coastal zone of Guangdong Province.In order to improve the accuracy,we use Terra SAR-X image data with higher resolution for correction,and the experimental results were verified by field investigation.At the same time,in order to prove the veracity of the Seg Net and U-Net network for remote sensing image classification,the minimum distance method,maximum likelihood estimation,kmeans and other classification algorithms are used for comparison.Through the comprehensive analysis of the experiment,the following conclusions can be drawn:(1)Sentinel's single satellite(including A and B) has a revisit period of only 12 days.Using Sentinel-1A image data as the main experimental data,more datas can be obtained.At the same time,the image data is provided by ESA freely,so it can also save economic costs.The usage of Terra SAR-X image data for correction can compensate for the error caused by the low accuracy(compared with Terra SAR image data) of Sentinel-1A image data.(2)Traditional clustering algorithms such as kmeans do not need to label the data,and directly use the internal characteristics of the samples to classify,which can greatly reduce the workload,but the accuracy is low,and the classification result is rough.In recent years,the academic community has generally begun to apply the met hod of deep learning to the classification of remote sensing images.For example,so meone uses FCN,Seg Net,Unet and other networks for land classification.These met hods use multi-layer convolution,which can extract features of different scales of images and enrich the feature,which is beneficial to improve the accuracy of classification.Secondly,due to the popularity of hardware platform GPU(for convolution operations),network training is easier and more efficient,which can improve the speed of the algorithm,and can achieve deeper network training and improve classification a ccuracy.(3)Through the investigation activities with the mapping company,the real land types of different locations in Nansha District are obtained,which can verify the reliability of our method.It can be seen from the mapping results that the model can better reflect the land type in the study area,and has certain guiding significance for the utilization and planning of local natural resources,which has a good development prospect.