Updated January 18, 2022:
Project Report:
This research is a collaborative effort between weed science experts at Iowa State University and remote sensing experts at Montana State University (MSU). All experiments and imaging were conducted at ISU. Data processing and machine-learning algorithms were developed at MSU.
Experiments under artificial lighting:
Greenhouse and laboratory experiments were started in late fall 2020 to identify spectral reflectance of different biotypes of waterhemp plants resistant to ALS inhibitors, atrazine, and/or glyphosate. Seeds used in these experiments were collected in fall 2019 for the survey of herbicide-resistant waterhemp populations in Iowa. Waterhemp seeds were planted in 25 cm × 51 cm × 5 cm flats containing Sunshine Mix #1/LC1 potting soil (Sun Gro Horticulture, Agawam, MA, USA) with sand (4:1 ratio) in the ISU Agronomy Greenhouse. Individual waterhemp seedlings were then transplanted into long-narrow plastic cones. Transplanted seedlings were fertilized and watered regularly. Eight plants (two from each biotype), with a height of 7-8 cm and similar leaf numbers were selected for the hyperspectral imagery (Figure 1). With the remaining seedlings, a standard whole-plant dose-response bioassay was conducted to confirm the sensitivity of these biotypes for three herbicides, imazethapyr, atrazine, and glyphosate.
This allowed us to obtain reflectance spectra for different biotypes. A Pika L software (SpectrononPro) was used to process the images. Results obtained from the controlled experiments are summarized in Figure 2. The reflectance spectra represent the average and standard deviation of spectra obtained from each plant. The curves represent spectral reflectance of different waterhemp biotypes. As evident in Figure 2, significant spectral differences between biotypes would require a more sophisticated analysis to develop accurate classification systems.

Weeds were differentiated from soybean with >95% accuracy in the field (Figures 5 and 8). Waterhemp was differentiated from other weedy species with 81% accuracy. One-way resistance to glyphosate (EPSPS) and HPPD inhibitors was identified with 83% and 27% accuracy, respectively (Figures 6 and 7). Two-way resistance to ALS inhibitors + PS II inhibitors, and glyphosate + ALS inhibitors were identified with 16% and 37% accuracy, respectively (Figure 7). The three-way resistance to glyphosate + ALS inhibitors + PS II inhibitors was identified with 46% accuracy.
Results for HPPD resistance, two-way resistance, and three-way resistance were less consistent, and would require further evaluation to achieve classification accuracies >80%. For future research, classification accuracies of weed species and multiple herbicide-resistant waterhemp biotypes can be increased by integrating NDVI filters on hyperspectral images (Figure 9) and increasing training data collection of plant spectra.

View uploaded report

Overall, the results indicate that hyperspectral imaging and neural networks hold promise for early detection of herbicide-resistant weed biotypes in soybean production fields, especially glyphosate-resistant biotypes. This will ultimately lead to development of UAV-based weed maps for timely implementation of integrated weed management (IWM) programs for managing herbicide-resistant weeds in crop production fields.