随着生活水平提高，人们更加关心生活品质，对绿地的需求逐渐增加，传统的基于俯瞰视角的绿地评价指标已不能充分衡量人们对绿色空间的需求与感知。快速提取城市绿地类型，确定需进行绿色改造的区域，为城市规划提供信息参考与指导，对提高城市居民生活品质具有重要意义。
本文以北京市五环内区域为研究区，基于SPOT6影像利用海洋捕食者算法（Marine Predators Algorithm, MPA）优化机器学习模型提取城市绿地类型。根据街景图像计算绿视率（Green View Index, GVI），研究GVI的影响因素。分析比较250m×250m格网尺度下两种视角对应的指标——绿地覆盖率与GVI的空间分布，提出经济可行的改善GVI方案。本文的主要工作及成果如下：
1.构建MPA优化的机器学习模型，实现快速提取城市绿地类型。为提取公园绿地与附属绿地，根据SPOT6影像构建特征空间并建立城市绿地类型数据集，利用MPA建立了三种优化后的模型MPA-KNN、MPA-SVM、MPA-RF，根据验证集精度选择最佳超参数组合，以AUC值、测试集精度等指标比较三种优化后模型。结果表明三者中MPA-RF模型表现最佳，AUC值与测试集精度分别为0.983与89.93%。使用MPA-RF基于SPOT6影像提取北京五环内城市绿地类型，识别北京市东城区、西城区大型公园准确率可达80.8%。MPA优化后的机器学习模型在城市绿地类型分类中展示了出色的分类能力，对城市绿地类型的精确识别具有重要意义。
2.研究通过街景图像提取GVI方法，并分析GVI影响因素及分布特征。首先，对已有的街景图像预处理流程进行了细化，保证所使用的街景图像可代表人在实际中对绿色空间的感受。其次，利用图像分割算法计算街景图像内各地物占比。从街景图像内部特征及街景采样点地理环境方面，分析GVI的影响因素，具体包括：（1）总体上，各植被类型贡献率显示树木对GVI的贡献率达94.64%，灌木和草累计贡献率为4.55%，人对绿色空间的感知主要源于树木，灌木和草具有调节作用。（2）GVI与视野内天空、房屋建筑占比具有显著的负相关性，与路面呈较弱的正相关，与人与车占比无关。（3）提出衡量沿路与垂直路方向的GVI指标——绿视率视向比(The viewing direction rate of GVI, VDRGVI)，当GVI较低时GVI主要为沿路方向所得；当GVI较高时，VDRGVI接近于0，即沿路与垂直路方向GVI均为高值，且垂直方向GVI多高于沿路方向。（4）随着缓冲区面积增大，绿地覆盖率与GVI的相关系数先增后减，当缓冲区范围为250m×100m时（沿路×垂直路），绿地覆盖率与GVI正相关系数最大。（5）在北京市五环范围内，GVI与建筑密度呈负相关，在四环至五环区域内负相关性最强，而在二环至三环范围内，GVI与建筑密度无相关性。
3.通过多视角绿色空间分布比较研究，提出绿色空间建设区建议。在精细空间（250m×250m格网）尺度下，根据绿地覆盖率与GVI间的等级对应关系，将分析格网分为三类绿色空间建设区：绿色舒适区、绿色重点关注区、绿色待改进区，分别占65.27%、9.00%、25.73%。通过分析建议绿色待改进区中的典型街区，在GVI低的地区通过改善植被结构、适当增加相对矮小植株等方式提高人视野所及范围的绿色。
综上，仅以俯瞰视角建立绿地评价的指标已不能充分反映人们对绿色的感知，应加入GVI等代表人的观察视角的绿地评价指标，建立综合的绿色评价系统，为城市规划与可持续发展提供合理依据。

With the improvement of living standards, people are more concerned about the quality of life, and the demand for green space is gradually increasing. The traditional green space evaluation index based on the overlooking perspective can not fully measure people's demand and perception of green space. Quickly extracting the types of urban green space and determining the areas that need to be green transformation can provide information reference and guidance for urban planning, and is of great significance to improve the quality of life of urban residents.
