Research On Land Cover Classification Method Based On Satellite Radar Remote Sensing Data

Land cover classification has an essential impact on land use,ecosystem degradation,environmental and water availability that have been attached great importance.Object-oriented image classification method is one of the cutting edge image processing technologies attracting significant attention in remote sensing and Synthetic Aperture Radar(SAR)possess a high potential for classification of earth surface.The research of land cover classification based on SAR imaging is the frontier and hotspot of remote sensing image processing.This thesis presents a novel Polarimetric SAR databased object-oriented land cover classification approach,using a new multi-channel Watershed region-merging image segmentation method and an ensemble of Extreme Learning Machine(Ensemble-ELM)image object classifier.Sentinel-1A SAR data of three different areas(Bengbu,Singapore,Xilinhot)are used to evaluate the proposed land cover classification method.Multi-channel segmentation algorithm is specially designed for SAR image segmentation task which contains two steps.Firstly,the Watershed algorithm is applied to the input image and obtain the initial segmentation objects.Then a region merging technique is used to merge these initial objects in a hierarchy way.This method not only overcomes the over segmentation issue caused by the traditional watershed segmentation algorithm in practical scenarios,but also can keep the edge of the object in the SAR image.The advantage of the multi-channel segmentation algorithm is verified by comparing it with the well-known Simple Linear Iterative Clustering(SLIC)superpixel algorithm using an image segmentation evaluation based on the homogeneity within and the heterogeneity between the segmentation objects.Following the multi-channel segmentation,an ensemble ELM learner is developed to perform the classification of image objects.Besides,four other classification methods: single ELM,Random Forest,KNN and SVM,are set to compare the performance of the proposed classifier.Quantitative analysis with respect to ground truth information available for the test site shows that the Ensemble-ELM learner achieves an accuracy of 92.52% for Bengbu area,95.94% for Singapore area and 88.28% for Xilinhot area in classifying the considered classes which outperforms the other four algorithms.The proposed method is proven to be a novel and competitive tool for SAR-based land cover classification task in real application.