top of page
Student Speakers
Remote sensing to identify early season annual ryegrass (Lolium rigidum) in wheat (Triticum aestivum)
Justina Serrano* (UWA Masters student), Kenneth Flower (UWA School of Agriculture and Environment), Nik Callow (UWA School of Earth and Environment), and Michael Ashworth (Australian Herbicide Resistance Initiative and UWA School of Agriculture and Environment)
Early intervention is the most cost-effective way to control annual ryegrass. For that purpose, we hypothesized that drone images could be used to detect early stages of ryegrass in wheat. Weeds were seeded in a field and in glasshouse pots to simulate different patch densities and growth stages. A drone equipped with an RGB and a multispectral camera was flown over both sites at different altitudes (15 to 120m). Pot images were interpreted on a visual scale and field orthomosaics were transformed into vegetation indices and compared. In the pots, seedlings could not be detected even at 15m because individual plants/leaves were too narrow with pixels representing mainly soil. During early-tillering, the impact of mixed-pixels (soil+plant) was lower because weeds formed a patch; therefore, imagery resolution could be reduced flying at 30m. These same principles were tested in the field, but drought and crop-competition reduced ryegrass establishment impeding its detection. Other broadleaved weeds grew in the site at higher densities; therefore, with a complete ground cover, bigger pixels and higher heights offered a better spectral separability of patches. The multispectral camera had less spatial resolution but a higher capacity to differentiate vegetation against soil compared to the RGB. Our findings showed some limitations of drone imagery to detect early-ryegrass stages and practical limitations when scaling this method into broadacre crops. Further research will need to confirm these results with pure-ryegrass patches within a field scenario where other factors (stubble, crop-competition) can affect image interpretation and reflectance of such small targets.
Breeding Program Optimization for Genomic Selection in Winter Wheat
Megan Calvert, Byron Evers, Xu Wang, and Jesse Poland Kansas State University
Developing methodologies in the fields of phenomics and genomic prediction have the potential to increase the production of crop species by influencing germplasm improvement. The integration of these technologies into germplasm improvement and breeding programs requires evidence that there will be a direct economic benefit to the program. We determined a basic set of parameters, such as prediction accuracy greater than 0.3, the ability to genotype over 7 lines for the cost of one phenotypic evaluation, and heritability levels below 0.4, at which the use of genomic selection would be of economic benefit in terms of genetic gain and operational costs to the Kansas State University (KSU) winter wheat breeding program. The breeding program was then examined to determine whether the parameters benefitting genomic selection were observed or achievable in a practical sense. Our results show that the KSU winter wheat breeding program is at a decision point with regards to their primary means of selection. A few operational changes to increase prediction accuracy would place the program in the parameter space where genomic selection would be of greatest benefit to the program.
Statistical machine learning methods for cotton yield classification in the US Southeast
Daniel Perondi, Clyde W. Fraisse, Nikolay Bliznyuk
The estimation of crop yields has been the focus of several studies, and different approaches to estimate crop yields have been established. The development of crop models that simulate the physiological processes of crops was one of the first approaches. However, crop models require detailed observed data for calibration and evaluation. Usually, detailed observed crop data is not available in spatial resolution. As an alternative for crop yield estimation, the use of statistical machine learning techniques may provide a close inference on spatial crop yields. Statistical machine learning methods offer benefits for the estimation of response variables without having to develop a mathematical model that represents the crop physiological processes. According to the literature, these methods have been showing good results, and they could provide reasonable predictions of crop yields. In this study, we evaluated machine learning methods for cotton yield classification as high, normal, and low yields in the US Southeast states of Alabama, Florida, Georgia, and South Carolina. Weather data and climate indicators summarized for every county were used as predictors. County yield from the USDA-NASS database for several years was retrieved to define high, normal, and low yields. Machine learning methods such as Random Forests, Bagging, Boosting, Generalized Additive Models, and Support Vector Classifier were evaluated. According to the results in this study, Random Forest had the highest accuracy in the test dataset for cotton yield classification.
