Introduction
In recent decades, the growth of cities and the development of urbanization have intensified in different countries, including Iran. The importance of the city and urban planning from the point of view of improving the environment in the context of a healthy city has been considered more than ever. In particular, urban green spaces play a role in environmental and ecological functions such as controlling air pollution, regulating microclimate, reducing heat island effects, purifying water quality, and increasing urban resilience in response to environmental and climate changes.
In the past, land cover maps were done using ground mapping and spending a lot of time and money, but today, despite satellite images and drones, it is possible to prepare these maps with high accuracy. Based on recent advances such as high spatial resolution imagery, remote sensing provides a valuable set of tools that can minimize the need for field surveys even in highly heterogeneous and complex urban environments. The aim of the current research is to use images with high spatial resolution in the preparation of urban green space level maps. The studies that were reviewed in the resource review section mostly used Landsat and Sentinel satellite images with spatial resolution of 30 and 10 meters to separate urban land cover, while considering the small area of urban green space, especially in desert cities like the study area, it causes an error. Therefore, in this study, in order to improve previous studies, images with high spatial resolution were used to prepare urban green space maps. The results of this study will be useful for planners and policy makers in the field of metropolitan decisions and green spaces. It can also be used in locating and building new green spaces.
Methodology
In the current research, Landsat 8 satellite images and high spatial resolution images prepared by airplanes from the study area were used. The resolution of Landsat 8 spectral images is 30 m and images with high spatial resolution is 10 cm. Landsat 8, which was launched on February 11, 2013, has 2 sensors, OLI and TIRS, and performs imaging with a 16-day sequence. OLI includes 9 spectral bands that are used for environmental monitoring studies
Classification methods are conventionally divided into supervised and unsupervised classification. Supervised methods require basic information such as the number of classes, their characteristics, as well as known examples of each class. On the other hand, unsupervised methods are more automatic, which do not need known samples and make decisions about their classification based on the values of the pixels themselves. The use of object-oriented methods in processing satellite images has increased the potential of the applicability of environmental remote sensing research. In general, object-oriented classification has three stages: 1) image segmentation 2) classification 3) classification accuracy evaluation. A segment means a group of neighboring pixels within an area whose similarity (such as numerical value and texture) is the most important common criterion. Classification is done in object-oriented methods using different algorithms including: random forest, decision tree, support vector machine, KNN and Bayes.
Conclusion
To better compare the accuracy of object-oriented classification methods in all methods, the number of training samples was considered the same in different land covers. The accuracy of different methods is different and it shows the importance of comparing the accuracy of these methods in order to choose the optimal method. In the decision tree method, a large part of the area is wrongly considered as road cover, which mostly includes vegetation and shadow of buildings. In the random forest method, a part of the wrong area is covered by green areas, while the shadow is residential areas. The presence of shadows in images with high spatial resolution is a big problem for classification. In order to solve the shadow problem in the images, it is necessary to take the imaging exactly at the local noon hour with half an hour increase and decrease, but because it is not possible to take pictures in large areas in this limited time frame, and if the imaging is done on different days at this time in addition to being time-consuming, the existence of different weather conditions causes errors. In some studies, shade is considered as a separate land cover, which has been rejected in many studies and is not practical. Comparison of the object-oriented land cover image using KNN and SVM methods with the original image shows that in these two methods, in some cases, roads are considered as unbuilt land and these areas are mostly in newly built areas and land cover is changing. It is to the building. In these parts of the studied area, the road is partially covered by the soil of the adjacent barren lands and has a different reflection than other pixels with road coverage, which these methods were not able to distinguish.
Kappa coefficients and overall accuracy were used to compare object-oriented classification methods in the preparation of land cover maps of urban areas using images with high spatial resolution. These methods are estimated by comparing classified maps and ground reality, and the closer to one, the higher the accuracy of the method. The results of comparing the kappa coefficients and the overall accuracy of the object-oriented methods are shown in Table 1. The comparison of the kappa coefficients in the object-oriented methods showed that the BAYES method showed higher accuracy than other methods, the kappa coefficient in this method was 0.83 and the overall accuracy It is 87 percent. In the land cover map, this method has the ability to separate different land covers. After that, SVM and KNN methods showed kappa coefficient of 0.78 and 0.76, respectively. Among the object-oriented classification methods, the decision tree method and random forest showed the lowest accuracy with kappa coefficients of 0.70 and 0.71. In order from the most accurate to the least accurate, object-oriented methods include BAYES, SVM, KNN, decision tree and random forest. The results of this research can be used for planners and decision makers in urban science and urban green space.