In this paper, the area within the fifth ring road of Beijing is taken as the study area, and the Marine Predators Algorithm (MPA) is used to optimize the machine learning model based on SPOT6 images to extract the types of urban green space. The Green View Index (GVI) is calculated from street view images, and the influencing factors of GVI are studied. The spatial distribution of green space coverage and GVI corresponding to the two views under the 250m×250m grid scale is analyzed and compared. An economical and feasible scheme to improve GVI is proposed. The main work and results of this paper are as follows:
1. A machine learning model optimized by MPA is constructed to realize the rapid extraction of urban green space types. To extract the park green space and attached green space, three optimized models (MPA-KNN, MPA-SVM, MPA-RF) are established by MPA. The best combinations of hyperparameters are selected according to the accuracy of the validation set. and the three optimized models are compared by AUC value, test set accuracy and other indicators. The results show that MPA-RF performs the best among the three, with AUC value and test set accuracy of 0.983 and 89.93% respectively. MPA-RF is used to extract the types of urban green space in the fifth Ring road of Beijing based on SPOT6 images, and the accuracy of recognizing large parks in Dongcheng and Xicheng district of Beijing can reach 80.8%. The machine learning model optimized by MPA shows excellent classification ability in the classification of urban green space types, which is of great significance for the accurate identification of urban green space types.
2. The research extracts GVI from street view images and analyzes the GVI influencing factors and distribution characteristics. Firstly, the existing street view image pre-processing process is refined to ensure that the street view images used can represent the human feelings of green space in reality. Secondly, the image segmentation algorithm is used to calculate the percentage of objects within the street view images. The influencing factors of GVI are analyzed in terms of the internal characteristics of street view images and the geographical environment of street view sampling points, specifically: (1) Overall, the contribution rate of each vegetation type shows that trees contribute 94.64% to GVI, and the cumulative contribution rate of plants and grasses is 4.55%, and the human perception of green space mainly originates from trees, and plants and grasses have a moderating effect. (2) GVI has a significant negative correlation with the percentage of sky and buildings in the field of view, a weak positive correlation with road surface, and no correlation with the percentage of people and vehicles. (3) The viewing direction rate of GVI (VDRGVI), which is a measure of GVI along and perpendicular to the road, is proposed. When the GVI is low, the GVI is mainly obtained along the road; when the GVI is high, the VDRGVI is close to zero, and the GVI is high in both directions, and the vertical GVI may be higher. (4) With the increase of buffer area, the correlation coefficient between green space coverage rate and GVI increases first and then decreases. When the buffer area is 250m×100m (along the road × perpendicular to the road), the positive correlation coefficient between green space coverage rate and GVI is the largest. (5) Within the fifth ring road of Beijing, GVI is negatively correlated with building density, and the negative correlation is strongest in the fourth ring road to the fifth ring road, but there is no correlation between GVI and building density in the second ring road to the third ring road.
3. Through comparative study of green space distribution from multiple perspectives, the suggested zones for green space construction are proposed. Under the scale of 250m×250m grid, based on the rank correspondence between green space coverage and GVI, the analyzed grids are divided into three types: green comfort zone, green focus zone, and green to be improved zone, accounting for 65.27%, 9.00%, and 25.73%, respectively. By analyzing the typical neighborhoods in the green to be improved areas, the greenness of the areas with low GVI can be improved by improving the vegetation structure and appropriately adding relatively dwarf plants within the reach of human view. In summary, the indicators of green space evaluation established only from the overlook perspective can no longer fully reflect people's perception of green, and green space evaluation indicators representing the human observation perspective such as GVI should be added to establish a comprehensive green evaluation system to provide a reasonable basis for urban planning and sustainable development.