Study Of Multi-spectral Remote Sensing Imagery Classification And Urban Expansion Based On Machine Learning

With the rapid development of space technology, remote sensing imagery has beenwidely used in many fields in21st century. In recent years, people are facing more andmore pressure in the man-environment relationship because of the ever-increasingurbanization in China, followed by the increasing pressure of land-use monitoring. Landuse in city is the illustrated results on earth surface induced by the interaction of theenvironment and human being activity. Different types of land-use will be resulted fromdifferent backgrounds of social culture and conditions of politics-economy andenvironment. The traditional statistics of manual field require a lot of manpower andresources, remote sensing imagery is emerging as a new method for monitoring dynamicchanges of land-use, simulation and prediction of the situation of urban expansion, whichis the trend of intelligent processing of remote sensing imagery. Many scholars have alsodone a lot of research work in this field, which provides a solid foundation for the study.At present, the dynamic monitor of land-use and the simulation of urban expansionoften involve the problem of remote sensing image classification. With the recent rapiddevelopment of the aerospace technology, the accuracy of remote sensing imagery hasbeen significantly improved, and the amount of data is also growing rapidly. The dataset of remote sensing image contains the following features: large amount of data, highambiguity, small ground-sample data, high redundancy of data and low signal-to-noiserate. Therefore, in the remote classification of remote sensing imagery, it is an importantresearch direction to find useful information from massive data set by using machinelearning methods. Manifold learning is one of machine learning methods, whichdemonstrates significant advantages in finding the internal structure of complex datasets. Scholars have used manifold learning method in remote sensing imageclassification. The information of cities, rivers, and farmland extracted in remotesensing images affects the simulation results of the urban spatial expansion. Therefore,this paper improved the classification accuracy of remote sensing imagery, and used theinformation extracted from the classification results for urban spatial expansionsimulation.This paper mainly includes two aspects: firstly, the paper introduced the methods offeature extraction about multi-spectral data, which are based on semi-supervisedmanifold learning; secondly, the information from the classification results is applied to urban expansion simulation and prediction. To improve the accuracy of remote sensingimage classification by incorporating labeled and unlabeled samples, a new manifoldlearning method, called semi-supervised manifold discriminant embedding (SSMDE), isproposed in this paper. This method fused the class information by prior knowledge ofthe user into the semi-supervised manifold learning. This method constructs tworelational graphs through data point labels: A within-class graph and a between-classgraph are used to encode the class relation information that indicated in the labeled datapoints; two weighted matrices are computed based on the two graphs. Then, we utilizethe labeled and unlabeled data points to construct the total scatter matrix to describe theinformation of all data points. Finally, the projection matrix of SSMDE can be obtainedby solving an optimization problem. SSMDE method can not only take into account thediscriminant information of labeled data, but also preserve the global structureinformation of all data points. The proposed algorithm achieves a semi-supervisedlearning ability by considering the label data and unlabeled data information. Theexperimental results on both synthetic and remote sensing images show that theproposed method can improve the classification accuracy rate and reduces the cost forachieving the labeled sample.The information of the urban areas, rivers and roads is extracted from theclassification result of the remote sensing image. Then, the information is applied tourban expansion simulation and prediction. At first, the paper analyses the mathematicalmodel of urban simulation. Then, this paper selected Chongqing district as the studyarea, which has a rapid rate of urbanization and Eco-environmental pressure. Thefundamental data of this study is the remote imagery and urban planning imagery ofChongqing city in1988,1993,2000and2007.After using RS and GIS for dataprocessing, the model calibration is applied to obtain best parameters for the SLEUTHmodel. Then, the SLEUTH model is used for the simulation of the urban growth overthe past20years. On the basis, the SLEUTH model is applied to predict the situation ofurban expansion of the region for the next10years. Finally, the paper put forwardreasonable proposals and measures based on analyzing the simulation results.