Remote Sensing Image Semantic Segmentation And Landscape Pattern Change Analysis Based On Improved FCN

Landscape is a geographical entity composed of a series of different land units or ecosystems,which has obvious visual characteristics.Urban landscape pattern change is closely related to land urbanization,which has become an effective method to study the process of land urbanization.With the development of remote sensing technology,the high resolution remote sensing image is increasingly rich,which makes remote sensing image become an important data source of landscape pattern change analysis,and remote sensing image landscape classification has become a very important step of landscape change analysis.However,the original remote sensing image data can't be directly used due to the characteristics of complex information,more noise and so on.The next step can be started only after the information required for research is extracted manually Analysis.The main purpose of this paper is to use more efficient and accurate methods to classify the landscape of remote sensing images,and then analyze the change of landscape pattern visually according to the classification results.Although the traditional classification method of remote sensing image performs well in low resolution and single scene,there are some problems in the classification of highresolution remote sensing image with different shooting time and lighting conditions.In order to improve the accuracy of remote sensing image classification,this paper constructs a FCN based deep learning model with pixels as basic units to extract highlevel abstract features of remote sensing image,and finally uses the FCN network to output landscape classification map for landscape pattern change analysis.The main contents of the paper are as follows:(1)Obtaining remote sensing images with different landscape features and using Arc GIS for image interpretation,and then establish the data set for semantic segmentation training and output through image clipping and data enhancement.(2)Using the improved network model based on FCN as the basic framework of remote sensing image semantic segmentation,aiming at the problem that the general FCN model can't take into account both the global features and the detailed features at the same time in feature extraction,based on the U-Net network,a new encoder method is added to ensure that the perception field is unlimited,and a stronger feature extraction network is constructed to lay a foundation for the subsequent landscape pattern change analysis Solid foundation.(3)Using the deep learning semantic segmentation network,we train the remote sensing image classification model for CCF open data set and the data set made in this paper.Compared with SVM and other semantic segmentation networks,the improved FCN network achieves 91% of the test average pixel accuracy(MPA)in the CCF data set.Compared with SVM and other semantic segmentation networks,this method has better performance in the details of vegetation,buildings and the integrity of roads.The improved FCN network achieves 85% of the test average pixel accuracy(MPA)in the data set of this paper,which proves the availability of this data set.(4)Applying the trained model to the analysis of the landscape pattern change in the Xindu District of Chengdu.Based on the segmented landscape classification maps of different years,using the landscape pattern related methods to analyze the changes of the landscape pattern in the study area on the time and space scale Evaluation of the stability of the landscape pattern,and good results are achieved.