Skydipper Model Training PoC

An spanish version of this document is provided here

Description

Skydipper is a Simbiótica S.L. (D.B.A. Vizzuality) project, cofinanced by the Ministry of Energy, Tourism and Digital Agenda as part of the National Scientific Research, Development, and Technical Innovation Plan 2013-1016 with expedient number TSI-100504-2017-1.

The aim of the project is the development of an open source platform for the detection of environmental risks through the application of Machine Learning models to Earth Observation data.

This Proof of Concept (PoC) implements a data pipeline that takes data from a cloud provider (Google Earth Engine in this case), performs the necessary transformations on the original data and trains a computer vision model able to discriminate crops. An extra validation process is provided where the model is tested with a fraction of the original data.

The satellite imagery is taken from the Copernicus 2 satellite data (an initiative of the European Union), and the crops labels are taken from the Croplands data source, provided by the USDA.

Running the PoC

This process is implemented in docker-compose. As part of the stack configured the dockerfiles will download, compile and install several software applications: GDAL, Google Cloud libraries and tensorflow. Your host environment won’t be changed. It’s extremely recommended to run this pipeline in a host with GPU acceleration, either in-cloud (e.g. CPU-enabled GCE instances) or in a workstation with a CUDA enabled graphics card; and nvidia-docker support. More details are provided in this document, but for more details on how to set this up, refer to your platform documentation.

We provide a convenience script that is able to provision the instances tasked with the different steps of the data pipeline. Usage is as follows, given a properly configured environment:

▶ ./train_model.sh
Usage: train_model.sh {generate_data|preprocess|train|ingest}

The four steps (generate_data, preprocess, train, and ingest) are meant to be ran in order, as they’ll produce artifacts in the filesystem – i.e. the first step will create a directory called ‘samples’, to which it will download data from the cloud. Data placed there will be transformed for model traininga and the final model will be outputted to the ‘networks’ directory.

Installing necessary components

You’ll need to install docker and docker-compose to run the PoC. `nvidia-docker` is needed to provide GPU acceleration to the training process, so you’ll want to install it. Refer to the docker documentation and the docker-compose documentation for instructions on how to set these up in your environment.

While it can be ran with CPU support, it’s generally recommended to use a GPU for the model training – otherwise, the model training will take up a lot of time. You’ll need to set up nvidia-docker to add GPU support to the training process. Notice that nvidia-docker needs a suitable graphic card to run (either on a workstation or in a cloud instance). You’ll need to set the docker runtime to nvidia-docker as default or modify the docker-compose file to use this engine.

Access to Google Earth Engine and credentials

Data comes from Google Earth Engine, a platform running on Google Cloud that provides a multi-petabyte catalog of satellite imagery. Google Earth Engine is not an open access service, so you’ll need proper credentials when using this ETL tool. For more information on this, check out the home page of the project, where you can request access. Once you are provided with access to GEE, you’ll have to provide your own authentication credentials. Follow this guide for instructions.

In addition to this, the first step of the process needs credentials to upload the data to Google Cloud. The pipeline is meant to run in the Skydipper platform, but by default the first step needs not to be run (i.e. there are already results stored in the cloud for this step). In case you would like to use your own cloud platform, be advised that you’ll need to adapt the code to reflect the proper locations. In this case, modify the docker-compose file to provide the instances with a $GOOGLE_APPLICATION_CREDENTIALS pointing to a valid service account.

Running the process

Running the process in order will give out the following results:

• ./train_model.sh generate_data will set up a series of areas of interest and instruct Google Earth Engine to upload this data to a bucket in tiff format.
• ./train_model.sh preprocess will download this data, merge it and transform it so it’s in a format suitable for tensorflow to compile a model, in addition to other tasks (e.g. the train-test split).
• ./train_model.sh train will take the preprocessed data and train a neural network on it. It will output the serialized models to the networks directory.
• ./train_model.sh ingest will take the generated neural network model and run it on the testing set.

Model evaluation and results

An example of this workflow is provided in a computational notebook. A rendered version of it can be found in the file workflow.html placed in this same repository.

(c) Vizzuality- Simbiotica S.L. 2017-2019