Validating a window view simulation engine based on open data and open source using semantic segmentation

The visual access to urban green spaces through window views plays a key role in increasing well-being, particularly for those with limited mobility. This study verifies a window view simulation engine around the Green Window View Index (GWVI) that combines open source approaches with open geospatial data. Using a pretrained DeepLab V3+ model on Cityscapes data set for semantic segmentation, the study compares the accuracy of simulated window views to photorealistic semantic segmentations. A total of 40 window views were examined, with 0.1 m and 2.0 m distance to the window. The validation metrics consist of overall accuracy (OAcc), mean accuracy (mAcc), mean intersection over union (mIoU), and individual IoU values for vegetation, sky, and buildings. The statistics show an mIoU of 0.53, with class-specific IoU values of 0.52 for vegetation, 0.64 for the sky, and 0.43 for buildings, an OAcc of 0.68, and an mAcc of 0.74. The approach has a low variance in visibility values, with a minor underestimating of vegetation (-6%), and an overestimation of sky (+5%) and buildings (+3%). These findings indicate that the simulation engine performs well, outlining its potential for analyzing window views in a variety of urban scenarios. Future large-scale crowdsourcing experiments are recommended to statistically support these findings.