Speaker
Description
Man-made climate change is leading to an increase in average air temperatures. This has a particular impact on densely populated and sealed urban areas by intensifying the urban heat island effect. An important instrument for adapting the associated temperature changes is the development, preservation and establishment of green spaces in urban areas. The creation of a green space mapping system is essential for the inventory and further development of green spaces. However, the heterogeneity and diffuse distribution of green spaces in the spatial context make mapping them considerably more difficult. It is necessary to define the concept of green spaces. Subsequently, with reference to the challenge of a complete mapping of all green spaces on site, the extent to which such areas can be identified, delimited from each other and recognised as locally climate-active with the help of artificial intelligence can be investigated.
The initial definition of green spaces is carried out at a general level due to the heterogeneous nature and characteristics of the areas. Various scientific demarcations are used for this purpose. Based on this, they can be categorised as grass or shrub areas, water areas, large shrub areas and vegetation-free ground areas as well as existing trees. The defined green spaces are analysed using a pre-trained deep learning model to identify the general land cover.
The application of this pre-trained deep learning algorithm can identify green areas based on the general land cover. The best values are achieved for inner-city locations. However, the green areas located outside the built structures are often incorrectly delineated due to their nature and labelled as water areas, so that the pre-trained model achieves an overall accuracy of 46%. After assessing the results, the deep learning algorithm is to be retrained to potentially improve the model. Re-training the pre-trained deep learning model using transfer learning leads to a general improvement in overall accuracy to 78%. In this case, the retrained model has a higher level of detail in the assessment of out-of-town green spaces - inner-urban green spaces are identified more poorly compared to the pre-trained model. A subsequent combination of the modelling of the pre-trained and the post-trained deep learning algorithm leads to an overall accuracy of 86%.
Based on this, the climate activity of the green spaces is determined with the help of combination modelling. This is characterised by the potential cooling capacity of the identified green spaces using the proportion of shade, the ground surface and the absolute area size. The green spaces with low climate activity can be found in the heavily compacted and sealed inner-urban locations. With an increase in unsealed areas, the absolute area size and thus the climate activity increases from an initial cooling capacity of less than 1 K in heavily sealed areas to > 3 K in unsealed areas outside the settlement, which then also have a transfer effect beyond the actual green space due to the high cooling capacity.
| Keywords | Artificial intelligence; green spaces; urban planning; climate adaption; smart city |
|---|---|
| Best Congress Paper Award | No |
