Non Technical Summary
The sustainability of irrigated agriculture in the arid Western U.S. is being threatened by increasing water demands from a growing population, diminishing aquifers, recurring droughts, and the effects of a changing climate. These conditions are forcing alfalfa hay producers in the region to irrigate alfalfa without meeting its full water demands. This practice, known as deficit irrigation, inherently leads to a reduction in yield, which in turn endangers the livelihood of rural communities in the Western U.S. and threatens the access to affordable food for urban communities in the same region.The long-term goal of this research project is to help alfalfa farmers in the Western U.S. to ameliorate the economic impact of producing alfalfa with a limited water supply by developing an alfalfa hay Yield Forecasting Tool (YFT) that can be used to analyze many irrigation management scenarios and identify a decision for an upcoming irrigation event (i.e., don't irrigate or how much to irrigate) that maximizes yields without exceeding a water quota. To achieve this goal, the overall objective of this project is to identify a computational tool capable of estimating the effects that an irrigation management decision will have in the seasonal alfalfa hay yield with a satisfactory level of accuracy and limited data inputs.We'll identify such a tool by comparing four computational tools that can be used to estimate alfalfa hay yield: two artificial intelligence (AI) methods and two computer programs, known as crop growth models. We'll train the AI methods to estimate alfalfa hay yield and calibrate the crop growth models for the same purpose using data collected in Northern Nevada and the Texas High Plains from two historical and two ongoing alfalfa experiments. We'll identify the AI method or crop growth model that more accurately estimates alfalfa hay yield, and we'll incorporate the best performing AI method or model into the YFT. The resulting YFT will be a computer program that agricultural researchers or consultants in the Western U.S. can use to estimate the effects that different irrigation management scenarios will have in the alfalfa hay yield obtained at the end of a growing season. We'll release to the public the computer code of the YFT so that other agricultural researchers or consultants can improve it and/or modify it to fit the needs of other crops or regions.

Classification

Knowledge Area (KA)	Subject of Investigation (SOI)	Field of Science (FOS)	Percent
102	1640	2020	60%
405	0210	2080	40%

Knowledge Area
102 - Soil, Plant, Water, Nutrient Relationships; 405 - Drainage and Irrigation Systems and Facilities;

Subject Of Investigation
1640 - Alfalfa; 0210 - Water resources;

Field Of Science
2020 - Engineering; 2080 - Mathematics and computer sciences;

Goals / Objectives
The long-term goal of this research project is to help alfalfa farmers in the arid Western U.S. to ameliorate the economic impact of producing alfalfa with a limited water supply by developing an alfalfa hay Yield Forecasting Tool (YFT) that can be used to analyze many irrigation management scenarios and identify a decision for an upcoming irrigation event (i.e., don't irrigate or how much to irrigate) that maximizes yields without exceeding a water quota. To achieve this goal, the overall objective of this project is to identify a computational tool capable of estimating the effects that an irrigation management decision will have in the seasonal alfalfa hay yield with a satisfactory level of accuracy and limited data inputs. To achieve this overall objective, the specific objectives of this research project are to:1.Compare the accuracy and data input requirements of two existing alfalfa crop growth models and two machine learning (ML) algorithms when used to forecast alfalfa hay yields of historical and ongoing alfalfa experiments2.Develop an open-source alfalfa hay Yield Forecasting Tool (YFT) that uses the best performing alfalfa crop growth model or ML algorithm and a network of weather and soil water sensing systems to estimate alfalfa hay yield and water use3.Evaluate the YFT using experimental data obtained from ongoing experiments applying full irrigation (FI), i.e., full replenishment of the crop's water needs, conventional (constant) DI, and managed (variable) DI treatments to alfalfa

