Mapping Greenness Visibility in the Netherlands: A Viewshed-Based Approach and Socio-economic Modeling
The visibility of green spaces and features, particularly natural elements such as trees, is known as 'greenness visibility' and is associated with numerous health benefits. However, previous studies have largely focused on smaller scales, such as cities or regions, when examining visibility metrics, partly due to their computational challenges. In response, this study presents a methodological approach to compute tree visibility at a national level in the Netherlands, while assessing its spatial pattern as well relation to selected socioeconomic factors.
The aim of this study is threefold:
- compute the tree visibility using viewshed-based approach for the entire Netherlands in a reproducible manner;
- analyse tree visibility inequality using weighted GINI index;
- model the relationship of tree visibility with selected socio-economic variables.
To successfully run this project, the following dependencies are required:
geopandas: A library for working with geo data.numpy: A library for scientific computation.osmium: A library for working with OpenStreetMap data.shapely: A Python package for manipulation and analysis of geometric objects.rasterio: A library for reading and writing raster data.contextily: A library for adding basemaps to geo plots.missingno: A library for visualizing missing data.cbsodata: A library for extracting CBS Open Data.libpysal: A library for spatial analysis and modeling.mgwr: A library for geographically weighted regression.seaborn: A library for statistical data visualization.
The workflow of this project follows a sequence of steps as illustrated below.
Step 1 - data extraction: The data was collected for Digital Surface Model (DSM) and Digital Terrain Model (DTM) from Actueel Hoogtebestand Nederland (AHN) in 1100 tiles for each model. Tree coverage data was obtained from Atlas Living Environment, street data using Python's library 'pyrosm', and socio-economic variables using the 'cbsodata' library.
Step 2 - preprocess data: a) both DTM and DSM data models are merged into one tif file, which is subsequently divided into larger tiles consisting of 10 individual tiles each. The division is done by defining a spatial extent of the area (the Netherlands in this case). This step is necessary for computational efficiency, while minimazing the need for cropping the street data (had the initial 1000 tiles were preserved) (see below).
b) The street data is obtained in a '.pbf' format, hence it is converted and preprocessed into '.gpkg'. The street geometries data are sampled at 50 metres intervals; if its linestring is shorter than 50 meters, a point is taken at its centroid (see below).
c) The tree dataset is converted into a binary based on cell threshold 10 to accurately represent the degree of detail of tree coverage (see below).
d) An additional sub-step involves checking the resolution of the DSM, DTM, and tree datasets. Since the tree dataset has a resolution of 10x10, while the DSM and DTM had a resolution of 5x5, we downsampled the tree dataset to match the granularity of the DSM and DTM files.
Step 3 - data quality check (optional): this step follows a function that assesses the presence of missing data in the DSM and DTM files, as these files are often subject to missing data. A buffer of a specified length (specify according to the defined radius in the viewshed) is generated as a mask surrounding the streets. Consequently, only the cells relevant to the VGVI calculation are evaluated.
Step 4 - tree visibility computation: this step is done in R for computational efficiency reasons. Folders with preprocessed DSM, DTM, tree data and street points are inputted into a function, where VGVI score is calculated for each tile. Note that the street data input is included in its entirety as it has a relatively smaller size compared to the DSM and DTM LiDAR and tree datasets. Moreover, some tiles may contain only NaNs (e.g., a result of being in the ocean), hence the function checks for such a scenario prior to the computation. This check is necessary as the original VGVI function does not account for this particular condition.
Step 5 - VGVI validation: the reults of the step 3 are validated by performing a Pearson's coorelation between VGVI and Normalized Difference Vegetation Index (NDVI).
Step 6 - neighbourhood-level aggregation: the resulting tree visibility is spatially merged with two geodataframes obtained from cbs. One contains geometries and data for "buurt" (the smallest district) and the other one for "wijk" (a neighborhood composed of multiple buurts).
Step 7 - Modeling socio-economic variables and tree visibility: Modeling is done at the buurt level, with three sets of models: spatial autocorrelation models (Moran's Statistic, LISA, GI*), GINI index and GWR/MGWR. For the MGWR modeling a special fastgwr package needs to be used to handle the data volume(see dependencies).
The below image represents the final calculation of the VGVI for the entire Netherlands at three levels:
The below image shows significant clusters of LISA's local spatial autocorrelations, where high–high (HH) represents hotspots, low-low (LL) represents coldspots, and LH and HL represent spatial outliers:
The below image shows the weighted Gini index by population and its results on the neighbourhood-level VGVI.
The below image shows the MGWR coefficients of the factors examined in the spatial context. The greyed-out area represents a non-significant spatial extent of coefficients.
Creator: O. Mlynarcik (Ondrej), Masters student Applied Data Science, Faculty of Science, Utrecht University (ondrejmlynarc@gmail.com)
Supervisor: S.M. Labib (Labib), Assistant Professor of Data Science and Health, Faculty of GeoSciences, Utrecht University (s.m.labib@uu.nl)