Allometric models to predict plant biomass in maize: a comprehensive analysis using data from published field studies
Diego Hernán Rotili, Pedro Maximiliano Tognetti, Gustavo Ángel Maddonni
Plant allometry links ontogeny with physiological processes and underpins crop architectural modeling. Since Andrade et al. (1999), allometric models have allowed to predict vegetative (stem + leaves + tassel) and reproductive (ear) biomass of individual maize plants. While linear models have been locally adjusted, a cross-experimental estimation of fixed predictors and random effects originating from genotype (G), environment (E), and their interaction (GxE) is lacking. Using a comprehensive experimental database (21 site-x-year environments and 230 maize genotypes from 20 published papers), we estimated the fixed allometric predictors and random G, E, and GxE effects for vegetative and reproductive biomass. We run linear and nonlinear mixed-effects models, with 70/30 estimation/validation proportion. For vegetative biomass, the best model included stem volume as a fixed-effect predictor, with a quadratic term (R2marg./cond.=0.84/0.96; NRMSEmarg./cond.=0.33/0.21; n = 6630) which reduced slope variance by 34% (E), 12% (G) and 26% (GxE) compared to a linear model. In the quadratic model, E (74%), more than G (15%), or GxE (11%) explained slope variance. For reproductive biomass, the best model was an exponential regression with maximum ear diameter as a fixed effect (R2marg./cond. = 0.77/0.86; NRMSEmarg./cond. = 0.44/0.30; n = 2894). Using data’s central 80% to exclude extreme values, models did not improve with N-addition or hybrid/inbred information (𝚫AIC<2). Though in a reduced magnitude, broad environmental conditions, rather than genotype or local management, modulated maize allometric parameters. In conclusion, these models could be reliably applied under diverse research and production scenarios.
Keywords: mixed-effect models; maize; allometry; data management; genotype; environment; genotype x environment.
Mapping soybean-corn rotation utilizing remote sensing crop classification layers
Luan Pierre Pott; Telmo J. C. Amado; Raí A. Schwalbert; Ignacio A. Ciampitti
Crop rotation is one of the key principles of no-till farming systems, commonly linked to boosting soil health, and its spatial variation can contribute to optimize agricultural systems contributing to crop modeling, classification and yield forecast. The aim of this research is to characterize the soybean-corn rotation in northwest Rio Grande do Sul, (RS) Brazil. Crop rotation information was obtained through crop classification utilizing remote sensing [Sentinel 2, Sentinel 1 – SAR, and Digital Elevation (SRTM)] data associated with ground truthing achieved a 0.94 overall accuracy. As a candidate region, we have selected Não-Me-Toque city due to the large cropland area representing the Northwest region of the state of RS. Field boundaries of permanent agricultural fields from the Rural Environmental Registry (CAR), n=913 fields, were assessed to extract the centroid of each agricultural field associated with crop type information data for 2017-2018, 2018-2019, and 2019-2020 growing seasons. For crop rotation, soybean area represented the major crop in the three growing seasons measuring up to ~85%, while ~15% of the crop area was for corn crop. Continuous soybean fields (3-yr in a row) represented 68% of the fields, while continuous corn was only 0.1%. Crop rotation, evaluating the three growing seasons, with two soybean growing seasons achieved 30% of the crop areas, and just 1.9% of the areas with two corn growing seasons. The methods proposed to evaluate spatial crop rotation provided results that can be aggregated with crop modeling, classification and yield forecast implementing crop monitoring in regional scale.
Vegetable Variety Navigator: An Interactive Grower Decision-Support Tool
Ali Loker & Dr. Sam Wortman
Local vegetable variety trial results help growers make decisions about which varieties to plant on their farms. However, there are vegetable growers in nearly every county of Nebraska and every corner of the world, and researchers can’t keep up with the demand for location-specific variety performance data for so many specialty crops. We know every farm is different, but we also know that climate and soil are major drivers of crop adaptation and performance around the world. With that in mind, we set out to aggregate and analyze all publicly available variety trial data for three crops, to start. With this data, we developed the Vegetable Variety Navigator. The Vegetable Variety Navigator leverages data from nearly 300 variety trials of broccoli, cucumbers, and peppers across the world. We’ve analyzed these data to provide estimates of relative yield and quality potential for varieties included in at least three trials across distinct locations. The results of our analysis can be viewed through graphs or maps to explore geographic patterns in variety performance. The goal for this tool is to provide growers with information about variety yield and quality potential within and across specific locations. Using knowledge of their local climate and soil characteristics, growers can use information from this tool to make predictions about how a specific variety might perform on their own farm.
bottom of page