Source: WESTERN MICHIGAN UNIV submitted to
SYNOPTIC WEATHER FORECASTING AND WEB-BASED INFORMATION DELIVERY SYSTEMS FOR MANAGING CROP DISEASE RISK IN MULTIPLE REGIONS OF THE U.S.
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
TERMINATED
Funding Source
Reporting Frequency
Annual
Accession No.
0215069
Grant No.
2008-51101-19477
Project No.
MICR-2008-02925
Proposal No.
2008-02925
Multistate No.
(N/A)
Program Code
112.B
Project Start Date
Sep 1, 2008
Project End Date
Aug 31, 2013
Grant Year
2008
Project Director
Baker, K. M.
Recipient Organization
WESTERN MICHIGAN UNIV
(N/A)
KALAMAZOO,MI 49008
Performing Department
(N/A)
Non Technical Summary
The project will create weather-based disease risk forecasts for crop diseases in various regions of the U.S. We will specifically focus on leaf spot of peanut in Georgia and northern Florida, Fusarium head blight of barley in the northern Great Plains, and late blight of potato in Michigan. The resulting forecasts will be delivered through web-based systems and will be available to growers on a daily update basis. Risk forecasts can improve crop quality while at the same time reducing fungicide use by allowing growers to improve the timing of fungicide applications. Our goals in reducing fungicide use are of increasing product quality, limiting expenditures, and reducing the amount of chemical released to the environment. We will examine and quantify the accuracy, economic and environmental impacts, and usability of crop disease risk forecasts at both the synoptic and mesoscale (different weather forecasting scales) in the various regions of the U.S. By selecting three very different crop species (peanut, barley, potato) each heavily dependent upon fungicide sprays for the prevention, avoidance, and management of disease in various regions of the country (southeast, northern Great Plains, Great Lakes region), we plan to capture the overall potential of such forecasting systems for widespread use in the U.S. New technologies in the form of improved National Weather Service forecasts (since 2004) and access to high performance computer workflows give us the opportunity to exploit these advances funded through large government technology grants (the LEAD project alone cost $11.5 million) for the benefit of both major and minor agricultural commodity groups. The ultimate goal of any forecast system is reduction of uncertainty that can negatively influence decision making by users. In the case of disease risk forecasting for agriculture, decisions by the target group of users can negatively impact our food supply, environment, and economy by increasing use of pesticides. Increasing accessibility to forecast information through free, thoroughly tested forecast services in the web environment has the potential to dramatically decrease uncertainty in a variety of crop systems. In this project we examine three crops in three regions of the country, but the implications of the project's results are much more far reaching. Many crop diseases have similar meteorological triggers based on temperature and moisture requirements that may be mined from available forecasts. For a number of years, growers, stakeholders , commodity associations, crop consultants and land grant specialists have been requesting disease risk prediction models for all the major crops in the U.S. through Pest Management Strategic Plans. The public availability of weather forecasts and the improved infrastructure to promote the use of these forecasts make this the time to examine usefulness of forecasting as a national priority in agriculture.
Animal Health Component
60%
Research Effort Categories
Basic
25%
Applied
60%
Developmental
15%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
21213101160100%
Knowledge Area
212 - Pathogens and Nematodes Affecting Plants;

Subject Of Investigation
1310 - Potato;

Field Of Science
1160 - Pathology;
Goals / Objectives
The overarching goal of this proposal is to reduce the impact of weather sensitive diseases on profitable crop production in the US, while at the same time reducing fungicide input. We will do this by developing disease specific synoptic and mesoscale weather forecasts coupled with delivery systems designed to expand the options for integrated prevention, monitoring and suppression of disease. Synoptic forecast models (120 hours) for leaf spot of peanut in the southeast U.S. and Fusarium head blight of barley in the northern Great Plains will be developed, implemented and validated using a prototype developed for late blight of potato in Michigan. A prototype will be developed of a related mesoscale gridded forecast models (48 hours). Feasibility of mesoscale forecasting at various spatial and temporal scales will be examined. Concepts will be tested for a four week period of high disease risk in 6-10 county subregions of regional importance in Georgia and South Dakota. Web delivery and evaluation of web delivery and stakeholder use will focus on synoptic forecasts. Development of methods and workflows to operationally serve mesoscale forecasts will be tested in Michigan only, but the feasibility of operationalizing such forecasts at various scales and by various methods will be examined for each region. We will assess spatial and temporal patterns in accuracy, variability, environmental and economic benefit among crops and regions. Our proposed collaboration among institutions focused on varied regions and crop types provides an research setting which will in enable us to make broad recommendations regarding the adoption, use and economic and environmental implications of crop disease specific weather forecasting systems.
Project Methods
An artificial neural network (ANN) synoptic crop disease forecasting model developed by Baker and Kirk (2007) has proven successful in predicting potato late blight risk as estimated by potato late blight disease severity values. The initial phase of the current project will involve developing similar neural network models to interface with the daily NWS model output statistics for automatic updates of leaf spot of peanut and Fusarium head blight of barley grain values. While the usefulness and accuracy of synoptic scale predictions have been proven by our prior research, mesoscale forecasting is a newly available option to those entities who do not maintain their own networks of high performance computers. Linked Environments for Atmospheric Discovery (LEAD) makes meteorological data, forecast models, and analysis and visualization tools available to anyone who wants to interactively explore the weather as it evolves. The LEAD Portal brings together all the necessary resources at one convenient access point, supported by high-performance computing systems. This type of environment is ideal for gaining access to large amounts of data and computational power not traditionally associated with plant pathology. ANN models similar to those used at the synoptic scale will be incorporated with mesoscale LEAD outputs and spatially linked through a geographic information system (GIS). Initial validation of predictions made with both synoptic and mesoscale models will be examined through comparison of predicted disease risk with those computed based on the Unedited Local Climatological Data (ULCD) for 2005-08 growing seasons available in archived form from the National Weather Service. Spatial and temporal variability in risk predictions and associated model accuracy will be evaluated using standard statistical procedures and spatial statistics. Impacts of additional variables such as length of archive, regional climatic normals, and seasonal variation will also be explored. Field validation will be consistent with standard practices in plant pathology. Disease risk forecast results will be taken into account during field monitoring and fungicide spraying at research farms. Research sites and farmer's fields will be selected and monitored for disease. The data collected from these sites will be used to evaluate the accuracy of the disease model predictions. Economic and environmental impact assessments will be used to estimate the reduction of financial and chemical inputs to the various cropping systems as a result of access to model information. Spatial and temporal variability in weather patterns, as determined from regional climatic normals, will be examined with respect to spatial and temporal variability in forecast accuracy, resulting in uncertainty metrics for both the predicted and actual patterns for a region. A comparison of these metrics will allow some quantifiable measure of the reduction of uncertainty resulting from model use.

