2019 CGIAR Big Data in Agriculture Platform Convention

Earth observations (EO) provide reliable and temporally consistent data globally. Machine Learning (ML) techniques enable us to use these data and develop data-driven models to detect anomalies, identify patterns, and predict future states, among other applications. However, these techniques rely on training data to build a model. In the context of agriculture, training data can include variables such as the location of farms (e.g plot boundaries), crop type, temporal information of crop growth stage, crop yield, management practices, and many other categories depending on the problem of interest. A majority of these variables can only be measured with ground referencing which requires extensive resources. Therefore, it is necessary to standardize ground referencing best practices to maximize the use of these training data, and build benchmarks to evaluate models’ accuracy.

Existing ground referencing efforts are not aimed to develop training data for EO applications; as a result, in many cases the variables and collected metadata lack enough granularity to be used with EO to build ML models. This session aims to bring together ML, EO and agricultural experts to draft a set of best practices for ground referencing and metadata collection. Our goal is to review the challenges with existing approaches for ground referencing used by different groups (both at CGIAR centers and outside). Then, we will have a technical group discussion to develop a framework for future ground referencing campaigns. This framework will include specific metadata fields that would be required to generate training data as well as best practices to assess and report uncertainty in the data. We plan to share the framework with the broader community after the workshop to get feedback and progressively update it.

In this session, we will discuss the following questions:

1. Why do we need georeferenced data?
2. What does the ideal georeferenced data look like?
3. What is the current status of georeferenced data collection and management in CGIAR?
4. What are the challenges of collecting and sharing georeferenced data?
5. What are the strategies to address the challenges?
6. What are the best practices for managing georeferenced data and sharing them?

We will organize the session in the following order:

1. Alando (15 min): Introduction on why we need good georeferenced data and what the current status
2. Murali (15 min): Example at ICRISAT on what types of georeferenced data are collected and how they are used in research
3. Todd (15 min): Current status of georeferenced data collection in CGIAR
4. Rob (15 min): Best practices of georeferenced data collection, management, and sharing
5. Group discussion and Q&A (30 min): Strategies for addressing challenges and formulating recommendations