Zhou Zhang, assistant professor of biological systems engineering, was recently awarded three USDA NIFA grants that will help move forward her research program, which focuses on developing machine learning models and AI tools to increase agricultural productivity and sustainability. Read below for details about her newly-funded projects.

AlfAdvisor: A Web-Based Cyber-Platform to Estimate Alfalfa Yield and Quality To Support Harvest Scheduling (funded by USDA NIFA AFRP)

Project Description: The research team includes Zhou Zhang (BSE), Matt Digman (BSE), Paul Mitchell (Ag Economics) from UW-Madison, Jerry Cherney from Cornell and Jinha Jung from Purdue. Timing alfalfa cutting is a critical management practice to maximize production potential, quality, and profitability of a standing alfalfa crop. However, deciding when to harvest can be a difficult decision as it is affected by multiple and competing factors, such as, quality versus yield, drying rate and weather. Despite its importance and complexity, the decision to harvest is often made with crude guidelines (e.g., cut every 28 days). The goal of this project is to develop a free and public cyber-platform (AlfAdvisor) to help alfalfa growers plan for economically optimal harvest scheduling. Our regional collaborative team will (1) develop models to estimate alfalfa yield, quality and drying rate in real-time by combining satellite imagery and environmental factors; (2) develop a decision-making model to provide economically optimum harvest scheduling; and (3) create a web-based cyber-platform to disseminate tools for free public use. Training materials on using the AlfAdvisor will be created and distributed through multiple extension and outreach approaches.

Harnessing Machine Learning and Hyperspectral Imaging for High-Throughput Maize Silage Phenotyping (funded by USDA NIFA AFRI)

Project Description: The research team includes Zhou Zhang (BSE) and Natalia De Leon (Agronomy) from UW-Madison. Improved forage cultivars provide economic opportunities for livestock and crop farming operations and promote a more stable, sustainable agriculture. However, the existing laboratory-based forage quality assessment approaches are labor-intensive and time-consuming, and thereby greatly limit the genetic selection and forage breeding efficiency. To enhance forage phenotyping capacity, the overall goal of this project is to assess maize silage yield and quality traits in a high-throughput manner by melding cutting-edge remote sensing and machine learning technologies in a field setting. Specifically, the proposed research addresses two key plant phenotyping challenges from the data science perspectives by: (1) developing multi-temporal feature fusion approaches to fully exploit the potential of time-series hyperspectral data; (2) developing unsupervised domain adaptation strategies to increase the model transferability across different environments to avoid continuous label effort associated with environmental changes. The maize silage mix is more complex than other forage species as it includes grain and stover, therefore methods developed for this plant structure can potentially be applied to other forage commodities with less complex plant structures for biochemical composition assessment.

Developing an Integrated Deep Learning Modeling Framework For County-Level Crop Yield Prediction in support of USDA Nass Operation  (funded by USDA NIFA AFRI)

Project Description: The research team includes Zhou Zhang (BSE), Qunying Huang (Geography) from UW-Madison, and Zhengwei Yang from USDA NASS. Accurate and timely estimation of regional crop yield is of great importance for food security monitoring, market planning, and farm resource management. The USDA NASS publishes major crop yield predictions monthly at the state-level before harvest through NASS Crop Production Report. However, there are no monthly or weekly county-level crop yield estimates available from NASS report, though they are in heavy demand and critical to many stakeholders in the government, the academic community, and the private sector for their business decisions. Moreover, NASS’s crop yield estimation relies on in-season surveys which are costly, time-consuming, and labor-intensive. To fill this gap, this project aims to produce weekly in-season crop yield predictions over the continental United States (CONUS) at the county-level, by utilizing publicly available satellite remote sensing datasets and state-of-the-art deep learning (DL) technologies. Specifically, we will (1) Develop a Bayesian deep learning model to provide weekly crop yield predictions and uncertainties at the county-level in the CONUS; (2) Integrate the deep learning-based yield prediction model with unsupervised domain adaptation to improve the model performance in counties with no or limited reported yield data; and (3) Create a web-based cyber-platform to disseminate the integrated modeling toolkit and yield prediction results for free public use. The developed models predicated county-level yield will provide complementary information to the in-season NASS Crop Production Report.