Progress 09/01/08 to 08/31/13

Outputs
Target Audience: Our forecasting efforts have focused on crops which have traditionally required large fungicide inputs and for which growers, stakeholders, and extension workers have recognized the need for early-warning systems and have specifically requested development of such systems. The Michigan Potato Industry Commission (MPIC) which represents potato growers in Michigan have pushed for disease forecasting systems since 1999 (Ben Kudwa, Director MPIC, pers. comm.) when forecast models were not yet accurate enough for extensive use in agriculture. Similar support has been provided by the Georgia Agricultural Commodity Commission for peanuts. Extension pathologists in South Dakota, North Dakota, and Minnesota have received requests for barley Fusarium head blight information and management tools (pers. comm.) indicating the need exists for tools such as risk advisory/forecast systems. These potential users, including growers, extension agents and the like, have been continuously updated since the beginning of the grant through articles in commodity and extension specific publications and via an active twitter feed. From a broad research perspective, agro-ecosystem forecasters of all types are an audience to our results implementing National Digital Forecast Database data for specific stakeholder groups. Widespread availability of quality forecast weather data, climate models, geographic information systems technology and large scale computing has democratized the ability to create regional forecasts for disparate groups but has also increased the risk that accepted models are deployed directly on new datasets without retraining. An active effort has been made to disseminate relevant findings to the geography, plant pathology, agro-geoinformatics and extension communities through presentations at national and international conferences and submission of peer reviewed publications in each of these disciplines. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided? A number of students have been trained in environmental modeling, computer programming and integration of geographic information systems, plant pathology basics, and publicly available weather datasets. This training has resulted in three thesis projects, training of another 9 graduate students, a number of undergraduate research awards, one PhD. Professional development for students has included the opportunity to make national conference presentations and attend National Science Foundation workshops in cyberinfrastructure. Numerous publications in professional journals also have included graduate and undergraduate students as authors. How have the results been disseminated to communities of interest? During the 2013 growing season, both the Michigan potato late blight website hosted by Michigan State University and Dakotas Fusarium head blight of barley website hosted by North Dakota State University allowed access to daily updates of crop disease risk forecasts at various point locations throughout the region. The sites used the most recent incarnation of artificial neural network forecast models, which rely on 10 years of past data for verification. Access to gridded forecasts, based on the National Digital Forecast Database (NDFD) will begin during the 2014 growing season. Potential users, including growers, extension agents and the like, have been continuously updated since the beginning of the grant through articles in commodity and extension specific publications and via an active twitter feed. An active effort has been made to disseminate relevant findings to the geography, plant pathology, agro-geoinformatics and extension communities through presentations at national and international conferences and submission of peer reviewed publications in each of these disciplines. What do you plan to do during the next reporting period to accomplish the goals? Nothing Reported

Impacts
What was accomplished under these goals? Publicly available weather forecasting datasets have improved dramatically in skill and spatial resolution in the last decade. Significant changes in data sources have been made even since the inception of this project 5 years ago. Our overall goal has been to understand the ways in which these data sources can most effectively be used to implement integrated pest management (IPM) strategies, reducing pesticide use and the associated negative impacts on human health and environmental quality. We have followed advances in weather forecasting and developed increasingly accurate models based on those forecasts to make 5-day crop disease specific risk forecasts in the United States. We have also examined the most effective web-based information delivery systems for direct use of those forecasts by growers of economically important food crops. As a proof of concept that such a system will be beneficial in a variety of cropping systems throughout the U.S., we focused on development of web delivered forecasts for Fusarium head blight of barley in the northern Great Plains and late blight of potato in Michigan. Model calibration and comparison of model skill and data flows for different regions, such as the potato late blight model in the Great Lakes, New England, Red River Valley and Snake River regions of the US, also were used to understand limitations to consistent provision of reliable forecasts to growers at a wider scale. As a final step, implications of climate change, climatic oscillations and low frequency severe weather events (hurricanes) on the crop disease risk landscape of the US were examined. Each of these impacts the long term accuracy of forecasting, as newer datasets have less archived data available for model creation and implementation. Developed models ranged in accuracy from 75-85 percent and decreased stakeholder uncertainty by upwards of 50 percent. Project objectives 1.Develop, implement, and validate synoptic forecasting models. 120 hour risk forecast models for each of 3 crops and regions were developed at National Weather Service point forecast locations and implemented on regional crop servers. Accuracy of all models was over 75 percent and deemed acceptable for use by growers. Artificial neural network models proved most accurate when past weather, forecast weather and some indication of ‘normal’ weather (climate normals) were included as input variables. Results indicate that retraining and re-evaluating an operational model every few years in light of a longer data archive is far from a fruitless exercise. Although potato late blight risk forecast models had been thoroughly trained and had been accessible online for several years in Michigan, dramatic improvements were made when an extended dataset was available for testing and validation and a greater understanding of input variables, gleaned from experience with models in other crops and regions. The increase in accuracy on days with conditions conducive to disease development was particularly striking. Accuracy on the first forecast day nearly doubled and by the fifth forecast day more than tripled in forecast quality with the new model. These results are valuable to two groups general of stakeholders. From the perspective of potato growers in the region, improvements are most dramatic on days that are conducive to disease development. From the perspective of agro-ecosystem forecasters, this is one of many steps to understanding how an increased archive length and frequent reassessment of input archives can benefit model development. 2. Develop, implement, and validate mesoscale gridded forecast models. Gridded forecast models were developed and validated for late blight of potato in three large regions: Great Lakes, New England, and Pacific Northwest. Fusarium head blight of barley model was developed and validated in the northern Great Plains. These forecasts will be accessible to stakeholders online beginning with the 2014 growing season. From model development we learned broad lessons about integrating the National Digital Forecast Dataset into agroecosystem models. As the pace of new data availability and access to cyber-infrastructure increases, weather data inputs to practical application models have gone from point data to raster grids of varying spatial and temporal resolution. Certainly there is a benefit to widespread access of data, but transforming models developed at point locations to raster datasets is not trivial. This portion of the project served as a test case for examining larger patterns in the seemingly endless struggle to keep models functioning accurately with ever-new data availability. Significantly different results were obtained for the varying regions of the country with the same training techniques. Our results show that retraining and revalidating models on new datasets is important, regardless of how simple or how complex the initial model may have been. This constant upkeep of historically developed models has real costs that must be considered by model developers and funding agencies that support long term decision support systems, which are moving toward new datasets with increasing frequency. 3. Develop a web portal for synoptic and mesoscale forecasting of crop disease risk management which will integrate outputs from models. Regional groups of plant pathologists and extension agents decided that it was more sustainable in the long term for forecasting to be controlled and served regionally rather than through a centralized portal. However, with the increase in available data and cyberinfrastructure that enable regional land grant centers to maintain their own web sites comes the problem of a system design that is expandable and portable. This problem has become a key component of web access to weather-based agroecosystem risk forecasting models which did not exist at the grants inception. We embraced this question by comparing sustainable storage system designs. Three versions, including a traditional relational database management system (PostgreSQL), a NoSQL database system (MongoDB) and a scientific file format version (netCDF), of a single crop disease risk modeling system in one region of the country were designed and compared for speed. To test expandability, another crop disease risk modeling system was added. The models differ in their timespans and spatial areas, resulting in the retrieval and processing of more data for the second model. Speeds for the three types of systems were fairly similar. If overall system speed is paramount, scientific data formats like netCDF remain a preferred storage solution. If research personnel resources are small, PostgreSQL libraries use simple, swift code with easy interfaces. If creative model development and spatio-temporal flexibility of modeling inputs are a priority, NoSQL options such as MongoDB allow for the most flexibility in stored data contents. 4. Assess spatial and temporal patterns in accuracy, variability, environmental and economic benefit among crops and regions. Crop risk forecasting systems increasingly must be dynamic, enabling new datasets to be exploited in those spatio-temporal situations when they exceed the accuracy of prior data. Promoting widespread understanding of agroecosystem forecasting under a variable climate and distributing relevant, transferable methods is key in enabling innovative adaptation and mitigation for unique local to regional scale systems. The key component of this objective not addressed in the previous is the examination of environmental and economic benefits. As conditions for many crop diseases are exacerbated by climate change, the main benefit of forecasting systems becomes a reduction in grower uncertainty, leading to less reactionary practices and more strategic use of IPM. We determined that our five day forecasts on average reduced the seasonal area of risk (AOR), a measure of uncertainty, by 58 percent.

