PlasticDetectionService

This repository contains a cloud-based pipeline for predicting marine debris from Sentinel-2 L2A images using machine learning. The predictions are made using the PlasticDetectionModel. The output is stored in a PostGIS database hosted on AWS RDS. The predictions are displayed on our mapping application, deployed here: https://map.oceanecowatch.org

This repository is triggered by the OceanEcoMapServer via a Github Actions workflow endpoint.

Dependencies

• AWS account with configured access to RDS and S3 services.
• Rundpod account with a serverless endpoint.
• Access to Sentinel-Hub Processing API

Running the service manually

The service can be run manually by triggering the workflow_dispatch event here

Local installation

1. Clone the repository
2. Install the dependencies

pip install -r requirements.txt

3. Create a .env file in the root directory with the following variables:

DB_USER=
DB_PW=
DB_NAME=
DB_HOST=
DB_PORT=
SH_INSTANCE_ID=
SH_CLIENT_ID=
SH_CLIENT_SECRET=
AWS_ACCESS_KEY_ID=
AWS_SECRET_ACCESS_KEY=
AWS_DEFAULT_REGION=
RUNPOD_API_KEY=
RUNPOD_ENDPOINT_ID=

3. Greate the database and necessary tables

python -m scripts.reset_db

python -m scripts.add_tables.py

4. Run the service locally The job and model id will be 1 if you have run the reset_db and add_job_to_db scripts

python -m src.main --job-id <job-id> --probability-threshold <prob-theshold>

Development environment and testing

pip install -r requirements-dev.txt

1. Run quick unit tests

pytest -m 'not integration and not slow'

2. Run integration tests

pytest -m 'integration'

3. Run slow (real inference) tests

pytest -m 'slow and not integration'

Software Design Documentation

software_design_documentation