Quantifying greenspace using deep learning in Karachi, Pakistan

Greenspaces in communities are critical for mitigating effects of climate change and have important impacts on our health. Nowadays, the availability of visible spectrum satellite imagery data combined with deep learning methods, allows for automated greenspace analysis at high resolution. We propose a custom deep learning-based approach which includes novel green color augmentation to better detect as well as delineate types of greenspace vegetation (trees, grass) with satellite imagery. Our method outperforms gold standard methods, which use vegetation indices, by 33.1% (accuracy) and 77.7% (intersection-over-union; IoU). With the proposed augmentation technique, we also show improvement over popular deep learning-based segmentation methods for both classification of total greenspace as well as by vegetation type. We apply the method to high-resolution (0.15 meter per pixel) satellite images covering the entirety of Karachi, Pakistan. Detection across the city can inform planning needs based on where greenspaces exist and in what form; we find that greenspaces in Karachi are often linked to areas of development (Pearson’s correlation coefficient (r) shows a significant 0.352 correlation between greenspaces and paved roads, p < 0.001), with a slightly higher correlation between roads and trees versus roads and grass. Quantifying greenspace in Karachi also illuminates an important need; Karachi has 4.17 square meters of greenspace per capita, which significantly lags World Health Organization recommendations.