Publications

  • Type: Conference Papers and Presentations Status: Published Year Published: 2013 Citation: Nogueira, R. and K. Baker. 2013. Tropical Cyclone Events and Late Season Disease Risk Triggers: The Case of Late Blight in the Eastern US. AAG Annual Meeting, 9-13 April, Los Angeles, CA.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2013 Citation: Roehsner, P and K. Baker. 2013. Data Storage Alternatives for a Gridded Crop Disease Risk Forecasting System. AAG Annual Meeting, 9-13 April, Los Angeles, CA.
  • Type: Journal Articles Status: Under Review Year Published: 2014 Citation: Kathleen M Baker *, Ricardo Nogueira, Krishna Bondalapati, Cassandra Hoch, Jeffrey Stein. Climate Variability and Risk Of Deoxynivalenol Accumulation in U.S. Wheat and Barley: A Rationale for Early Warning Systems. Toxins, Special Issue: Recent Advances and Perspectives in Deoxynivalenol Research
  • Type: Theses/Dissertations Status: Other Year Published: 2013 Citation: Roehsner, Paul. Data Storage Alternatives for a Gridded Crop Disease Risk Forecasting System
  • Type: Journal Articles Status: Submitted Year Published: 2014 Citation: Baker, K, Roehsner, P, Lake, T., Rivet, D., Benston, S., Bommersbach, B., and W. Kirk. Point-trained models in a grid environment: Transforming a potato late blight risk forecast for use with the National Digital Forecast Database. Computers and Electronics in Agriculture. Computers and Electronics in Agriculture
  • Type: Journal Articles Status: Submitted Year Published: 2014 Citation: Bondalapati, K.D., Stein, J.M., and K. M. Baker. A Neural Network Model to Predict Deoxynivalenol (DON) in Barley using Historic and Forecasted Weather Conditions. Plant Disease.
  • Type: Journal Articles Status: Awaiting Publication Year Published: 2014 Citation: Baker, K.M., Lake, T. Improved Weather-Based Late Blight Risk Management: Comparing Models with a Ten Years of Forecast Data Archive. Journal of Agricultural Science.
  • Type: Journal Articles Status: Submitted Year Published: 2014 Citation: Roehsner, P. and K. Baker. Sustainable system design for gridded, spatio-temporal, agroecosystem forecasting models. Computers, Environment and Urban Systems.


