Multi-scale Segmentation Optimization Of Urban Green Space In The Typical Area Of Nanjing

As an important part of urban ecosystem,urban green space plays a pivotal role in the construction of eco-city,improve urban air quality and ensure the health of residents.Thus,the study on urban green space has been widely developed since 1980s.With the development of high-resolution remote sensing technology and object-oriented image analysis technology,the traditional urban green space monitoring methods which costs a lot of manpower,low efficiency,and access to low data accuracy are replaced by high-resolution remote sensing image which be data support of accurate urban greening survey.However,the objects produced by image segmentation are not necessarily associated with real objects,for the urban green space in high-resolution remote sensing image has obvious multi-scale characteristics.To establish the relationship between the image object and the real object through optimizing the segmentation results is an urgent problem in the object-oriented urban green space information extraction.This study focuses on the need of National Natural Science Foundation of China"High Resolution Remote Sensing Image Semantic Segmentation with Integrated Priori Scene Knowledge"(Grant No.41601366).The study area is located in the typical area in Nanjing.Firstly,the image representations of urban green space in IKONOS are discussed.Then the study of multi-scale segmentation and multi-scale optimization of urban green space is carried out.The main contents and conclusions are as follows.(1)The image representations of urban green space.The discusses of the functional and geometric characteristics of urban green space shows that,according to the different functions,the urban green space can be divided into dotted green space,green grassland,and ribbon green space.Different forms of green space are interdependent and complex in high-resolution remote sensing images,which are difficult for image segmentation and segmentation optimization.The analysis of the spectral response characteristics and the distribution characteristics of NDVI in the study area with typical samples shows that,NDVI was more discriminative to urban green space than DN.The single-scale segmentation results cannot achieve the optimal expression of urban green space at the same time because of the differences in area and spatial structure complexity.The results of the scale analysis of the green space in the multi-scale segmentation results show that,green land of different particle size is expected to be expressed at its optimal scale with integrating multi-scale segmentation results.(2)The multi-scale segmentation of urban green space.A multi-scale segmentation algorithm based on region merging is proposed.Specifically,the initial over-segmentation results are obtained from region growing.Based on the initial over-segmentation results,the area adjacency graphs are generated,the merger criteria are set according to the geometric characteristics of urban green space,the consolidation strategy based on priority queues is developed,the scale parameter sequence is set.Then,perform a region consolidation to obtain a uniform multi-scale segmentation result and use the object scale tree to establish the contextual relationship between the surface objects in different scale segmentation results to provide the steady for the subsequent optimization across multiple scales.(3)The multi-scale optimization of urban green space.Design a multi-scale optimization algorithm for urban green space.Firstly,the global optimal segmentation results are selected from the multi-scale segmentation results using the optimization index.Secondly,the objects in the global optimal segmentation result are divided into the under-segmentation and fine-segmentation.The under-segmentation is optimized by the object scale tree.Thirdly,the final result is obtained by synthesizing the under-segmentation with optimizing and fine-segmentation.The validity of the multi-scale optimization algorithm is analyzed qualitatively and quantitatively,which shows that the multi-scale optimization algorithm can effectively identify the fine-segmentation and under-segmentation,and separate the green objects in the under-divided region,thus reducing the segmentation error in the global optimal segmentation result.The result of evaluating the accuracy of multi-scale segmentation results and multi-scale optimization results using supervisory evaluation indicators show that,multi-scale optimization results are more accurate while achieving the optimal expression of different particle size green objects.In view of multi-scale characteristics of urban green space in high-resolution remote sensing images,this paper proposed multi-scale segmentation algorithm and multi-scale optimization algorithm for urban green space,combining the spectral response,geometry,and spatial context relations of urban green space,which provides a new idea for the study of multi-scale segmentation optimization of remote sensing image and has certain innovation in theory and application.