Field Boundaries for Agriculture Mid-point Progress Report

Jun 25, 2024

Strategic Intent

The Field Boundaries for Agriculture Initiative is advancing the methodologies to create, share and update datasets containing field boundaries for increased understanding of sustainable global agricultural systems. To that end, the project is seeding progress in AI and computer vision models to extract field boundaries from earth observation imagery. In parallel, we are outlining a framework for publishing data in the cloud for open, collaborative use to reduce the common resource drain associated with managing large global geospatial datasets.

Our method is to convene experts in data science, artificial intelligence, satellite imagery and the overlap of geospatial and agriculture technologies (data producers) with stakeholders who are working on related challenges and could benefit from this work (data consumers).  

We are executing the project in an open and transparent manner. We publish data and associated tools on an open platform implementing the most open licenses possible. We build communities of collaborators aligned to the mission. We leave behind a body of work that is useful and can be enhanced and expanded by a wide variety of stakeholders.

This post provides a mid-project update. The graphic below highlights our progress to date.

AI for Global Geospatial Datasets

AI models and algorithms are entirely dependent on high-quality training, sample, and ground truth datasets. Many of the current ML algorithms are trained on data that is too narrowly focused to ensure the model is broadly applicable, especially when applied to extract geospatial data across the entire globe. “There are no good models without good data (Sambasivan et al. 2021).” TGE’s Field Boundaries Initiative is enhancing the training data, models and benchmarks needed to extract field boundary delineations from satellite imagery using machine learning.

TGE is providing seed funding to two academic research teams whose body of work in machine learning and computer vision aligns with our strategic objectives and has identifiable commercial viability.

  • Dr Hannah Kerner – Arizona State University   Dr. Kerner is currently an Assistant Professor in the School of Computing and Augmented Intelligence at Arizona State University. Her research focuses on developing machine learning systems for real-world data and use cases with an emphasis on remote sensing and spatial datasets, fairness (particularly w.r.t. geographic bias), scientific discovery and exploration, agriculture, and food security. She is the AI/Machine Learning Lead for NASA Harvest and NASA Acres, as well as Center Faculty for the ASU Center for Global Discovery and Conservation Science (GDCS). 
  • Dr Nathan Jacobs – Washington University in St Louis   Dr Nathan Jabobs’ research centers on developing learning-based algorithms and systems for extracting information from large-scale image collections. He has applied this expertise in many application domains, with a particular focus on geospatial and medical applications. He earned a PhD in computer science at Washington University in St. Louis in 2010. Prior to re-joining the institution as a professor in 2022, he was a professor of computer science at the University of Kentucky.
Sentinel-2 images covering areas of 1536m x 1536m during the planting season and harvest season in France and Cambodia.  Ground-truth field polygons are shown on the right (randomly colored).   Comparing the areas in two countries highlights variation between cropping systems in different regions of the world in terms of field size, complexity, and appearance. The innovation being developed here is to enable models to be trained on geographically diverse field boundary data and provide a benchmark for evaluating model performance across diverse regions. Photo courtesy of Dr Hannah Kerner.

Bridge to Market

Routinely, the gap between the existence of an innovation and the broad adoption of an innovation is related to applicability, ease of use, access, and awareness. In the case of field boundaries, the innovation itself is expanding the applicability of a field boundaries dataset to be relevant to any region in the world through computer vision models, training data and benchmarks. TGE’s Field Boundaries Initiative is increasing ease of use, access and awareness in the following, tangible ways:

Ease of Use: The following tools have been developed to lower the barrier of entry for commercial impact. Tools are stored on Github (17 active repositories).  

  • Field Boundaries for Agriculture (Fiboa) schema: developed with input from 25+ stakeholders across global agriculture / ag tech industry
  • Data conversion tools: 20+ tools to translate data into fiboa compliant format 
  • Validator tool: validates compliance to fiboa schema
  • Custom extensions template + extensions to fiboa schema (e.g. tillage; AI ecosystem)
  • List of data available in fiboa format that will be used to create model training dataset
  • Resources: 4 tutorials, live demo sessions, advanced technical documentation

Access: Publishing global geospatial datasets in the cloud under open and permissive terms will facilitate innovation by making high-value datasets available to a much larger audience. Field Boundaries datasets are published on the Source Cooperative geospatial data sharing platform enabling collaborative contributions to global geospatial dataset development through cloud-native geospatial techniques.

  • Field Boundaries datasets (fiboa): 10 datasets (8 GB) covering 7 countries in Europe  
  • Field boundaries computer vision training dataset: built from over 35 source datasets; expected release date end of July 2024  
  • Field Boundaries extraction models and benchmarks: expected release end of July 2024

Awareness: The Innovation Bridge is designed to make industry more aware of relevant innovation emerging from academic efforts. The three ways this is accomplished is by participation, regular open communication, and publication in leading peer reviewed journals and conferences.

Join us!

The Field Boundary Initiative is an open community which strives to be as collaborative and transparent as possible.  Program resources and communication channels are linked below.