Progress 09/01/11 to 08/31/12

Outputs
OUTPUTS: In year 4 of the grant, we continued to work on the development of modeling and implementation techniques for crop disease risk using new digital gridded datasets, and continued the dissemination of point location results for all models. Our initial examination of techniques for using National Digital Forecast Database and Gridded Model Output Statistics datasets became a full blown analysis of system design and function for sustainable use in the cloud, long term. We determined best practice methodologies for the storage and processing this data by comparing a typical GIS system using standalone files, a system running a Relational Database Management System (RDBMS), a similar system running MongoDb, and finally using netCDF files as another alternative will be created. Michigan growers, extension agents, crop consultants and others who are interested in tracking the threat of potato late blight in Michigan can still follow point based model results from lateblight.org front page. In 2012, we refined the model used to do the forecasting so that it is more robust to long term climate trends in the region. The new model began forecasting and has been widely accessed via the internet beginning in May 2012 accompanied by a new color coding of markers on the forecasting map. In 2011-2012 we also developed a new section of the website which forecasts the chance of volunteer potato survival over the winter. Using soil temperature data from Michigan Agricultural Weather Network stations we developed a model that can accurately predict the risk of survival of volunteer potato tubers over winter. This model and explanation are now available on the lateblight.org website at www.lateblight.org/volunteer-risk.php. In addition, disease forecasts are posted on Twitter on a weekly basis or when the risk of a late blight outbreak is high. To date the late_blight Twitter account has been receiving on average 1 new follower a week and we now have over 128 followers. A lot of these followers are crop consultants who can then disseminate the information to their growers and clients. The Northern Great Plains Fusarium head blight model was further refined with 2011 and 2012 datasets. The NDSU barley pathology program is developing the FHB risk forecasting website and will have it up and operational for the 2013 growing season. A prototype web page displaying the five day FHB risk forecast and five day validated risk based on data from NDAWN weather station was successful. Markers are displayed for each weather station site across North Dakota are shown on a map and users can point and click on the marker to display daily and forecasted risk information for their growing region. In a similar way, the online spray advisory for leaf spot of peanut continued to be incorporated into the University of Florida AgroClimate site. We continue to improve the leaf spot of peanut forecasts and understand the interplay of leaf spot weather requirements and weather forecast variables. Initial steps have been taken to implement all three disease models (potato late blight, Fusarium head blight, leaf spot of peanut) in the grid based system using newly available datasets. PARTICIPANTS: Kathleen M. Baker, Western Michigan University, is the principal investigator on the project. Postdoctoral research assistant, Ricardo Noguiera, and researcher Susan Benston assisted with analysis and publication at WMU. Graduate and undergraduate students (Thomas Lake, Douglas Rivet, Paul Roehsner, Jason Smith, Jason Wengert, Joshua Williams) worked as research assistants on the project and received training in statistics, spatial analysis, and computer programming at Western Michigan University. William Kirk is the project director for Michigan State University and is assisted by Lee Duynslager. Jeffrey Stein resigned as the project director for South Dakota State University, passing that responsibility to Thomas Cheesbrough. Dennis Todey is an additional PI at SDSU and Krishna Bondalapati is a Research Associate II and doctoral student at South Dakota State. Mark Boudreau has taken over as project director for University of Georgia from Joel Paz and Gerrit Hoogenboom. Abraham Fulmer is a graduate student at UGA, working closely with Robert Kemerait, Joy Carter, and Ankit Arte. Robert Brueggeman is the PI on a subcontract with North Dakota State University. Phillip Wharton, University of Idaho, received summer salary as a consultant. Rabiu Olatinwo completed his postdoctoral research at University of Georgia. Beth Plale at Indiana University was a collaborator on the project, allowing us to implement workflows through the Linked Environment for Atmospheric Discovery (LEAD) portal. Robert Trenary is a consultant for neural networking solutions at Western Michigan University. Ilya Zaslavsky is a consultant on the project, providing us with data backup and archiving solutions. Three undergraduate students and four graduate students received training in statistical techniques, GIS, and the use of publicly available weather and climate data sources in mitigating the risk of crop disease. A graduate student also was trained on NetCDF file format at an NSF sponsored Unidata workshop at NCAR and on PySAL at GIScience. TARGET AUDIENCES: Our extension efforts aim to change the knowledge of growers as they make management decisions that impact actions such as fungicide spray quantities and frequencies. The forecast sites for potato late blight in Michigan and leaf spot of peanut in the Southeast US were both live this growing season. The most current models at point locations provided data for decision support. The site for Fusarium head blight in the N Great Plains will go live next growing season. The late blight of potato forecasting model is now more accurate at predicting risk and with the addition of color coded markers on the map visitors can get an instant overview of late blight risk in MI. A new volunteer risk model also provides growers with a new tool to help reduce the risk of a late blight outbreak. Growers can effectively manage volunteers in rotational fields (historical and current) and initiate early intervention crop practices such as seed treatment for late blight and early application of fungicides. In addition, disease forecasts are posted on Twitter on a weekly basis or when the risk of a late blight outbreak is high. People following the website on Twitter receive an SMS message to their mobile phone as soon as the Twitter message is posted. This is an excellent method of us to disseminate results and warnings to growers, extension agents and crop consultants in MI and elsewhere in the world. To date the late_blight Twitter account has been receiving on average 1 new follower a week and we now have over 128 followers. A lot of these followers are crop consultants who can then disseminate the information to their growers and clients. Research results have also been disseminated through state wide grower communities and national and international agricultural research conferences, especially in the field of agro-geoinformatics. PROJECT MODIFICATIONS: Nothing significant to report during this reporting period.

Impacts
Analyses performed in this grant year and lessons learned across spatial regions and various diseases improved general knowledge of the ways in which publicly available weather and climate data sources can be used to forecast crop disease effectively. Substantial progress was made in understanding the artificial neural network model training processes that relate to forecast accuracy and the influence of longer term data sets (10 years or more) on overall accuracy. This type of interaction is particularly important under climate change scenarios, and we began examining at the influence of long term trends like ENSO and AMO, as well as hurricane frequency, as influences on the long term disease risk variability. Risks associated with Fusarium head blight of small grains and late blight of potato, are assessed for the sixty year period 1950-2009 with respect to these aspects of climate variability. Each of these aspects of variability was significantly related regionally to daily crop disease risk forecasts. We also learned much about implementation of our previously developed models using the national digital forecast dataset and continue to contrast technological solutions to best disseminate resulting forecasts to growers. Certainly there is a benefit to widespread access of new data, but transforming models developed at point locations to raster datasets is not trivial. The dramatic improvements that can be made to models when an extended dataset is available for testing and validation are not always possible in an era of quickly changing datasets and modeling techniques. Each of our changes in knowledge changes the conditions under which growers will manage crop disease risk. By reducing grower uncertainty, we reduce the reactionary tendencies of those individuals placed in high risk situations and bring about a steady change toward action that relies on current and accurate forecast information. The new volunteer risk model also provides growers with a new tool to help reduce the risk of a late blight outbreak. Growers can effectively manage volunteers in rotational fields (historical and current) and initiate early intervention crop practices such as seed treatment for late blight and early application of fungicides. In 2012 for example, although volunteer risk was predicted using our system and late blight had been present in 2011, growers saved crops from late blight by practicing early intervention and no late blight was reported in Michigan in 2012. Stored potatoes are therefore free from late blight potentially saving losses which in severe years can by approximately $20 million. In particular we have found that extension agents use our online tools to better inform their clients and other agents in a particular region. In the Great Plains, we now have the ability to run the forecasting model from our NDSU servers and will have it linked to the barley pathology web site for easy access to producers and industry consultants. We are designing the one page extension bulletin that will be published and posted online to promote the barley FHB forecasting web site.

Publications

  • Kirk, W., Duynslager, L., Wharton P., Baker K. and B. Bishop. 2012. New website for estimating potential survival of potato volunteers in Michigan. http://msue.anr.msu.edu/news/new_website_for_estimating_potential_sur vival_of_potato_volunteers_in_michigan
  • Kirk, W., Duynslager, L., Wharton P. and K. Baker 2012. Potato late blight risk monitoring website updated July 2012. http://msue.anr.msu.edu/news/potato_late_blight_risk_monitoring_websi te_updated_july_2012
  • Kirk, W., Duynslager, L., Wharton P., Baker K. and B. Bishop. 2012. Updates to the five-day potato late blight forecast website. http://msue.anr.msu.edu/news/updates_to_the_five-day_potato_late_blig ht_forecast_website
  • Kirk, W., Duynslager, L. and K. Baker 2012. Five-day potato late blight forecast now available for Michigan. http://msue.anr.msu.edu/news/five_day_potato_late_blight_forecast_now _available_for_michigan
  • W. Kirk, and L. Duynslager, P. Wharton and K. Baker. 2011. Multiple resources will help you track potato late blight http://msue.anr.msu.edu/news/multiple_resources_will_help_you_track_p otato_late_blight
  • Kirk, W. 2011. Potential survival of potato volunteers in Michigan http://msue.anr.msu.edu/news/potential_survival_of_potato_volunteers_ in_michigan
  • Everman, W., Long, C. and W. Kirk. 2010. Volunteer potatoes http://msue.anr.msu.edu/news/volunteer_potatoes
  • Wharton, P.S and W. Kirk. 2009. Effect of climate change on the over winter survival potential of volunteer potatoes in Michigan. American Journal of Potato Research. DOI 10.1007/s12230-008-9064-9 http://www.lateblight.org/pdf/PAA-2008-volunteer-survival.pdf
  • Baker, K.M., Lake, T., Roehsner, P. and K. Schrantz. 2012. Forecasting Disease with 10-Year Optimized Models: Moving Toward New Digital Datasets. The First International Conference on Agro-geoinformatics 2-4 Aug, Shanghai, China.
  • Baker, K.M. and R. Nogueira. 2012. Multi-Decadal Variability and the Daily Agroecosystem Forecasting Accuracy Plateau. The First International Conference on Agro-geoinformatics 2-4 Aug, Shanghai, China.
  • Bondapalati, K., Stein, J., and Baker, K.M. 2012. Neural Network Model to Predict Deoxynivalenol (DON) in Barley using Historic and Forecasted Weather Conditions. The First International Conference on Agro-geoinformatics 2-4 Aug, Shanghai, China.


