Research On Water Extraction Of GF-1 Satellite Data Based On Deep Learning

The successful launch of domestic Gao Fen(GF)series satellites opened a new period of high-resolution earth observation in China,promoted the improvement of high-resolution earth observation satellites and their application level,and made China's earth observation level reach a new level,and the application of land resources monitoring,crop yield estimation,disaster monitoring and other aspects obtained more accurate data sources.Among them,water resource management is an extremely important aspect in remote sensing image application.Higher resolution remote sensing data can help to extract more accurate water information,which is conducive to better practice of water resource monitoring and disaster management.At present,there are many algorithms for water extraction from remote sensing images,but traditional water extraction methods have their own limitations.In addition,some traditional water extraction algorithms are not applicable to GF satellite data because the band distribution and texture information of GF satellite data are different from other remote sensing data sources.To solve the above problems,this study integrated the spectral features of remote sensing image into the algorithm of deep convolutional neural network(CNN),and proposed a convolutional neural network for water extraction of GF-1 satellite data.Firstly,we proposed a convolutional neural network based on Xception to identify the water bodies in Poyang Lake area,and compared the results with the Normalized Difference Water Index(NDWI);minimum distance,maximum likelihood,mahalanobis distance from supervised classification and K-means,ISODATA from unsupervised classification.The results showed that the algorithm of the convolutional neural network is far superior to the other methods.Secondly,we made an improvement on the original algorithm and proposed a deep convolutional neural network based on multidimensional densely connected neural networks(Dense Net).After obtaining the optimal number of network layers through many tests,the feature extraction of the model is carried out through the series of densely connected convolution modules,and the full convolutional neural network is constructed by transconvolution to achieve the goal of image semantic segmentation.On this basis,the feature utilization efficiency of the model is enhanced and the image recognition accuracy isimproved by adding the idea of multi-scale fusion.The results showed that compared with the convolutional neural network based on Xception,the water recognition accuracy was improved significantly.In addition,in order to verify whether the improved model is the optimal model in the convolutional neural networks,Dense Net's results were compared with classical convolutional neural networks(CNNs): Res Net,VGG,Seg Net and Deep Lab v3+,and compared with NDWI.We selected the training time and the losscurve of the training process to evaluate the training efficiency of the five deep networks.We selected four indexes of precision(P),recall(R),F1 score(F1)and mean Intersection over Union(m Io U)to evaluate the performance of all methods.The results showed that all deep convolutional neural networks are better than those of NDWI.Although the calculation of NDWI is simple and convenient and the time consumption is almost negligible compared with the training time of neural networks,but the recognition accuracy of NDWI is low,and the optimal threshold changes with time and place,and the generalization is far less than that of convolutional neural network.Therefore,in general,the five convolutional neural network methods are superior to NDWI.Among the five convolutional neural networks,Deep Lab v3+ algorithm has the shortest training time,but it is inferior to Dense Net in terms of model convergence speed and water recognition accuracy.Generally speaking,Dense Net has the best training efficiency and recognition accuracy.Among them,VGG and Seg Net performed poorly in distinguishing water from clouds and mountain shadows;Res Net could distinguish clouds and water bodies well,but it could not distinguish mountain shadows well.Deep Lab v3+ will not confuse the mountain shadow,but the ability to distinguish clouds is poor,and the recognition of the water edge is rough.Dense Net also misjudged some clouds in the process of identifying water bodies,but the misjudged area was far smaller than other models and the recognition effect of mountain shadows was better.Therefore,in general,for the water extraction task of GF-1satellite,the comparison results of multiple convolutional neural networks showed that Dense Net is the optimal model.Finally,we applied the selected optimal model to multi-temporal GF-1 remote sensing images to extract water information in Poyang Lake area.We analyzed the water area changes in the Poyang Lake area during wet and dry season from 2014 to 2018.Among them,the flood disaster events in the summer of 2016 were selected,and the flooded area of this disaster was analyzed in detail according to the land cover the research area.The results showed that deep convolutional neural network has a strong application prospect in the water extraction of high-resolution satellite,which can provide data support for water resource monitoring and protection,pre-warning and post-disaster reconstruction of flood disaster in the research area.