State of the Map 2022 - Academic Track

In addition to important ecosystem services such as clean air or local climate regulation, green spaces provide peace and recreation, contributing to a good quality of life for the population. In high-density urban areas, publicly accessible green spaces are used for a variety of recreational activities, which has become even more important, not least because of the COVID-19 pandemic [1]–[4]. In this context, the research project "Information and Navigation on Urban Green Spaces in Cities - meinGruen" examined publicly accessible green spaces with regard to a variety of criteria in order to assess their suitability for the pursuit of leisure activities, such as going for a walk or playing soccer [3], [5], [6]. The aim of this study is to derive a suitable polygon dataset to describe the spatial distribution of publicly accessible urban green spaces. The presented approach favors the use of OpenStreetMap data and intrinsic knowledge. Advantages of the use of OpenStreetMap data are the global availability, the often high completeness in urban areas as well as the unified open data license ODbL 1.0. In this way, problems with data availability and heterogeneity due to different responsible authorities can be avoided. Ludwig et al. [7] describe an approach to mapping public green spaces based on OpenStreetMap and Sentinel-2 satellite imagery in which barriers and land use changes are considered based on a priori (expert knowledge) assumptions for polygon generation. In the approach presented here, spatial delimitation is to be refined by describing barriers by probability values. The term "barrier" is first analyzed in an interdisciplinary way in order to then work out its meaning for the spatial delimitation of a green space. Here, barriers describe the action space of a recreational activity. While there are a number of object types (such as walls, fences, rivers, roads or railroad lines) can be assumed to be barriers with certainty, there are others (such as paths or the change of land use) for which knowledge is still lacking. The study area includes the city of Dresden in Germany, plus a buffer of five kilometers. OpenStreetMap represents the main data source. For training and validation, official cadastral data (ALKIS) as well as a dataset on cadastral parcels owned by the city of Dresden were used.
The methodology consists of six steps: First, according to defined rules, types of barriers were extracted from OpenStreetMap data. Second, we derived a land use layer without overlaps and holes from OpenStreetMap. Here, two options were compared regarding different target schemes for land use classification. Third, a mapping in terms of a “ground-truth“ in selected areas in Dresden followed in order to be able to evaluate the existence of a barrier on site for the extracted paths and changes of land use. Fourth, generic probabilities for the existence of a barrier were determined based on path type or land use change type. Fifth, a polygon mesh was created by applying thresholds to the determined barrier probabilities. Sixth, the generated polygons were enriched with attributes on the number of green space-related POI, such as benches, trash cans, or trees. Models for "greenness" and "accessibility" are thereby trained.
For the technical implementation mainly Docker, PostgreSQL/ PostGIS, Python (Geopandas, Scikit-Learn) and Jupyter Notebook were used. Data import was performed by osm2pgsql and ogr2ogr. For mapping we used the app QField.
Land use layers were successfully generated for the study area using a residual class. The results indicated that the land use classification according to the area scheme of the IOER-Monitor (option B) has a higher thematic accuracy with a maximum of 33 classes (433 original OSM tags were assigned) than the option A based on a classification according to osmlanduse.org/ Schultz et al. (up to 13 classes, based on 61 OSM tags) [8], [9]. The classes of arable land (A: 28.40% / B: 28.06% share of area) as well as forest (A: 21.81% / B: 23.33%) are dominant in both variants. While the residual class takes up 6.29% of the area in option A, it is only 4.88% in option B. For the “ground-truth”, a total of approximately 82.3 km of paths (with 408 line objects) and approximately 64.2 km of land use changes (1720 line objects) were evaluated for the presence of a barrier in two selected areas in Dresden. The land use changes are based on variant B. Data were collected on 61 different land use transitions and four different trail types. While bike lanes can be safely assumed to be a barrier, the "track" (96.8%), "footway" (92.7%), and "path" (86.0%) trail types have a slightly lower barrier probability. Among land use transitions, the forest-meadow (12.6%), meadow-sports facility (22.8%), meadow-park (24.6%), and forest-grassland (26.7%) transitions have the lowest barrier probabilities. Together with the barriers assumed to be safe at the beginning, a line pool is formed, from which different polygon meshes are generated based on different intervals for the barrier probability (p ≥ 0%; p ≥ 20%; p ≥ 40%; p ≥ 60%; p ≥ 80%; p = 100%). The lower the probability threshold, the higher the number of polygons created (whose area decreases). For the "accessibility" model, the number of benches, trash cans, public toilets and public internet were considered per polygon. The logistic regression achieved 76.7% accuracy here, similar to a Support Vector Classifier (SVC). The "greenness" model is based on number of benches, picnic tables, trees, and trash cans per polygon. The accuracy is about 92.3% (for logistic regression and also Support Vector Classifier).
This work successfully demonstrates a new approach to derive publicly accessible green spaces based on OpenStreetMap data considering different qualities of barriers in contact of green spaces. Based on the examined barrier probability of path types and land use transitions, more realistic spatial delineations of green spaces were made possible. The chosen approach is globally applicable due to the use of OpenStreetMap. In each case, locally prevailing climatic and cultural influences must be taken into account. The knowledge collected here can be applied in the Central European region. For other areas, a renewed “ground-truth” may have to be carried out on site. The schematic transformation of the land use into the area scheme of the IOER-Monitor leads to a reduction of classes compared to the original data. In addition to benefits in capturing barrier probability, it also simplifies comparability and transferability. Thus, other data could also be migrated into this scheme. The determined barrier probabilities correspond to the expectations. The polygon generation based on different barrier probabilities allows here a differentiated setting of the desired action space for the relevant leisure activities. The quality of the trained models is good, but can be improved. A variety of additional features can be calculated for each potential green space (polygon), such as path network density or density of path network intersections (see also Ludwig et al. [7]). Questions about the perception and use of green spaces can also be part of interdisciplinary research in the future.