Progress 09/01/10 to 08/31/11

Outputs
OUTPUTS: Results from preliminary analyses in Years 1 and 2 of the project led to real breakthroughs in modeling techniques, implementation and dissemination in Year 3. We established techniques for transferring modeling techniques and accuracy measures from previously limited datasets (WRF runs through LEAD on Teragrid) to the new national forecasting standard grid-based datasets. National Digital Forecast Database and Gridded Model Output Statistics datasets were archived throughout the 2011 growing season. These datasets will be used for testing grid implementation of the forecasting models for potato late blight model in the Great Lakes region, the Fusarium head blight model in the Great Plains, and the leaf spot of peanut model in the Southeast during Year 4. Michigan growers, extension agents, crop consultants and others who are interested in tracking the threat of potato late blight in Michigan can now receive a "5-Day Forecast" on the risk of a late blight outbreak by visiting the neuroweather website (www.lateblight.org/neuroweather.php). In addition, disease forecasts are posted on Twitter on a weekly basis or when the risk of a late blight outbreak is high. People following on Twitter receive SMS messages to their mobile phones. To date the late_blight Twitter account has been receiving on average 1 new follower a week and we now have over 94 followers, mostly crop consultants who can then disseminate the information to their growers and clients. Site statistics have shown that we received over 31,000 visits in the past year, mostly during the growing season. A breakdown of the traffic statistics shows that the largest number of visits is coming from Michigan which strongly suggests that we are reaching our target audience of Michigan growers, crop consultants and extension agents during the growing season. Usage data has also shown that the number of visits to the website from mobile devices has increased exponentially over the past year. Success of the initial potato late blight neuroweather site in Michigan is being mirrored by other parts of the project. The Fusarium head blight of barley model was posted to the North Dakota State University forecasting site during the 2011 growing season and analysis of site statistics is underway. The online spray advisory for leaf spot of peanut is being incorporated into the University of Florida AgroClimate site for dissemination throughout the 2012 growing season. Training sessions for extension agents and other stakeholders in all three regions discussed the importance of weather/climate forecasts in promotion or anticipation of forecast releases in the coming year. Research updates were also disseminated to growers via Newsline, and to international researchers pursuing similar results at the World Congress of Computers in Agriculture and other conferences. Three undergraduate students and four graduate students received training in statistical techniques, GIS, and the use of publicly available weather and climate data sources in mitigating the risk of crop disease. A graduate student also was trained on NetCDF file format at an NSF sponsored Unidata workshop at NCAR. PARTICIPANTS: Kathleen M. Baker, Western Michigan University, is the principal investigator on the project. Graduate and undergraduate students (Susan Benston, Thomas Lake, Douglas Rivet, Paul Roehsner, Jason Smith, Joshua Williams) worked as research assistants on the project and received training in statistics, spatial analysis, and computer programming at Western Michigan University. William Kirk is the project director for Michigan State University and is assisted by Lee Duynslager. Jeffrey Stein resigned as the project director for South Dakota State University, passing that responsibility to Thomas Cheesbrough. Dennis Todey is an additional PI at SDSU and Krishna Bondalapati is a Research Associate II and doctoral student at South Dakota State. Mark Boudreau has taken over as project director for University of Georgia from Joel Paz and Gerrit Hoogenboom. Abraham Fulmer is a graduate student at UGA, working closely with Robert Kermait, Joy Carter, and Ankit Arte. Robert Brueggeman will be PI on a new subcontract from North Dakota State University. Phillip Wharton, University of Idaho, received summer salary as a consultant. Rabiu Olatinwo completed his postdoctoral research at University of Georgia. Beth Plale at Indiana University was a collaborator on the project, allowing us to implement workflows through the Linked Environment for Atmospheric Discovery (LEAD) portal. Robert Trenary is a consultant for neural networking solutions at Western Michigan University. Ilya Zaslavsky is a consultant on the project, providing us with data backup and archiving solutions. Three undergraduate students and four graduate students received training in statistical techniques, GIS, and the use of publicly available weather and climate data sources in mitigating the risk of crop disease. A graduate student also was trained on NetCDF file format at an NSF sponsored Unidata workshop at NCAR. TARGET AUDIENCES: Our extension efforts aim to change the knowledge of growers as they make management decisions that impact actions such as fungicide spray quantities and frequencies. As was anticipated, the forecast site for potato late blight in Michigan was the first to go live and is being quickly followed by implementation of model results for decision support in the Southeast for leaf spot of peanut and in the Dakotas for Fusarium head blight of barley. In Michigan, with the neuroweather website now integrated into the lateblight.org website, clientele can now receive a "5-Day Forecast" on the risk of a late blight outbreak in the upcoming 5 days and also get an overview of what parts of the state are under the highest risk of an outbreak. This gives growers time and information to plan their crop protection measures for the following week. With the advent of Twitter, we now have another powerful tool to disseminate up to the minute information on late blight outbreaks directly to growers in a very cost effective manner (free). Site traffic statistics also show that we are reaching our target audience with most of the visits to the website coming from Michigan and occurring during the summer months. PROJECT MODIFICATIONS: As Teragrid is brought offline and associated projects, such as the Linked Environments for Atmospheric Discovery project are ended, our project has turned to new and impressive publicly available forecast datasets such as the National Digital Forecast Database and Gridded Model Output Statistics. The evolution of forecasting datasets across the US has provided us with an increasingly available and accurate pool of data that was not yet operational when the grant was written. Overall project objectives have not been modified, but some changes have been made regarding the extent of subcontracts. With the departure of a PI (SDSU) and a postdoc (UGA) from academia, work that can no longer be completed at SDSU and UGA will be moved to Western Michigan University and a new subcontract will be initiated with North Dakota State University (NDSU). The new PI at SDSU will be Dr. Thomas Cheesbrough and at NDSU will be Robert Brueggeman. Personnel changes have resulted in quite a few adjustments during the life of the grant but we do not expect that the overall quality of the work or scope of the objectives will be impacted. Our rate of expenditure has been reduced from what was originally planned and we anticipate the request of a no-cost extension to allow completion of work during the 2012-13 academic year.

