| Due to the acceleration of urbanization and industrialization and the frequent occurrence of natural disasters,China's natural resource security has been seriously threatened under the dual influence of human factors and natural factors.Therefore,timely understanding of the changes in farmland is of great significance to China's agricultural construction and sustainable development of natural resources.Farmland change detection is a process of determining farmland changes based on farmland information at different times in the same region of interest.With the rapid development of remote sensing technology and image processing technology,it is very convenient to use the high resolution remote sensing image to detect the change of the ROI.But the use of traditional pixel-based change detection method has been unable to apply to high-resolution remote sensing image with rich information features.Therefore,the use of object-oriented change detection method has become a key issue for scholars at home and abroad.This article selects Sheli Town of Da'an City as the study area and uses the GF-2 high-resolution remote sensing image of September 2015 and August 2017 as the change detection experimental datum.Using object-oriented classification method to complete the extraction of remote sensing image farmland information at two times and evaluated the accuracy of the classification results.Then used the object-oriented change detection method to achieve change detection of farmland information.The main research results of this paper are as follows:(1)The differences between boundary-based segmentation algorithm and regionbased segmentation algorithm were studied in detail.And analyzed that the multi-scale segmentation of region-based segmentation algorithm was superior when simultaneously extracted multiple target feature types from both objective and subjective aspects,then summed up the choice of multi-scale segmentation parameters.Different segmentation parameters were used for multi-scale segmentation in different image object layers.Using the trial-and-error method to find the optimal segmentation scale for the first layer of vegetation and non-vegetation was 90,the shape factor was 0.2 and the compactness factor was 0.5.The optimal segmentation scale for the second layer of dry land,paddy field,and other vegetation was 150,the shape factor was 0.2,and the compactness factor was 0.5.(2)Comprehensively utilized the spectral features,shape features and custom features of high-resolution remote sensing images to extract information on the target object type.After establishing the classification level of the image object,the rules for extracting the target features were established through the description and combination of the characteristics of the target features.The multi-scale segmented image object was the basic unit of classification.The combination of threshold classification and nearest neighbor classification enabled the extraction of farmland information in the study area.(3)The visual interpretation results of GF-2 remote sensing images in 2015 and 2017 were used as reference images to evaluate the classification results.The visual interpretation of the remote sensing image in the study area at two times was conducted and the farmland information was artificially read.Using confusion matrix to compare the automatic classification results of two times with the visual interpretation results and found that the overall accuracy was 91.6% and the Kappa coefficient was 84.8% of 2015.The overall accuracy was 90.8% and the Kappa coefficient was 84.3% of 2017.(4)Discussed three kinds of remote sensing image change detection methods and the object-oriented change detection method was used for the experiment.The changes of farmland information in the study area at two times were detected by using the algorithms of copy map,synchronize map,convert to sub-objects and copy image object level.The results show that this method is practicable and effective. |
PDF Full Text Request