AgriFieldNet Model for Crop Types Detection

First place solution of the of the Zindi AgriFieldNet India Challenge for Crop Types Detection from Satellite Imagery.

MLHub model id: model_ecaas_agrifieldnet_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
• Python 3.8
• R 4.2.2

Hardware Requirements

	Training	Inferencing
RAM	25 GB RAM	16 GB RAM
NVIDIA GPU	Required	Optional (but very slow for Unet model)

Get Started With Inferencing

Start by cloning this repository to get the model scripts and saved model checkpoint files:

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

To get started, the R and Python dependencies must to be installed locally in your environment. Alternatively you can look at the original AgriFieldNet Challenge solution scripts (.zip) which are linked in the model documentation.

R and packages

R 4.2.2 is required for the data preprocessing and feature engineering step of the model.

R -e "install.packages(c('devtools', 'plyr', 'tidyverse', 'raster', 'celestial', 'caret', 'fastICA', 'SOAR', 'RStoolbox', 'jsonlite', 'data.table', 'spdep'))"
R -e "devtools::install_github('OpenDroneMap/FIELDimageR')"

Python and packages

Python 3.8 is required for the model training and inferencing steps.

pip install --upgrade pip
pip install -r requirements.txt

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/ref_agrifieldnet_competition_v1: for inferencing:
       • ref_agrifieldnet_competition_v1_source: folder containing for the satellite imagery bands
         • ref_agrifieldnet_competition_v1_source_{folder_id}: e.g. ref_agrifieldnet_competition_v1_source_0a1d5
       • ref_agrifieldnet_competition_v1_labels_test: containing the field ids
         • ref_agrifieldnet_competition_v1_labels_test_{folder_id}: e.g. ref_agrifieldnet_competition_v1_labels_test_0a1d5
   • 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_ecaas_agrifieldnet_gold/data/input/"
   export OUTPUT_DATA="/home/my_user/model_ecaas_agrifieldnet_gold/data/output/"

3. Run the run_models.sh bash shell script.

   ./run_models.sh

4. Wait for the script to finish running, then inspect the OUTPUT_DATA folder for results. If you run into errors, or missing packages, alternatively you can look at the original AgriFieldNet Challenge solution scripts (.zip) which are linked in the model documentation.

Understanding Output Data

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