Quantifying greenspace of Karachi, Pakistan using high-resolution satellite imagery and deep learning
*Hover over each Union Council to see example greenspaces and per capita greenspace.*
*Click to zoom in to highlight the Union Council area.*
Greenspaces in communities are critical for mitigating effects of climate change and have important impacts on our health. In Karachi city, Pakistan, we demonstrate a deep learning-based approach which includes green augmentation for measuring and delineating types of greenspace, with 0.15m high-resolution satellite imagery. Given the global importance of greenspaces for improved planetary and human health, this method is relevant to, and can be extended to locations worldwide.
Fine-grained maps of per-capita greenspace
Highly granular greenspace mapping via this method enables us to compute how much greenspace is available by neighborhood in the city (Only populated areas are considered). We found that the greenspace availability also varies highly across union councils; some union councils have over 30 m2 per person (these are union councils on the periphery of the city with large agricultural lands) while 6 union councils with the lowest amount have less than just 0.1 m2 per person. The average greenspace per person across union councils (smallest administrative level in Karachi) is 2.84 m2/person which significantly lags World Health Organization recommendations (minimum of 9 m2 per person and an ideal value of 50 m2 per person).
How we did it
1) With expertise and local knowledge of Karachi surroundings, we first labelled a sample (over 400) of satellite images across the city at the pixel level to identify greenspaces by type: Tree, Shrub, and Grass.
2) Next, we trained a deep learning model which we enhanced using a new method to adjust green light in an image and better detect Trees and Grass.
The below figure shows the exact pipeline of how our method works. See our paper if you are interested in the details.
3) We then used the model to identify where all Trees, Shrubs and Grass are in unlabelled images covering the entire area of Karachi.
4) Finally, the below figure and table show how our method outperforms existing gold standard methods for vegetation detection, which use combinations of different colors of images to identify greenspaces (GRVI, VARI, GLI). DeepLabv3+ (GreenAug), which is our method with green color augmentation, segments 89.4% of the greenspace, and correctly assigns 90.6% of image pixels as greenspace or not, while the best performing vegetation index only detects 63.3% of the greenspace of which 64.0% of predictions are correct.
Related links
- Code for the project
- Link to the location of all images in the study
- Download of our Karachi image annotations
- Eartharxiv paper link
This project was conceived, co-designed and implemented in collaboration with the CITRIC Health Data Science Center at the Aga Khan University.