Open Land Use for Africa (OLU4Africa)
This is the first progress report of the team No. 3 of the Nairobi INSPIRE Hackathon 2019. The team is led by Dmitrii Kozuch.
You can watch the webinar recording of this team below.
In the first
stage the work is more oriented on the technical development rather than
contacts with the community. The work went into two parallel directions that
will be described in this progress report. The ultimate goal of these two
directions is to enable land use/cover classification and segmentation of
satellite imagery.
The first
direction is relying on creating classification models using Sentinel-2 imagery,
OpenStreetMap database and Keras (with Tensorflow backend) library.
The first
steps are relying in collecting training and validation data for different
types of land cover/use.
The example
of data collection of aerodromes is following with single steps:
1) Download
latest Kenya OSM database from the pages: http://download.geofabrik.de/africa.html
2) Import
osm file into PostgreSQL database using osm2pgsql tool:
3) Use
script to download aerodromes images:
As a result it is 70 sample images that could be use for training and validation of airports and classification model. The model itself will be trained in Keras library using Convolution Neural Network algorithm. There were lots of similar intentions for various image classifications implemented in Keras. Here is an example to classify CIFAR-10 dataset - https://keras.io/examples/cifar10_cnn/ . The dataset itself is composed of 6 000 images for each following class: airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks.
Obviously except the problem with classification we need to tackle problem with image segmentation. So far it wasn‘t time to research existing algorithms for this.
The second
parallel way that it is worked on is to use eo-learn library and Africover
dataset (http://www.un-spider.org/links-and-resources/data-sources/land-cover-kenya-africover-fao
) for the land cover classification. The tutorial doing this is found on the
official library web page:
https://eo-learn.readthedocs.io/en/latest/examples/land-cover-map/SI_LULC_pipeline.html
It has been
attempted to use the code provided, however there are some error emerging here
and there. Thus there is a need to debug the code. The eo-learn library is
quite young and thus it is natural to expect some bugs. All in all priority
will be given to the first way because it has been tested many times so far and
generally could give accuracy of classification higher than 70% . The only
thing is that sometimes it could be too few samples to train and validate
models so it will be need to download samples from the other countries in
Africa.