Land Cover Classification Of Wetland Based On Multi-source Remote Sensing

Wetland is a kind of important natural ecological system,which has the strong capacity of natural productivity and rich ecological resources.Wetland ecosystem provides vital survival environment for human beings and all the living things in the world.Sustainable development of wetland plays an important role in keeping ecological balance in a region.Wetland belongs to the transitional zone of aquatic ecosystem and terrestrial ecosystem.It is easy to be affected by the environmental changes.Some worse situation may lead to the degradation and disappeared of wetland.It is necessary to protect and manage wetland ecological environment by efficient and reasonable measures.The basic work of wetland environment monitoring is the classification of wetland land-cover types.It may obtain the mapping of wetland land use and provide basic datasets for the following work.Therefore,it is of great significance to search a good technology for wetland land-cover classification to monitor and manage the wetland environment.In the study of wetland land-cover classification,the remote sensing imaging characteristics and spatial relationship between different types are more complicated,especially the wetland subclasses.Some wetland types have very similar features and the boundary of objects is ambiguous.The situation makes classification work some difficult.The classification results directly affect the quality of following work.So it is the key step to extract and select feature variables that can reflect the different objects and make efficient models for classification,particularly for the types which are high similarity.Aiming at solving the main problems about the research of wetland land-cover classification,taking care of the latest technologies in the field of image processing,data mining and pattern recognition,this thesis made a deep study on the issue of the strategy of wetland land-cover classification,including the acquisition of remote sensing data source,feature extraction,variables selection and classification methods.The central area of the Wuyuer river wetland nature reserve in Heilongjiang province was chosen as the study area.The full polarized Radarsat-2 radar image and Sentinel-2A multispectral data were used as the main data sources.The main work and the research results in this thesis can be summarized as follows:1.Two kinds of remote sensing data resources were used for the wetland land-cover classification in this research,including the multi-spectral image obtained from the newly launched Sentinel-2A remote sensing satellite and the radar image obtained from the Radarsat-2 remote sensing satellite.In order to investigate the applicability of using Sentinel-2A multispectral data and Radarsat-2 radar data in wetland land-cover classification,diffierent associations of data sources were implemented.The results demonstrated that both of the multispectral data and radar data can be effectively used to distinguish different types,but the accuracy results were low and some categories were confused.However,the overall accuracy was remarkably improved when the muli-source remote sensing data was used because of the rich information.Some similar objects such as the marsh wetland with defferent types of vegetation can be better distinguished.2.For Radarsat-2 radar data,in order to investigate the application ability of different polarization decomposition methods in wetland land-cover mapping,a variety of classical polarimetric decomposition methods were used to extract polarimetric parameters for classification.For Sentinel-2A multispectral data,the spectral reflectance of different objects was used for classification.Because of the improved band setting of Sentinel satellite,new spectral variables were obtained.The extracted spectral variables and radar polarization variables were synthesized to form a high-dimensional dataset with 54 variables,which were used for the following steps.These variables were selected by a proposed feature selection method,and the optimized variable dataset can be used to distinguished different types of wetland land-cover and play a positive role on classification results.3.In order to obtain the variables which can reflect the different characteristics of wetland land-cover types,this study proposed a new feature selection method,which combined the most effective feature selection method called Relief F and the promising method called random forest algorithm.The variables selected through this method can effectively reflect the physical properties of the different land-cover types and help improve the classification results.To verify the effectiveness of the proposed feature selection method,the information reflected by the variables will be compared to the actual property characteristics.At the same time,the optimized variables dataset was applied to different classifiers to detect the classification accuracy of wetland land-cover types.The results showed that the selected variables by proposed feature selection method can effectively describe the characteristics of different land-cover types,and they also help for the classification.4.The shallow machine learning methods were used to distinguish the wetland land-cover types.The selected classification methods were widely recognized in the field of machine learning,including random forest algorithm,support vector machine algorithm,extreme learning machine algorithm and ensemble learning algorithm.The experimental results showed that ensemble learning algorithm obtained the best classification results,followed by random forest algorithm,support vector machine algorithm and extreme learning machine algorithm.The ensemble learning algorithm took the advantages of different classifiers and obtained best classification results with an overall accuracy above 85%.5.The deep machine learning methods were used to distinguish the wetland land-cover types.The selected deep model was a whole convolution neural network structure called Unet,which is promising and widely used in the field of deep learning.In order to train the deep model easier,a residual structure was added in the basic model.The optimized model called Res Unet.Unet and Res Unet model were both used for classification.The results showed that the overall accuracy was 94.45% and 94.53% by using Unet and Res Unet,which was much better than the shallow machine learning methods.The deep learning models had strong ability of extracting data features by itself.A lot of time used for extracting auxiliary variables can be saved and the features extracted from deep learning models played the best role on classification and visualization.