Impacts
Analyses performed in this grant year improved our knowledge of the ways in which weather and climate data sources can be used to forecast crop disease effectively. Both the potato late blight and Fusarium head blight models showed trends of increased accuracy when at least two of three datasets (weather forecast model output statistics, climate normals, and previous day risk) were used in conjunction with one another. Substantial progress was also made in understanding the artificial neural network model training processes that relate to forecast accuracy. Previous overfitting of the potato late blight forecast model was corrected through adjustment of the training procedures. We implemented weight decay and early-stopping as part of the training regimen. After these changes, the model was able to train for many more iterations, producing an increase in overall accuracy. Most importantly, it was found that a majority of this increase occurred on days when crop disease risk was forecast. In a similar way, the Fusarium head blight model for barley was improved. A more robust risk model based on the two-dimensional Weibull function for pathogen infection and mycotoxin accumulation was considered during the peak growing season (June and July). A single hidden layer NN model was developed and tested to predict the risk five days in advance using the combination of historic, forecasted weather and climate normals. Different combinations of weather variables were generated and added to the existing NN model. Then the performance of the model was evaluated in terms of total accuracy. In Georgia the modified AU-Pnut model, relying solely on forecast data, was field tested with the more traditional approach which requires growers to also track the number of recent rain events. It was shown that all disease severity among treatments was significantly less than in the untreated control. The weather-based fungicide programs required fewer fungicide sprays as compared to the standard 7-spray fungicide program. In both irrigated and non-irrigated conditions, there were no significant differences between the AU-Pnut and the Modified AU-Pnut fungicide programs at any of the test locations. The Modified AU-Pnut program differed significantly from the standard program at the Attapulgus irrigated trial, but was not significantly different at the other locations. This simplified approach will make the leaf spot risk forecasts easier to implement using methods developed for late blight of potato and Fusarium head blight of barley. Each of these changes in knowledge changes the conditions under which growers will manage crop disease risk as these results are disseminated and become part of decision support systems in the 2012 growing season. By reducing grower uncertainty, we reduce the reactionary tendencies of those individuals placed in high risk situations and bring about a steady change toward action that relies on current and accurate forecast information.

Publications

  • Bondalapati, K.D., J.M. Stein, and K.M. Baker. 2011. Using Forecasted Weather Data and Neural Networks for DON Prediction in Barley. Poster session presented at: Joint Statistical Meetings, Miami, FL.
  • Kirk, W., Duynslager, L., Wharton, P. and K. Baker. 2011. Updates to the Five-Day Potato Late Blight Forecast, Twitter-Page for Michigan for 2011. Newsline (a publication of the Michigan Potato Industry Commission).
  • Rivet, D. and K.M. Baker. 2011. Non-parametric interpolation in weather-based disease forecasting systems. Association of American Geographers, 12-16 April, Seattle, WA.
  • Baker, K.M., Williams, J, Lake, T.L., and W.W. Kirk. 2011. The role of climate normals in crop specific weather forecasts. In E. Gelb and K. Charvat (Eds.), EFITA/WCCA'11: Papers presented at the 8th European Federation for Information Technology in Agriculture, Food and the Environment / World Congress on Computers in Agriculture (pp.493-498). Prague, Czech Republic: Czech Center for Science and Society.


Progress 09/01/09 to 08/31/10

Outputs
OUTPUTS: Activities continued to focus on analysis of National Weather Service data sources for use in crop forecasting systems. For leaf spot of peanut during 2010, modifications were made to 2009 trials. Daily extended-range forecast model output statistics guidance was compared to on-site weather stations and risk of favorable conditions for leaf spot infection was calculated in accordance with the AUPnut model alert. For Fusarium head blight, a more robust risk model based on the two-dimensional Weibull function for pathogen infection and mycotoxin accumulation was considered to estimate the risk at each day during the peak growing seasons. A single hidden layer neural net (NN) model was developed to predict the risk five days in advance using the combination of historic, forecasted weather and climate normals. The performance of the NN model was evaluated using field data obtained from several location*years in the region. For late blight of potato, detailed analyses of NN model interactions with forecast and climate data included varying the learning algorithms, data representations, number of hidden nodes, and growing season sets used to train the NN. Climate variability was examined as a factor in determining disease variability. Indicator and probability kriging were examined as methods to increase the spatial resolution of working NN's. The neuroweather website was redesigned, new scripts were written and the back-end database was reprogrammed to enable the changes necessary for the new website design and new forecasting features went live on May 1, 2010. An account was also set up on the Twitter website with a link to this account posted on the front page of the lateblight.org website. Michigan growers, extension agents, crop consultants and others who are interested in tracking the threat of late blight in Michigan can now receive a "5-Day Forecast" on the risk of a late blight outbreak by visiting the neuroweather website. In addition, disease forecasts are posted on Twitter on a weekly basis or when the risk of a late blight outbreak is high. People following the website on Twitter will receive an SMS message to their mobile phone as soon as the Twitter message is posted. This is an excellent method of us to disseminate results and warnings to growers, extension agents and crop consultants in MI and elsewhere in the world. To date the late_blight Twitter account has been receiving on average 1 new follower a week and we now have over 48 followers. A lot of these followers are crop consultants who can then disseminate the information to their growers and clients. In Georgia, knowledge gained was disseminated to growers at local events, including trials in demonstration plots at a University of Georgia field day attended by growers, extension specialists/agents, scientists, and university administrators. Overall results from forecasting techniques that use existing NSF cyberinfrastructure were disseminated at the Midwest Weather Working Group, Microsoft Corporation, and a Cloud Computing Workshop. One graduate students was graduated (MA, Geography) with an emphasis on climate impacts on crop disease variability. PARTICIPANTS: Kathleen M. Baker, Western Michigan University, is the principal investigator on the project. Jeffrey Stein is the project director for South Dakota State University, and Dennis Todey is an additional PI; William Kirk is the project director for Michigan State University; Mark Boudreau has taken over as project director for University of Georgia from Joel Paz and Gerrit Hoogenboom. Phillip Wharton, University of Idaho, received summer salary as a consultants. Krishna Bondalapati is a Research Associate II at South Dakota State and Rabiu Olatinwo is a postdoctoral researcher at University of Georgia. Graduate and undergraduate students (Susan Benston, Cassandra Hoch, Douglas Rivet, Steven Schultze, Jason Smith, Joshua Williams, Magdalena Wisniewska) worked as research assistants on the project and received training in statistics, spatial analysis, and computer programming at Western Michigan University. Beth Plale at Indiana University was a collaborator on the project, allowing us to implement workflows through the Linked Environment for Atmospheric Discovery (LEAD) portal. Robert Trenary is a consultant for neural networking solutions at Western Michigan University. Ilya Zaslavsky is a consultant on the project, providing us with data backup and archiving solutions. TARGET AUDIENCES: We have begun to serve our target audience of growers, extension agents, crop consultants and others who are interested in tracking the threat of crop disease by making our potato late blight forecasts available in Michigan. "5-Day Forecast" on the risk of a late blight outbreak is available at our neuroweather website, linked from lateblight.org. This extension effort aims to change the knowledge of growers as they make management decisions that impact actions such as fungicide spray quantities and frequencies. In addition, disease forecasts are posted on Twitter on a weekly basis or when the risk of a late blight outbreak is high. People following the website on Twitter will receive an SMS message to their mobile phone as soon as the Twitter message is posted as part of our effort to disseminate results and warnings to growers, extension agents and crop consultants in MI and elsewhere in the world. Similar efforts are under developement for leaf spot of peanut and Fusarium head blight of small grains. PROJECT MODIFICATIONS: While we have not changed the objectives of the project, we have reduced our rate of expenditure from what was originally planned. South Dakota State University and University of Georgia both delayed expenditures until their initial model development (not funded by this project) for Fusarium head blight of barley and leaf spot of peanut were complete. Grant related work is now intensifying as these initial models are complete. We have also made additional changes to the distribution of funds between institutions. Continued health problems in the Michigan State University team have necessitated a reduction in workload and funding there. Another change in PI at UGA has resulted in a reduction in workload for that team as well. Work reductions at UGA and MSU will be made up by the team at WMU, and the WMU budget will be increased accordingly.