Project Methods
Data from two historical and two ongoing experimentswill be used to calibrate and evaluate two alfalfa crop growth models and to train and evaluate two Machine Learning (ML) algorithms. The alfalfa crop growth models that will be evaluated in this project are the CROPGRO-Perennial Forage Model (PFM)and the Agricultural Production Systems Simulator (APSIM). The ML algorithms that will be evaluated in this project are the classical Autoregressive Integrated Moving Average (ARIMA) algorithm and a neural network-based method, such as Recurrent Neural Network (RNN) or Transformer Neural Network (TNN).The first historical experiment was conducted in Fallon, NV, from 1973 to 1981. Data collected during these years was recorded in peer-reviewed publications and in theses available in UNR's library system. A total of 27 crop years (3 lysimeters × 9 years) will be available from this experiment. The second historical experiment was conducted at the USDA-ARS Conservation and Production Research Lab (CPRL) in Bushland, TX, from 1996 to 1999. Alfalfa was irrigated using a linear move irrigation system and subjected to Full Irrigation (FI) and Deficit Irrigation (DI) during the first three and the last year of this experiment, respectively. Co-PI Evett will share the quality-controlled dataset obtained from this experiment. A total of 16 crop years (4 lysimeters × 4 years) will be available from this experiment.The first ongoing experiment has the objective of evaluating the response of two alfalfa varieties to FI, mild DI (80% of FI), and moderate DI (60% of FI) in Northern Nevada. The alfalfa varieties are marketed as being drought tolerant (Ladak II) and high forage yield (Stratica). The experiment is being conducted at the Valley Road field Laboratory (VRFL) in Reno, NV. Experimental plots with dimensions of 25 ft × 5 ft were planted in the Fall of '20. Three years of experimental data (growing seasons '21 to '23) will be available from this experiment irrigated with a drip system, for a total of 54 crop years (18 plots × 3 years). PI Andrade-Rodriguez and Co-PI Solomon are conducting this experiment and will share its datawith the rest of the team. The second ongoing experiment has the objective of evaluating the response of alfalfa to mild DI (80% of FI) and moderate DI (60% of FI) in the Texas High Plains. The experiment is being conducted at the CPRL. Two years of experimental data (seasons '22 to '23) will be available from this experiment irrigated with a center pivot system, for a total of 16 crop years (8 plots × 2 years). The eight treatment plots are 30 ft × 30 ft and contain a neutron access tube centered within the plot. Weekly neutron probe readings will be taken to replenish 80% and 60% of crop water use to FC. Co-PIs Campbell and O'Shaughnessy are conducting this experiment and will share itsdatawith the rest of the team.Two datasets will be created with the 113 crop years available from the two historical and two ongoing experiments described. The first dataset will contain data readily available to crop managers and will include weather data, irrigation amounts and application dates, harvesting dates and alfalfa hay yields obtained from each harvest. The second dataset will contain, in addition to the data in the first dataset, data not readily available to crop managers that will include Leaf Area Index (LAI), plant height, biomass, growth stage, and evapotranspiration (ET). Both datasets will be used to calibrate and validate the CROPGRO-PFM and APSIM alfalfa crop growth models and to train and validate the ARIMA and RNN/TNN algorithms. The datasets will be divided in 75% that will be used for calibration/training and 25% that will be used for validation (25%).For the calibration of the crop models, a sensitivity analysis will be performed to identify the most important parameters that will be tuned during the calibration. The tuning of parameters will be achieved following a Bayesian calibration procedure. The outcome of the calibration will be two crop growth models (CROPGRO-PFM and APSIM) that are tuned to estimate alfalfa hay yield in Northern Nevada and the Texas High Plains. A graduate student under the supervision of PI Andrade-Rodriguez will be responsible for the training and evaluation of the crop models.Grid searches and cross validations will be performed to identify satisfactory parameters for the ARIMA and RNN/TNN algorithms. Multiple ARIMA and RNN/TNN will be trained and the most accurate will be kept for their comparison against alfalfa crop growth models. Ensemble strategy, where the consensus result is generated by averaging the results from models trained by subsets of data, will be also applied to identify the best ARIMA and RNN/TNN. The outcome of this process will be two ML algorithms (ARIMA and RNN/TNN) that are trained to estimate alfalfa hay yield in Northern Nevada and the Texas High Plains. A graduate student under the supervision of Co-PI Nguyen will be responsible for the training and evaluation of the ML algorithms.The calibrated alfalfa crop growth models and the trained ML algorithms will be used to forecast dry alfalfa hay yields using the 25% of the datasets reserved for validation. The Root Mean Squared Error (RMSE) between dry yields measured and estimated will be used as an indicator of the accuracy of the model/algorithm. The alfalfa crop growth model or ML algorithm that consistently shows to be more accurate when provided with data readily available and data not readily available to crop managers will be selected for its incorporation into the Yield Forecasting Tool (YFT) that will be developed by this project.Computer code will be written using the Python programming language within the Anaconda scientific computing environment (Anaconda Software, Austin, TX) to implement a YFT capable of performing a recursive execution of the best performing alfalfa crop growth model or ML algorithm. The YFT will automatically gather weather data from public sources of weather data and soil water content data from sensing stations consisting of Time Domain Reflectometer (TDR) sensors (TDR-315H, Acclima, Meridian, ID) connected to the Internet-Of-Things (IoT) using a low-cost node-gateway system. The YFT will store the data collected by this network of weather and soil water sensing systems in text files using a format compatible with the alfalfa crop growth model or ML algorithm selected to be incorporated into the YFT.Data from two experiments will be used to evaluate the YFT. Data collected during the '24 growing season at the CPRL will be used together with two years of data collected from a new experiment that will be conducted at the VRFL using a linear move irrigation system. Plots with dimensions of 25 ft × 5 ft will be assigned one of the following irrigation treatments: i) FI, ii) mild conventional DI (80% of FI), iii) mild conventional DI (60% of FI), iv) mild managed DI (variable amounts restricted by a water budget matching 80% of FI), and v) moderate managed DI (variable amounts restricted by a water budget matching 60% of FI). There will be three replicates for each irrigation treatment for a total of 30 plots (5 irrigation treatments × 2 cultivars × 3 replicates). The experimental design is a Randomized Complete Block Design (RCBD) with a split-plot structure, with irrigation treatments as the main plots, and the alfalfa varieties as the subplot. Plots were planted in the Spring of '22. A FI treatment was applied to all plots during the '22 growing season to ensure a good stand establishment. Irrigation treatments will start during the '23 growing season. The performance of the YFT will be assessed by calculating the RMSE between the alfalfa dry hay yields measured during each cut obtained from these experiments and the corresponding alfalfa dry hay yields estimated by the YFT.