Detection of Zilla Parishad(ZP) School(s) from Satellite basemap imagery under the Guidence of UNICEF Data Science Team (https://www.unicef.org/innovation/innovation-fund-geospoc-geospatial)
GNU AGPLv3 https://choosealicense.com/licenses/agpl-3.0/
RS Imagery can be used for multiple applications amongst which building and road detection are common. But School detection is one of the complex problems that needs to be addressed given the variety of features they cover. This repository marks the begining of the school detection project in this direction.
The data for the schools is input in the form of POI's in the csv file. The user has to input the csv file for downloading mapbox data at specific Latitude & Longitude for training data.The csv file contains necessary Latitude and Longitude and accordingly the script loads csv file and downloads the tiles at zoom level 17 from Mapbox.
Mask for each image(tile) is obtained using roi-poly package in Python.The data path inside the directory needs to be changed inside label_unicef.py file as dpath and target directory where mask is stored as mpath then the image will be plotted on the window and user can label out the area for school.On the console,the user has to run :
python label_unicef.py
Then,user can label the image which is plotted as a matplotlib plot.
The model training and prediction phase utilises the satellite unet model implemented through keras-unet package.The script utilises the following packages with their versions :
- python (3.7)
- keras (2.3.1)
- keras-unet (1.15.2)
- scikit-image (0.16.2)
- pillow (8.2.0)
- matplotlib (3.3.4)
The script is tested with keras 2.3.1 with tensorflow 1.15.2. For inference,the weights are available inside the weights directory.
The packages required are mentioned in requirements.txt ,these packages are in addition to the ones mentioned above in the Model Training section.Further,for testing purposes,coveralls,coverage.py and pytest are utilised.