Skip to main content

UDiscoverIt Downloads — Agriculture, Soils, and Food Systems

All projects

Agriculture, Soils, and Food Systems
A densely-packed corn field in Wisconsin, USA. The plants look wilted and dry, and are brown/golden colored.

A corn field in Wisconsin, USA. Crop yields are highly sensitive to extreme weather conditions such as heatwaves and droughts. Soil quality plays a major role as well.

While the world population is projected to increase to 10 billion by 2050, our food systems are struggling to keep up. Feeding our rapidly-growing population in a sustainable way presents a major challenge, as food production is highly sensitive to extreme weather and contributes to environmental degradation. In order for policymakers to make informed decisions about this, accurate predictions are crucial. Forecasting crop yields and monitoring crop health helps policymakers and farmers prevent food insecurity and stabilize supply chains. Aquaculture (fish farming) can provide another sustainable source of protein, but may result in deforestation and land degradation; we need accurate maps of aquaculture expansion to help policymakers navigate these tradeoffs. Soils also have the potential to sequester carbon dioxide and mitigate climate change, but we need clear information on how long carbon stays in the soil in different environments. While these problems have proven challenging for traditional AI methods due to limited or biased data, we are developing state-of-the-art AI methods that leverage spatiotemporal structure, scientific knowledge, and weak supervision to tackle these challenges, giving us better predictions and scientific insights.

A GNN-RNN Approach for Harnessing Geospatial and Temporal Information: Application to Crop Yield Prediction

Accurately forecasting crop yields helps policymakers and farmers make decisions to prevent food insecurity and stabilize supply chains. However, this prediction task is exceptionally complicated since crop yields depend on numerous factors such as weather, land surface, and soil quality, as well as their interactions. While simple machine learning models have begun to be adopted in this domain, they struggle with the high dimensionality and complex spatiotemporal structure of the data. Our team has proposed a GNN-RNN framework for the complex spatiotemporal prediction task of forecasting crop yields. Our model adapts graph, recurrent, and convolutional neural networks into a state-of-the-art deep architecture that captures the spatial and temporal structure of our dataset. We provide Python code that implements the GNN-RNN model in PyTorch and trains it to predict corn and soybean yields for US counties. We also release a comprehensive dataset of county-level yield, weather, soil moisture, and soil texture data for over 3000 counties in the US, ranging from 1981–2020.

Files

Related Publications

Publications — Agriculture, Soils, and Food Systems

2025

2024

2023

2022

2021