Recipient Organization
UNIVERSITY OF FLORIDA
G022 MCCARTY HALL
GAINESVILLE,FL 32611
Performing Department
Agricultural and Biological Engineering
Non Technical Summary
The AgroClimate project (http://agroclimate.org/) at the University of Florida works on developing new knowledge and decision support tools based on climate monitoring, short term and seasonal forecasts, as well as long-term climate projections. Emphasis is given to developing methodologies for monitoring and forecasting the effects of extreme events, suchas droughts, extreme temperatures, and heavy rainfall, on crop development and yield, as well as on the occurrence of plant diseases. The main goal of this research project is to explore alternative methodologies, including models and artificial intelligence methodologies to calculate the most relevant climate indicators for agriculture and implement alert systems to inform farmers and agricultural managers about the potential effects of climate variability and change on crops. Ultimately, we intend to help farmers reduce production risk and increase resource use efficiency and the profitability of their operations.
Animal Health Component
70%
Research Effort Categories
Basic
10%
Applied
70%
Developmental
20%
Goals / Objectives
The main goal of this research project is to explore alternative methodologies, including empirical and process-based mechanistic models as well as artificial intelligence algorithms to calculate climate indicators for agriculture and implement alert systems to inform stakeholders of trends, current conditions and forecasts based on short-term weather forecast, seasonal outlooks and long-term climate projections.Specific objectives include:1.Review and identify the main climate indicators that are relevant to decision making in agricultural production systems;2.Implement new or enhance existing monitoring tools on AgroClimate to track climate indicators for agriculture;3.Explore the use of artificial intelligence algorithms to estimate crop yield and plant disease indicators such as leaf wetness duration;4.Develop and conduct workshops to engage stakeholders and demonstrate the use of AgroClimate tools and information to reduce production risk associated with climate variability and change.
Project Methods
Successful approaches to climate challenges require innovative and collaborative work. Education and participatory outreach play critical roles because stakeholders must be actively engaged in order for changes to be adopted and accepted. Our approach will include a review of climate indicators for agriculture, translation of existing scientific findings into information and decision support tools, identification of critical gaps in knowledge and research, as well as effective communication of promising findings within the research and extension communities.1.Review and identify climate indicators that are relevant to decision making in agricultural production systems: Temperature, precipitation amounts and distribution, timing of freezes, drought stress, environmental conditions for disease development and many other climate indicators can greatly affect crop yield and quality.Under this objective climate indicators for agriculture will be reviewed for relevance by project team members, collaborators, and regional stakeholders. We will also review existing regional and national sources of information related to the most relevant indicators. Climate indicators may have to be customized to ensure relevancy for different production systems and geographies. There is also a need to define standards and data sources to generate indicators that are broadly accepted by stakeholders. Under this objective we will review the methodology for calculation and data sources to calculate the indicators prioritized by the AgroClimate group and collaborators.2.Implement new or enhance existing monitoring tools on AgroClimate to track selected climate indicators: Under this objective we will implement new web-based tools or enhance exiting tools to track the selected indicators in the AgroClimate platform. The selected indicators will be calculated in accordance with the methods and data sources defined in Objective 1. Results will be presented in map and graph formats and include current and recently observed conditions as well as the deviation from the long-term averages to facilitate the comparison and evaluation of potential effects on crops. An initial list of indicators to de developed is presented below as an example. It includes:Growing degree-days;Heat degree-days;Night temperature above selected thresholds;Heat index alerts to inform about potential risk for field labor;Chill hours (< 32°F , between 32°F and 45°F , and partial chill units);Agricultural Reference Index for Drought (ARID);Leaf wetness duration (LWD);Plant disease infection indices;Others selected under objective 1.3.Explore the use of artificial intelligence (AI) algorithms to estimate crop yield and plant disease indicators such as leaf wetness duration (LWD): Timely and reliable crop yield prediction is important for regional and global food security. However, forecasting crop yields is a difficult endeavor that is becoming more challenging with the increased occurrence of anomalous weather conditions (e.g. droughts, heat waves, freezes, and floods) in major crop-producing regions of the USA and the world. Weather and climate variation magnify the importance of robust science-based tools that identify, measure, and monitor the effect of climate variability and extreme weather events on crop yields during the growing season.Encouraging results on the use of machine learning approaches for crop yield forecasting have been recently reported. Cai et al. (2019) combined satellite and climate data to evaluate various empirical models for wheat yield prediction based on three mainstream machine-learning methods (support vector machine, random forest, and neural network). According to the authors the machine-learning based methods outperform the well-known LASSO regression method in modeling crop yield. They concluded that combining climate and satellite data can achieve high performance of yield prediction.Reported AI-related research on plant diseases have mostly focused on image classification for disease detection and identification (Mohanty et al., 2016; Ferentinos, 2018; Lee et al., 2020). However, a recently published study evaluated the use of machine learning approaches to estimate LWD for prediction of apple scab (Wrzesie? et al., 2019). Most common pathogens in agricultural production systems, fungi and bacterium, require free water to grow and penetrate the plant tissue, starting the infection process, while temperature influences the speed of their metabolic reactions and development. LWD is defined as the period with free water available on crop leaves and is usually related to dew formation, rainfall and irrigation. Temperature and LWD are widely used to predict the risk of crop disease development.Under this research project we intend to explore the use of Machine Learning (ML) to predict crop yield (cotton) and estimate LWD. ML projects are generally composed of three phases or activities: 1. Data collection; 2. Training of the models; 3. Deployment of the models. During the last decade we have built a solid database of primary and derived weather variables that combined with observed and historical crop yield datasets and leaf wetness duration observations made at the well maintained weather stations serving the AgroClimate project provide a unique opportunity to explore ML approaches to estimate crop yield and LWD. The machine learning methods will be employed in this project for prediction contain random forest (RF), support vector regression (SVR) and neural network (NN) which have been used widely in multivariable predictions, including crop yield modelling. RF as an enhanced tree structure classification and regression methods obtains better performance than the original decision tree method. By determining the size (number of trees), the modelling efficiency of RF could be optimized between computation speed and accuracy. SVR, derived from support vector machine (SVM), uses support vectors to define the margin and hyperplane for fitting the dependent. A linear model will be found by increase the dimension of the input data using a kernel function, predefined in an SVR. It has high tolerance of outliers when the margin is properly defined (Smola and Schölkopf, 2004). NN was developed on the biological concept of neural network in recognition (Specht, 1991). A NN contains an input layer, a hidden layer and an output layer. While an initial input layer receiving raw data, each node (neuron) within a hidden layer will use a series weights (coefficients) to define the significance of input and pass to a transfer function for being converted into outputs. Containing multiple neurons, a single hidden layer can be stacked with another to realize a deep learning process. Additional algorithms may be included as the project progress and related research is reported in the literature.4.Engage stakeholders, demonstrate the use of AgroClimate tools during extension workshops and produce educational materials on the use of climate indicators to reduce production risk: The current success of AgroClimate is based on three main pillars: 1. Sound scientific basis; 2. Simple and intuitive web interface; and 3. Continuous engagement with growers and extension faculty. During this project we intend to further strengthen our interaction and engagement with stakeholders, reinforcing the two-way communication channels between growers, extension and research faculty. We will also develop and participate in training workshops and produce educational materials related to climate risk in agriculture and the use of AgroClimate tools.