Cyberinfrastructure for machine learning applications in agriculture: experiences, analysis, and vision

Area of Science:

• Agricultural Science
• Computer Science
• Data Science

Background:

• Machine learning (ML) advancements are rapid, but agricultural applications are limited by a lack of high-quality, diverse datasets and reusable cyberinfrastructure (CI).
• Existing compute resources and ML algorithms are capable, but data collection, processing, and training infrastructure lags behind, hindering agricultural innovation.

Purpose of the Study:

• To address challenges in agricultural data by exploring the collection, processing, and training of ML models using a multimodal dataset.
• To propose a vision for agriculture-focused CI to accelerate the development and deployment of ML solutions in the agricultural sector.

Main Methods:

• Collected 1 terabyte (TB) of multimodal data during the 2023 growing season, including Unmanned Aerial System (UAS) imagery (RGB and multispectral) and soil/weather sensor data.
• Focused on corn and soybean crops, training ML models for predictions of crop growth stage, soil moisture, and final yield.
• Developed four key CI components: a UAS imagery pipeline, a tabular data pipeline, an adapted vision transformer (ViT) model architecture, and a data visualization prototype.

Main Results:

• Created four reusable CI components designed to enhance ML model accuracy in agriculture.
• The UAS imagery pipeline improved processing time and image quality.
• The tabular data pipeline aggregated and aligned data from multiple sources, and the ViT adaptation incorporated domain expertise, improving data trust and outlier identification.

Conclusions:

• The developed CI components offer a pathway to higher accuracy ML predictions in agriculture.
• Further work is needed to mature these components and implement them on high-performance computing (HPC) infrastructure.
• Open questions remain on how best to leverage CI to meet the agricultural community's needs for accelerated ML application development.