Progress 08/01/16 to 09/30/17
Outputs
Target Audience:1. Alfalfa Grower Community 2. Alfalfa seed companies 3. USDA Researchers 4. University Researchers 5. General Public Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?Although this project is not intended or designed for training or professional development, some training and development opportunities were realized. 1) Under the guidance of USDA ARS scientists, students/young researchers and technicians gained experience in monitoring the alfalfa plant progression and developing techniques for estimation of plant floral resources. 2) Under the guidance of Kairosys, undergraduate students seeking careers in Agriculture gained valuable experience and professional development. How have the results been disseminated to communities of interest?Agronomists: Kairosys presented results from the study to agronomistswithin the S&W (strategic partner) network. Growers: Kairosys and S&W Agronomists presented partial results to participating growers. The participating growers were already using the Incubation monitoring solution developed by Kairosys. The impact of the holistic solution combining the incubation product and the bloom prediction product was disseminated and received well by the grower community. What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts
What was accomplished under these goals?
IMPACT The purpose of this proposal is to conduct research that will provide alfalfa seed growers a pollination management tool that enables synchronization of Alfalfa Leafcutting Bee (ALB) release with prediction of alfalfa bloom. These protocols will help increase yield and improve bee health and recovery. Most U.S. growers outside of California use ALBs for pollination because they increase yield by more than 50% over honey bees. Growers deploy ~$400/acre worth of ALBs. Kairosys' customer diligence indicates that a further 250 lbs/acre yield can be gained by responding adequately to changing field bloom conditions. Kairosys will develop a heuristic for alfalfa bloom prediction using data at grower sites, USDA plots, and its commercial partner's greenhouse. Employing imaging and machine learning techniques, we will track stages of plant development and compare spectral images with manual observation. Sensors deployed to calculate the Growing Degree Days and photoperiod needed through the development progression will enable predictive models for bloom onset using bud development markers as a basis. Sensors also track weather, weed, and pest infestations to provide better control windows for bee release and to localize and reduce pesticides. When commercialized, growers will receive a reliable forecast for bloom onset and bee emergence through a smartphone application. This will roughly double the management window for growers to make the two most important pollination decisions; 1) bee release timing to synchronize with bloom and 2) pesticide application, resulting in significantly increased yield and revenue while simultaneously improving bee health. Accomplishments: The primary goal in Phase I was to create a preliminary heuristic model for understanding bloom progression and prediction of bloom onset. Our technical objectives were to explore: Aerial imaging using small Unmanned Aerial Systems (sUAS) of alfalfa plots Ground-level imaging of alfalfa test plots using RGB and multispectral camera Environmental monitoring using ground sensors - temperature, humidity, photoperiod Image analysis to create a heuristic model for onset of bloom Predictive models for onset of bloom To accomplish these objectives, the Kairosys team gathered a variety of data (sUAS, RGB, NIR, Hyperspectral, Environmental sensors, Manual Floral counts, and seed yield) from multiple locations (three grower sites, two greenhouse experiments, and two USDA plots). At the request of S&W and participating growers, we included yield measurement and the associated correlation to bloom prediction models in Phase I studies (originally targeted in Phase II). Color threshold modeling of bloom progression: RGB images from USDA plots (2 plots, 4 sites each, 14 days), grower fields (3 locations, ground and aerial, 35 days), and S&W greenhouse (39 locations, 40 days) were used to compute a bloom intensity based on pixel counts meeting a color thresholding criteria of hue and saturation (RGB color thresholding model fit for flower counts (270 0.5; v > 0.5)).R statistical software with imager and caret packages was used for this analysis with 30% of the data as training samples. The Adjusted R2 for the test sample is ~ 0.80 with the residuals showing mostly underestimation of predicted values. At high manual floral counts, this could be due to flowers being shielded from the camera by vegetation. In one growerfield, two bloom levels were simulated by cutting back 100 ft2 in 4 plant rowsand designating an adjacent 100 ft2 as control. The bloom in the cutback cohort was delayed by ~15 days. Pro-rated yield for the cutback region was 224 lbs/ac, while the uncut region yielded 592 lbs/ac. For two other uncut fields,yields of 1303 lbs/ac and 362 lbs/ac, respectively, were reported. The predicted bloom intensity using our models show a 3x difference andcorrelate well with thedifference in measured yields. The estimation of bloom progression from the color thresholding model as applied to three grower fields, shows that when the bees are introduced and pollination begins, there is a rapid reduction in flower number. At this point there is an interplay of the loss of flowers and the creation of new flowers, which could translate directly to yield. If the new flower production is not sufficient to sustain the bees, then bees could drift away (Pitts-Singer 2013b). If pollination occurs too rapidly due to surfeit of bees, resulting stress could shut down subsequent flower production (Carlson 1928, Free 1993, Strickler 1999). Hyperspectral imaging of leaves, buds and flowers: Significant spectral differences (using ASD 350-2500 nm with contact probe) are observed among isolated targets of leaves, buds, and flowers. These include samples from two locations (grower site and greenhouse) collected over two days, and sample variance is minimal. These observations indicate that Shortwave Infrared (SWIR) bands may not be necessary, and imaging bands in the visible and NIR will be sufficient to delineate several predictive wavelengths and vegetation indices. This is a promising result, since SWIR filters are significantly more expensive than visible/NIR filters. Based on this preliminary finding, the hsdar R package was used to derive several light-use efficiency, leaf pigment, narrowband greenness, and broadband greenness vegetation indices (Section 8). Figure 6 shows four of the many indices that with significant differences in the response to leaves, buds, and flowers, namely, Normalized Difference Vegetation Index (NDVI), Structure Insensitive Pigment Index (SIPI), Anthocyanin Index (ACI), and Red-Edge Position (REP). NDVI and ACI indices obtained from grower fields show critical differences in plant development and senesence. Established stands showed much higher NDVI vs spring seeding at the beginning of the season and fields with greater seed showed a rapid decline of NDVI later in the season, which could be used as a signal for harvest. The ACI indices used in combination with NDVI showed a strong correlation to the differences in yield between grower fields. Aerial Imaging: A DJI Phantom 3 small Unmanned Aerial System (sUAS) was used to capture 12 Mp RGB and NIR-GB images at a height of 35 ft at two grower sites on 35 days from June 6 to August 11. In addition, RGB images were obtained in "hover" mode over a specific spot that included a reflectance marker. The color thresholding model applied to the hover images shows strong correlation (>0.9 R2) to bloom intensity predicted by ground images, while those at 35 ft did not possess adequate resolution for predicting bloom intensity. The NIR-GB images are useful in predicting the progress of green vegetation and are able to detect post-pollination senescence in the crop. Machine learning algorithms for bud and bloom detection: Multispectral and Hyperspectral data was used in conjunction with RGB images to develop a supervised machine learning algorithm to detect bud and bloom progression. Sample point software was used to count plant features and a PLSR algorithm (written in R software) was used to extract the principal wavelengths of interest from the multispectral and hyperspectral data. This work will further be fine tuned with data from the subsequent seasons. The output will be to create unique filters which respond to alfalfa plant phenology. Conclusion Models generated from imaging can predict the progression of bloom in the alfalfa plant. Ongoing work will strengthen these models and also include the prediction of progression of buds. Bloom differences between fields predicted by the models are accurately able to predict differences in yields between the fields. Hyperspectral data can be used to get unique combination of wavelengths for creating custom filters to monitor plant progression
Publications
|