Location: Cereal Disease Lab
Project Number: 5062-21220-025-031-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jul 1, 2024
End Date: Jun 30, 2025
Objective:
The goal of this project is to provide and validate comprehensive and timely wheat rust disease information at the field level and predicting risk of spread using multi-modal data including historical, weather, drone, and satellite data. The objectives are to synthesize existing wheat disease data, weather data and specific vegetation layers from satellite data to develop deep learning- based models for early rust risk prediction; validate the high-risk areas using satellite remote sensing immediately after prediction; obtain granularized disease information by drone remote sensing or by field level human observations.
Approach:
In the past a collaboration on drone remote sensing and deep learning modeling for wheat stem rust disease detection and severity classification has been established. Results of this work has been published in a peer-reviewed article using a dataset with 960 trial plots detailing significant progress. The group plans to expand on that work by collecting data from several states and disease nurseries (i.e., not sprayed with fungicide). The multi-year, multi-state datasets will be analyzed to lay the foundation for the proposed work. We plan to develop a novel rust disease prediction model with multi-modal inputs including historical disease report data collected and maintained by the USDA, weather data including air temperature, precipitation, humidity, volumetric soil moisture, growing degree days, downward shortwave radiation flux, and satellite vegetation layers including evapotranspiration, normalized differential vegetation index (NDVI), red-edge spectral index and sun- induced chlorophyll fluorescence. Advanced deep learning networks such as Long Short-Term Memory (LSTM) will be trained using disease records, daily meteorological data, and existing satellite layers to generate a historical information embedding, and to predict the near-future disease risk by employing the forecasted meteorological data extracted from the Global Forecast System (GFS). If disease risk is predicted in a region (e.g. a county), satellite multispectral images will be retrieved for field-scale disease area detection. Depending on the size and location of the disease areas detected by the satellites, drone flights or local, in person observations will be deployed to capture granular details and validate the detection results. Further satellite imagery will be used for monitoring the progress of the disease after detection and intervention. The process is scalable as disease detection will improve along the increased data volume used for modeling. We will conduct initial data acquisition and analysis for three wheat production states: Minnesota, Kansas, and North Dakota, and expand the use of our predictive and detection models to other wheat production states.