Impacts
Analyses performed in this grant year improved our knowledge of the ways in which weather and climate data sources can be used to forecast crop disease effectively. Analysis of 2009 growing season data for leaf spot of peanut took place in Fall 2009. The application of the AUPnut model based on the forecasts resulted in fewer fungicide applications than the traditional program, which typically had seven spray applications. The results from 2009 field trials served as input for the 2010 field experiment, aimed at evaluating the accuracy of the spray application guidance produced from the forecast for leaf spot management. Preliminary evaluation of the forecast for leaf spot from 2004 to 2009 at the KABY National Weather Service station in Albany, GA showed seasonal and annual variability in the ratio of risk/no-risk days. We observed similar variability at 23 other MOS stations across Georgia. The information on inter-annual variability of risk/no-risk days is important for model optimization and for development of long-term risk assessment tools for management of leaf spots in peanut. New forecast model results for Fusarium head blight in the Great Plains showed that 22 percent of the days were estimated as risk days in 36 locations over 9 years (2001-9) during the June and July. The NN model had a total prediction accuracy of 91%, sensitivity of 73% and specificity of 96% on training data set. The model performed equally efficient on the validation data set. Removal of historical weather variables from the NN model reduced the overall accuracy as well as sensitivity and specificity. Moreover, removal of either climate normals or forecasted weather variables from the model reduced the performance of the model in predicting the true risk day (sensitivity dropped). These indicate that the variables from three categories were required to predict the risk five days in advance. The NN model yielded 89% of the accuracy when tested on an independent field data obtained from 55 location*years in the region. The sensitivity and specificity on the testing data were 71% and 92%, respectively. Continued general testing of NN model workings with weather and climate data took place using potato late blight in Michigan as a test case. Normalization of variables, combining inputs, and increasing forecast day count each significantly increased the accuracy of the NN model. Spatio-temporal variables when substituted with climate normal variables yielded very similar results. Including climate normals increased the usability of models across broader regions. A change in condition occurred with the launch of new forecasting capabilities for potato late blight. With the neuroweather website now integrated into the lateblight.org website clientele can now receive a "5-Day Forecast" on the risk of a late blight outbreak in the upcoming 5 days. This gives growers time and more information to plan their crop protection measures for the following week. With the advent of Twitter, we now have another powerful tool to disseminate up to the minute information on late blight outbreaks directly to growers in a very cost effective manner (free).

Publications

  • Baker, K.M., Wharton, P., Stein, J., Paz, J., and W.W. Kirk. 2010. Building True Forecasts into Crop Disease Management Systems, Chapter 9 In Bundgaard, K. and L. Isaksen, (eds.), Agriculture Research and Technology. Nova Science Publishers.
  • Baker, K, Benston, S., Schultze, S., and D. Rivet. 2010. Forecast Accuracy in Crop Disease Early Warning Systems. Abstracts of the AAG Annual Meeting, Washington, DC.
  • Baker, K.M. 2010. Forecasting Crop Disease Risk Using Mesoscale Weather Forecasts. Microsoft Environmental Research Workshop Proceedings, Redmond, Washington.
  • Bondalapati, K.D., J.M. Stein, K.M. Baker, and D.G. Chen. 2009. Using Forecasted Weather Data and Neural Networks for DON Prediction in Barley. Poster: Proceedings of the 2009 National Fusarium Head Blight Forum, Orlando, FL. Canty, S., Clark, A., Mundell, J., Walton, E., Ellis, D., and Van Sanford, D. (Eds.), University of Kentucky, Erlanger, KY. pp. 30-32.
  • Hoch, Cassandra. 2010. Impacts of El Nino Southern Oscillation Phases on Fusarium Head Blight Disease of Wheat. Abstracts of the AAG Annual Meeting, Washington, DC.
  • Kirk, W., Duynslager, L., Wharton, P., and K. Baker. 2010. Five-day potato late blight forecast now available for Michigan. MSU Crop Advisory Team Alert. MSU CAT Alerts are published for vegetable growers on the web and in a weekly newsletter via mail weekly throughout the growing season.


