Spatio-Temporal Attention-based Unet for Field Boundary Detection

First place solution for NASA Harvest Field Boundary Detection Challenge

MLHub model id: model_nasa_rwanda_field_boundary_competition_gold_v1. Browse on Radiant MLHub.

ML Model Documentation

Please review the model architecture, license, applicable spatial and temporal extents and other details in the model documentation.

System Requirements

• Git client
• Docker with Compose v1.28 or newer.

Hardware Requirements

	Training	Inferencing
RAM	25 GB RAM	16 GB RAM
NVIDIA GPU	A100 80GB	Optional (but very slow)

Get Started With Inferencing

First clone this Git repository.

git clone https://github.com/radiantearth/model_nasa_rwanda_field_boundary_competition_gold.git
cd model_nasa_rwanda_field_boundary_competition_gold/

After cloning the model repository, you can use the Docker Compose runtime files as described below.

Pull or Build the Docker Image

Pull pre-built image from Docker Hub (recommended):

docker pull docker.io/radiantearth/model_nasa_rwanda_field_boundary_competition_gold:1

Or build image from source:

docker build -t radiantearth/model_nasa_rwanda_field_boundary_competition_gold:1 -f Dockerfile .

Run Model to Generate New Inferences

1. Prepare your input and output data folders. The data/ folder in this repository contains some placeholder files to guide you.

   • The data/ folder must contain:

     • input/: input folder containing the tile imagery. This folder has the following naming convention: {dataset_id}_{tile_id}_{year}_{month}. For example, for a dataset_id (nasa_rwanda_field_boundary_competition), a tile_id (00) and a year (2021), we'll have the following structure:

           nasa_rwanda_field_boundary_competition_source_test_00_2021_03/
           nasa_rwanda_field_boundary_competition_source_test_00_2021_04/
           nasa_rwanda_field_boundary_competition_source_test_00_2021_08/
           nasa_rwanda_field_boundary_competition_source_test_00_2021_10/
           nasa_rwanda_field_boundary_competition_source_test_00_2021_11/
           nasa_rwanda_field_boundary_competition_source_test_00_2021_12/
       

       Notice that the months are fixed and have to be : ['2021_03', '2021_04', '2021_08', '2021_10', '2021_11', '2021_12'].

     • The output/ folder is where the model will write inferencing results.

2. Set INPUT_DATA and OUTPUT_DATA environment variables corresponding with your input and output folders. These commands will vary depending on operating system and command-line shell:

   # change paths to your actual input and output folders
   export INPUT_DATA="/home/my_user/model_nasa_rwanda_field_boundary_competition_gold/data/input/"
   export OUTPUT_DATA="/home/my_user/model_nasa_rwanda_field_boundary_competition_gold/data/output/"
   export DATASET_ID="nasa_rwanda_field_boundary_competition"

3. Run the appropriate Docker Compose command for your system

   docker compose up model_nasa_rwanda_field_boundary_competition_gold_v1

4. Wait for the docker compose to finish running, then inspect the OUTPUT_DATA folder for results.

Understanding Output Data

Please review the model output format and other technical details in the model documentation.