| Surface water is an indispensable resource for human survival and social development It is of great significance to monitor surface water quickly and effectively,and to comprehensively grasp the spatial distribution of surface water.With the continuous development of space remote sensing technology and the successive launch of various types of high-resolution satellites,the use of remote sensing technology for water body identification and dynamic monitoring has become an efficient and fast way.At present,there are many methods for remote sensing water body extraction,such as visual interpretation method,water body index method,supervised classification method,inter-spectral relationship method,decision tree classification method,etc.The above methods have their own advantages and disadvantages,but most of them rely on expert experience,and have low capabilities in terms of automation,accuracy,and generalization capabilities.Therefore,the above methods are difficult to carry out rapid and effective remote sensing water body information monitoring on large-area and high-temporal remote sensing data.In recent years,deep learning has made great achievements in speech recognition,image target recognition,image target detection,and natural language processing.Therefore,many scholars have applied it to the field of remote sensing.Typical representatives are remote sensing image scene classification and remote sensing image target detection.,Semantic segmentation of remote sensing images(ground feature classification),etc.Deep learning can automatically extract shallow and deep features in high-dimensional data,automatically fit the classification results,and is suitable for multi-band remote sensing image information extraction.It has become an effective way to extract remote sensing image information.In this paper,for the problem of extracting remote sensing water body information from Gaofen-6 satellite imagery,the effect of FCN-8S and U-Net water body extraction is compared and analyzed,and the U-Net structure is optimized The optimized fully convolutional network VGGUnet uses a combined loss function to alleviate the problem of shallow data feature loss and data and imbalance in the U-Net network.This method alleviates the problems of shallow data loss and data and imbalance in the U-Net network,strengthens the network’s ability to extract image features,improves the sensitivity to detail water bodies in remote sensing,and achieves better water bodies Extraction effect.The method in this paper is applied to the GF-6 satellite wide-area camera surface coverage monitoring and rapid detection of surface coverage change technical subject,and provides support for the subject’s automatic monitoring of water body changes.The main research points and contents of this article are as follows:(1)Analyzed the effective band of water extraction of Gaofen-6 satellite image.In this paper,different types of samples,such as bare soil,vegetation,water bodies,and buildings,are selected from the pre-processed GF-6 WFV image.Then the average surface reflectance of each feature is counted,and the diffrence in surface reflectance of different features in different wave bands is analyzed.This paper analyzes the separability of different features in different band combinations by statistically analyzing the J-M distance between feature samples.The above two methods provide theoretical possibilities for deep learning to extract water bodies;(2)This paper studies and improves the deep network model used to extract the water information of GF-6 satellite image.In order to realize the automatic extraction of the image water body of high-resolution GF-6 high-resolution panchromatic multispectral sensor(PMS)and wide field of view(WFV),three kinds of neural networks,including full convolutional neural network(FCN-8s),U-Net and U-Net optimization(VGGUnet),are constructed for water extraction studies.Based on the water extraction results,the best model is determined as VGGUnet;then a VGGUnet network model based on the combined loss function Focal-Dice-Water loss(FD-Water loss)is proposed.Based on the comparative analysis of VGGUnet and single-band threshold method,multi-band spectrum relationship method,water body index method and other methods,further confirm the effect of VGGUnet network water extraction;(3)Based on the above algorithm model,this paper developed the automatic extraction system of water information of GF-6 remote sensing image。The system has basic functional modules such as remote sensing image opening,zooming,and saving.At the same time,it integrates the remote sensing image water body extraction model proposed in this paper to realize the automatic extraction of remote sensing image water body. |
