Field Mapper (Field of The World)

Project Summary

This repository shows the pipelines implemented for performing transductive learning leveraging the semantic information in the Field of the World Datasets transferred to selected areas in Africa, built on the Lacuna Funds Labels hosted by the Agricultural Impacts Research Group at Clark University for the Task of Crop Boundary Segmentation. Large scale crop boundary segmentation datasets are often scarce over the African region. For example, the Field of the World Crop Boundary Segmentation datasets, covers only two selected countries in the African region. In this project, we experiment with the large scale datasets and their ability to adequately transfer to new regions using datasets of different modalities. Hence, this repository implements boundary segmentation over selected areas in Africa leveraging pretrained models based of the Field of the World datasets.

Methodology

The approach taken in this project is implemented in two phases:

1. Transferring the weights of the FTW Models (UNet Attention and DeepLab v3+) to the African Labels from Lacuna Fund,
2. End to End Training using the Models from Agricultural Impact Research Group and the FTW datasets and testing different transfer learning scenarios for the African Labels from Lacuna Fund.

Datasets description

Fields of the World

Fields of the World

Lacuna Labels

Lacuna Labels

To access the data, use the AWS CLI to download the data to a local directory. We recommend making a new directory called “data” in your home directory, changing into that, and then using the sync function, as follows (note: this assumes a *nix-based terminal.

cd ~
mkdir data
cd data
aws s3 sync s3://africa-field-boundary-labels/ . --dryrun
aws s3 sync s3://africa-field-boundary-labels/ . 

If the second to last line previews a successful download, run the last line, which will download all data on the bucket into your directory.

The AWS S3 console can also be used to download the data.

How to run pipeline

FTW Model Weight: https://huggingface.co/torchgeo/fields-of-the-world/tree/main

1. How to run notebook in an attached docker file,
2. Report in Quarto

Contributors

1. Rufai Omowunmi Balogun
2. Tanmoy Chakraborty
3. Isaiah Taylor

References

1. A region-wide, multi-year set of crop field boundary labels for Africa
2. Fields of the World (FTW): A Comprehensive Benchmark Datasets for Agricultural Field Boundaries