Progress 09/01/08 to 08/31/09

Outputs
OUTPUTS: Output from the first year included activities in data analysis, development and assessment of new methods, and field trials. In the U.S. Northern Great Plains region Quality Controlled Local Climatological Data (QCLCD) and extended range forecast MOS were collected and hourly weather variables were calculated using the daily maximum and minimum conditions for 36 locations (June-July, 2001-08). Performance of the barley-FHB logistic regression model was examined when varying days of QCLCD was replaced with forecasting weather data. Artificial neural network (ANN) models were developed separately for predicting the risk various days in advance with the variables from the logistic regression model as input variables. Traditional regression methodologies identified the best set of predictors to use as input variables for ANN model. In the Great Lakes region a prototype data mining software for cataloging available data and calculating derived variables from National Weather Service alphanumeric messages was developed. A GUI was developed, accompanying an underlying workflow, which will generate input for two ANN systems based on often incomplete publicly available forecast data containing both primary and derived variables. The website prototype (www.lateblight.org/neuroweather.php) using the ANN computer model for potato late blight in Michigan developed by the project was completed. Based on input from users, a redesign was initiated to make the data more user friendly and understandable. New scripts and reprogramming of the back-end database have been initiated. Workflows were also developed and used to acquire and analyze atmospheric data and initiate numerical forecasts that run on TeraGrid. A Linked Environment for Atmospheric Discovery (LEAD) workflow created for the crop disease forecasting project initiates the Weather Research and Forecasting (WRF). These spatially gridded WRF outputs become input to a geographic information system (GIS) workflow for each crop disease. Daily forecasts were run through both the LEAD and GIS workflows for select months, regions, and crops for initial testing. In the Southeast region, field trials were conducted at Tifton and Midville, GA in 2009 to evaluate the possibility of controlling leaf spot of peanut (LSP) with fewer spray applications. The extended range forecast model output statistics (MOS) guidance was obtained from the National Weather Services (NWS) on a daily basis for the two nearest stations. Additional data were collected from on-site weather stations at each location. High and low risk spray application programs were included for comparison with the modified AU-pnut and the UGA-MOS Model. A total of five fungicide spray applications were conducted at Tifton, and four applications at Midville based on the GFSX MOS forecasts, fewer than the traditional seven spray program. Events included: Southeast Georgia Research and Education Center Field Day and presentations as the Potato Assoc. of America, American Phytopathological Assoc., World Congress of Computers in Ag., Assoc. of American Geographers Annual Meeting, and Supercomputing 2008. PARTICIPANTS: Kathleen M. Baker, Western Michigan University, is the principal investigator on the project. Jeffrey Stein is the project director for South Dakota State University, and Dennis Todey is an additional PI; William Kirk is the project director for Michigan State University; Gerrit Hoogenboom has taken over as project director for University of Georgia from Joel Paz. Phillip Wharton, University of Idaho, and Ding-geng Chen, South Dakota State, received summer salary as consultants. Krishna Bondalapati is a Research Associate II at South Dakota State and Rabiu Olatinwo is a postdoctoral researcher at University of Georgia. Graduate and undergraduate students (Cassandra Hoch, Thomas Overly, Jason Smith, Magdalena Wisniewska) worked as research assistants on the project and received training in statistics, spatial analysis, and computer programming at Western Michigan University. Beth Plale at Indiana University has become a collaborator on the project, allowing us to implement workflows through the Linked Environment for Atmospheric Discovery (LEAD) portal. Robert Trenary is a consultant for neural networking solutions at Western Michigan University. TARGET AUDIENCES: Nothing significant to report during this reporting period. PROJECT MODIFICATIONS: While we have not changed the objectives of the project, we have reduced our rate of expenditure from what was originally planned. South Dakota State University and University of Georgia both delayed expenditures until their initial model development (not funded by this project) for Fusarium head blight of barley and leaf spot of peanut were complete. We also have changed the distribution of funds between institutions. Much less of the research will take place at Michigan State University than originally envisioned due to health problems of the PI there. This research, especially involving potato late blight and some web development, will now be completed at Western Michigan University and the monies for those activities are now included in the WMU budget. Consultants were not paid in the first year of the project because of difficulties getting them approved by both WMU and USDA. This problem seems to have resolved itself in the last couple of weeks, but not in time for consultants to receive payment during the first year. Another change is in the project directorship at University of Georgia. Joel Paz, originally director at UGA, accepted a position at Mississippi State University and will no longer be involved in the project. Gerrit Hoogenboom has taken over as project director for University of Georgia.

Impacts
In the first year, the majority of project outcomes were in change of knowledge. In the Great Plains, the performance of the logistic regression model of FHB was 98% when predicting risk for one day in advance and 93% when predicting risk for five days in advance. However, five days advance prediction accuracies were skewed due to high proportion of non-risk days and the accuracy in predicting the true risk case (sensitivity) was only 52%. The prediction accuracy of ANN models ranged from 97% in case of one day advance predictions to 83% for five days in advance predictions. Six potential weather predictors were selected by traditional regression methods and were divided into two subsets and four ANN models were developed with optimal number of hidden nodes for each subset of predictors. The prediction accuracies of four ANN models varied from 81% to 84%. Since four models had approximately equal prediction accuracies, the model with fewer parameters was selected for further evaluation (four inputs and five hidden nodes). The prediction accuracy of the selected ANN model in Red River Valley region was 89% and was 82% in the completely dry region and was 82% in eastern South Dakota region. In the Great Lakes region, results of initial testing of LEAD WRF forecasts against ANN models indicate that PLB risk forecasts made using the LEAD workflow coupled with a GIS application of a standard disease model were highly accurate. Overall accuracy in August and September for 55 validation points at NWS station locations was 71 percent and 81 percent respectively. These accuracies are quite comparable with the current artificial neural network model in operation at NWS station locations. LEAD results exhibit a higher range of accuracy values based on location from over 90 percent (in southern Michigan) to 16 percent (close to the shores of the Great Lakes). In the Southeast region, data from the leaf spot of peanut field experiments will be used to evaluate the accuracy of the spray application guidance produced from the MOS forecast for disease management. Furthermore, data from the on-site weather station at the two locations will be compared with the nearest Georgia Environmental Monitoring Network (AEMN) and NWS stations actual observations. Preliminary analysis of the actual rainfall events compared with the extended range forecast MOS guidance show between 73% and 64% accuracy for the initial 96 hrs forecasts. The information gained from these trials and comparison of weather data from the three sources (NWS forecasts, AEMN and on-site weather station) will be useful for model optimization and for selecting additional locations for field trials.

Publications

  • Baker, K.M., Zaslavsky, I., and B. Plale. 2009. Towards cyberinfrastructure for multi-regional crop disease early warning systems. In Proceedings of the World Congress of Computers in Agriculture, 22-24 Jun, Reno, NV.
  • Smith, J., Baker, K.M., and R. Trenary. 2009. Mining derived weather forecast variables for crop disease risk prediction. In Proceedings of the World Congress of Computers in Agriculture, 22-24 Jun, Reno, NV.
  • Baker, K.M., Wharton, P., Kirk, W., and B. Plale. 2009. Late blight forecasting as a test case for multi-regional, multi-scale forecasting cyberinfrastructure. Potato Association of America Program and Abstracts.
  • Baker, K.M., Stein, J., Wharton, P., Paz, J., Kirk, W., and B. Plale. 2009. Cyberinfrastructure challenges to multi-regional, multi- scale weather forecasting for crop disease early warning systems. American Phytopathological Association Abstract CD.
  • Baker, K.M. 2009. Climate Change and the Demand for True Agricultural Disease Forecasts. AAG Annual Meeting Abstract CD.
  • Baker, K.M., Hoch, C.L., and I. Zaslavsky. 2008. Towards Cyberinfrastructure for Multi-scale Crop Disease Early Warning Systems. International Conference on High Performance Computing, Networking, Storage and Analysis (SC08) Proceedings and Abstracts.