Research On Multi-feature Remote Sensing Image Water Body Identification In Yili River Basin

The continuous development and progress of aerospace technology and remote sensing sensors have brought new opportunities and challenges to the effective management and application of remote sensing big data.Remote sensing image water identification is a kind of processing and application of massive satellite image resources.The traditional method is time-consuming and laborious to obtain better precision,while the progress of artificial intelligence technology promotes the progress of remote sensing image processing and makes the surface water resource identification more efficient.Deep learning is applied to the research of water information extraction in the yili river basin of xinjiang from satellite images,so as to provide data support for water resource information monitoring and realize timely water ecological analysis and risk warning.This is of practical significance to the study of the environmental evolution and sustainable development of water resources protection,development and rational utilization in the yili river basin.This paper aims to study the application of deep learning technology in the field of artificial intelligence in the water identification of remote sensing image.Aiming at the needs of the field of remote sensing,this paper improves and optimizes the deep learning model and structure to build an intelligent water identification model suitable for remote sensing image.The main work of this thesis includes:(1)Through research and analysis,the stability of the model is improved by using multiple extended features and enhancing the image.Because of the high dimension of remote sensing image,it is difficult for DBN model to measure the full dimension of high dimension image.In view of the disadvantages,a feature expansion method is designed,which is simple and effective compared with the complex model modification.(2)In the traditional multi-stage method of water body extraction of multispectral remote sensing image,there are bottlenecks in feature construction and selection.In view of this phenomenon,an end to end neural network model for remote sensing image recognition based on Bidirectional Long Short Term Memory Network(BiLSTM)and the Modified Convolutional Neural Network(M-CNN)is proposed.So by improving the model to reduce labor.(3)In order to cope with the huge amounts of unlabelled remote sensing data,using the characteristics that the auto-encoder could unsupervised learning the correlation representation of input data,an unsupervised remote sensing image segmentation method which combined with Atrous Convolution Auto-encoder and BiLSTM Auto-encoder(Bidirectional Long and Short Time Memory Auto-encoder)network was proposed.This method can proactively obtain the extraction information of target resources from remote sensing images without any prior knowledge,which improves the efficiency and automation degree of remote sensing image information extraction.The data sources are mainly Landsat satellite images.By making full use of the advantages of the deep learning model,deep features are excavated through training,spectral features and multi-scale spatial texture features of remote sensing images are studied comprehensively with supervise or unsupervised learning,so as to realize automatic learning and extraction of features and obtain objective recognition results.The experiment results show that the classification accuracy of the proposed model for image water recognition reaches 99.01%,and the average accuracy of clustering reaches 83.25%,which further reduces the artificial intervention and resource consumption,and verifies the effectiveness and applicability of the model.