Home

Awesome

Aerial Mapping with Drones & Deep Learning in Zanzibar, Tanzania

Motivation:

Open source R&D notebooks of all the steps (deep learning and otherwise) to create a state of the art deep learning building detector & classifier from high-resolution aerial/drone imagery. Something like this:

Interactive version: https://alpha.anthropo.co/znz-demo

static/znz-demo.gif

7/25/2019 Update:

In process of rewriting everything as a series of interactive geospatial deep learning tutorials on Google Colab.

See 1st tutorial published 7/25/2019 for a complete data creation, model creation, inference, and evaluation workflow for building segmentation:

Prior dev notebooks can be found in /archive with details preserved below the line:


Results:

As of 1/18/2019 (internal val only):

Mean F1 scoreFoundation F1Unfinished F1Completed F1All Buildings F1
Internal Val (grid 042)0.7280.7180.7550.7100.796

Top 2 in WeRobotics' Open AI Tanzania Challenge

Mean F1 scoreFoundation F1Unfinished F1Completed F1All Buildings F1
Final Test (grids 059, 066)0.6970.7440.6920.6550.723
Internal Val (grid 042)0.6960.6830.7490.6560.757

Background:

https://blog.werobotics.org/2018/08/06/welcome-to-the-open-ai-tanzania-challenge/

Maps are absolutely essential for decision support. Knowing where buildings are located is a fundamental input for urban planning, public safety, public health, disaster response, environmental protection, sustainable development and census data, for example. Some of these applications typically require timely and high-resolution maps.

https://blog.werobotics.org/wp-content/uploads/2018/08/zmi-geonode.png

Take the following scenario: a local organization that provides low-cost solar panels to low-income households in rural Tanzania is evaluating a large neighborhood with many small houses. They have to determine how to best optimize the installation and distribution of their panels. So they need to know which of the residential structures are oriented in a way that makes them more suitable for solar panels. Knowing where these small houses are and what their orientation is vis-a-vis the sun and surrounding trees, what their roofs look like to determine where to place the panels, and what materials the roofs are made of are all key inputs for their planning. This is just one of many applications for high-resolution maps.

https://competitions.codalab.org/competitions/20100#learn_the_details

Open AI Tanzania — is a partnership with our friends at the State University of Zanzibar (SUZA), WeRobotics, World Bank, OpenAerialMap and Tanzania Flying Labs. Open AI Tanzania invites data scientists to develop feature detection algorithms that can automatically identify buildings and building types using high-resolution aerial imagery collected by Tanzanian drone pilots through the Zanzibar Mapping Initiative (ZMI). The goal of this challenge is to correctly segment and classify building footprints under various stages of construction.

Source imagery & training data license

We request that all participants fill out a Google Form.

The imagery data is released as OpenData using the Creative Commons Attribution 4.0 International license, attribution must be given to: Commission for Lands (COLA), Revolutionary Government of Zanzibar (RGoZ)

Overview:

Training workflow:

static/overview_train.png

Prediction workflow:

static/overview_predict.png

Notebooks (7/25/19 moved to archive):

archive/znz-segment-buildingfootprint-20190108-alldata.ipynb

archive/znz-classify-buildings-20190118.ipynb

archive/znz-inference-20190118.ipynb

archive/znz-postprocess-20190118.ipynb

archive/znz-eval-20190118.ipynb

Ready-to-train preprocessed datasets

znz-segment-z19.zip (212 MB):

znz-classify.zip (259 MB):

Thanks

TO DO