Current projects

1. Precision mechanical weeding on horseradish

Illinois has a rich history as a leading producer of horseradish, dating back 150 years. A persistent challenge in horseradish production is the lack of effective weed management solutions throughout the growing season. The application of post-emergent herbicides is restricted. Currently, weed control relies heavily on manual labor and tool implements that can only pull weeds taller than the crop. This project aims to develop AI vision-based weed control solutions explicitly tailored for horseradish producers, addressing the weed challenges they face. We are developing a mechanical weeding robot based on Farm-ng’s Amiga platform, integrating advanced technologies for improved weed management. Our approach comprises three stages: lane detection and autonomous navigation, weed detection through AI-based object detection, and an actuator mechanism for removing weeds between rows.

Students

Past students: Ken Chen; Yudai Yamada; Janmejay Rathi; Pavankumar Dabilpuram; Paul Osuma; Xinyuan Luo; Yuhang Liu; Benchy Detreuil; Moises Rodriguez; Jusjeev Bhurjee; Caleb Zwilling

2. Generalized yield prediction using remote sensing

Crop yield prediction has been extensively studied for the past 4–5 decades, but developing a generalized approach remains a challenge. Models built for one region often fail to scale or adapt to other regions due to region-specific methodologies. Additionally, the impact of image and yield data resolution, as well as the influence of erroneous ground data on the transferability of prediction models, is underexplored. Most studies rely on ground truth data directly from combines, which often include yield data with delays of several seconds. The implications of using such erroneous data on yield prediction accuracy have not been adequately investigated. This research aims to examine the sources of variability in yield monitor data and assess how inaccuracies in yield data affect prediction models. It also addresses year-to-year variations to enhance model robustness and improve transferability across diverse agricultural regions.

Beyond ground data quality, prediction accuracy also depends heavily on the spatial resolution and platform of the imagery used. County-scale models built from coarse satellite imagery can reliably estimate productivity across millions of acres, but this resolution limits their usefulness for field-level decisions. High-resolution satellite and UAS imagery offer greater potential for site-specific management, yet how prediction performance shifts from county to field scale, and across sensor platforms, remains poorly understood. This research compares NDVI derived from multiple satellite and UAS platforms against measured yield at both county and field scales, and develops transformation methods to map NDVI values between datasets. This work aims to close the gap in scaling and translating yield prediction models across regions, platforms, and crops.

Students:

3. UAS-based monitoring of Western Corn Rootworm

Western corn rootworm is corn’s costliest insect pest, often called the “billion-dollar bug” for the yield losses and lodging risk it causes by feeding on corn roots. Effective management depends on monitoring: growers use sticky traps to track rootworm pressure and decide whether the added cost of Bt corn, about $25 per acre, is justified that season. Walking fields to count insects on sticky cards by hand is slow, so many growers skip monitoring and plant Bt corn every year regardless of actual pest pressure. This project develops a two-stage computer vision pipeline to automate rootworm monitoring using UAS imagery. A first model locates each sticky trap within the drone imagery, exploiting the trap’s bright yellow color against the green canopy to achieve 100% detection accuracy. A second model then works within the detected trap region to count individual insects, using tiled inference to resolve small, closely packed bodies and distinguish rootworm from flies, other insects, and debris. This stage reaches 83% accuracy for identifying western corn rootworm. Future work will extend this system toward a fully autonomous smart trap, with mechanical hardware to renew the sticky surface after each scouting pass. Paired with automatic frame selection from video, this would remove the remaining field-labor step and give growers a scalable, low-cost way to monitor pest pressure across their fields.

Students:

4. Automated field boundary delineation

Field boundary delineation is a critical component of modern agricultural management, enabling farmers, researchers, and policymakers to make informed decisions based on accurate data. Field boundary delineation serves several essential cases of use within the agricultural landscape that include: (i) Tracking changes in field sizes over years; (ii) Detecting regional field shape variations; and (iii) Tracking crop rotations. Despite years of research, accurate field boundary delineation remains a challenge. The USDA’s recently released Crop Sequence Boundary (CSB) dataset, derived from the Cropland Data Layer, is freely available but it has low and coarse accuracy and rough edges, failing to capture finer field details visible in satellite imagery. There are a few research studies that showed AI-based delineation methods, reproducibility is limited due to unaccessible training data. Commercial products offer higher accuracy but are often expensive and inaccessible to researchers and smallholder farmers. Our proposed method addresses these limitations by using traditional image processing techniques that offer a simple, effective, and reproducible approach to field boundary delineation.

Students:

5. Hybrid AI approach for Fusarium-damaged kernel detection

This project develops a hybrid approach to detect Fusarium damaged kernels in wheat, a disease that threatens grain quality and food safety through toxins like deoxynivalenol. The approach is called hybrid because it trains an AI model to detect kernels using only five labeled images, then applies traditional textural features, mimicking human quality metrics, to identify the damaged ones. Building on earlier work that used phone images for damage detection, the model is trained to recognize specific visual traits, such as creasing, roughness, and bleaching, that indicate disease. By moving from broad labeling to trait-based detection, this work aims to make kernel disease screening more interpretable and require far less manually labeled data.

Students: