Fusion Of Pixel And Object For High-resolution Remote Sensing Image Building Change Detection

Change detection is the process of identifying changes after multiple observations of a feature or phenomenon.Buildings and roads account for the largest proportion of highresolution remote sensing images in urban areas,and buildings are listed as the most important urban elements.The pixel-level building change detection result is prone to "salt and pepper phenomenon",and the post-classification comparison method classifies two multi-temporal images separately and then performs change detection,which is likely to cause error accumulation.The selection of the segmentation parameters in the object-level change detection method is the most important.If the segmentation scale is too small,the segmentation will be fragmented,that is,the same feature can be mistakenly divided into multiple objects to cause false detection;and the excessive segmentation scale may result in an area comparison.Small buildings with spectral features similar to those of adjacent objects are merged into adjacent objects,and the features may be split and lead to missed detection.High-resolution remote sensing imagery is complicated by building structure and spectral diversity.This paper uses two sets of data sources to verify the effectiveness of the method: World View2/3 high-resolution remote sensing imagery of Changsha City,Hunan Province and Xinyu City,Jiangxi Province high-resolution remote sensing image of airborne visible light.The experimental results were evaluated by the accuracy of the drawing,the user precision,the Kappa coefficient,etc.,and compared with the experimental results of the pixel level and the object level.The Kappa coefficient is improved by 0.3697 and 0.2892 in the World View2/3 high-resolution remote sensing imagery factory area compared to the pixel level and target level,and the residential area is increased by 0.2028 and 0.6104;the airborne illuminable image factory area is increased by 0.0385 and 0.3769.The residential area has increased by 0.0798 and 0.3137.The experiment proves that the fusion object level and pixel level change detection method is beneficial to improve the building change detection accuracy and eliminate the “salt and salt phenomenon”.The building extraction boundary is clear and the shape is complete.The main research work and results of this paper can be summarized as follows:1.A reasonable multi-dimensional feature vector construction method is explored for images,and spectral correlation factors(such as normalized plant index NDVI and morphological building index MBI)are added.NDVI facilitates the distinction between vegetation and buildings,effectively avoiding Plant-related changes are erroneously doped in plant changes;MBI is used to distinguish between buildings and other impervious layers,and can somewhat reduce the likelihood that changes in other impervious layers such as roads are mistakenly detected as buildings.In order to prove the effectiveness of the proposed algorithm,this paper uses the recall rate,accuracy and overall accuracy F-Measure to evaluate the feature effectiveness.The experimental results show that the accuracy of the image is significantly improved after adding spectral factor features.2.After constructing the front and back phase image feature vectors,select the building changes,non-building changes and unchanged samples,and then use the random forest classifier machine learning algorithm to change the building extraction to effectively avoid the error transmission,and calculate the random The OOB average drop accuracy and the Gini index in the forest classifier measure the feature importance distribution.3.In this paper,the method of iterative optimization to determine the segmentation parameters is used to automatically determine the optimal segmentation parameters,which effectively solves the problem that the segmentation boundary and the segmentation scale of the object-level building change detection are difficult to determine.4.Through the mathematical statistical method,the object-level segmentation result is used to constrain the change building extracted by the pixel level,and the threshold range is determined by the normal distribution,and the value is intermittently every 0.5 to determine the most suitable for different image experimental areas.Good constraint threshold.