<b>HORTICULTURAL CROP MODELING FOR DECISION SUPPORT</b>

<p dir="ltr">Annual fruit and vegetable production occurs within a complex system of interacting genetic, environmental, and management factors. By necessity, growers make management decisions with incomplete information, which can lead to unforeseen losses. Digital decision support tools (DSTs) have the potential to assist growers when making management decisions by synthesizing new information or predicting likely outcomes. DSTs may be driven by crop models which possess the quantitative rigor to navigate complex production systems. These models, which may be process-based (PB), reduced-order (RO), or utilize a machine learning algorithm (ML), have been historically designed as research instruments, but must be trained and calibrated with data analogous to commercial production systems to fulfill this purpose. Variety trials (VTs) of commercial cultivars represent the largest and most diverse source of publicly available data emulating commercial horticultural production, though they were collected for the narrow purpose of cultivar performance evaluation. The purpose of this research is to identify a pathway for crop models to provide applied decision support in a commercial production context so they may be utilized to power future DSTs.</p><p dir="ltr">This was accomplished by first designing and carrying out a novel data suitability assessment to systematically evaluate the potential of using VT data to train and calibrate a set of PB, RO, and ML models for pepper (Capsicum annuum), tomato (Solanum lycopersicum), and watermelon (Citrullus lanatus) crops. The PB and RO models were found unsuitable due primarily to their reliance on variables that require destructive sampling, such as leaf biomass.The VT data was found suitable for training and calibrating ML models, though the inclusion of raw VT data in addition to trial reports substantially improved suitability.</p><p dir="ltr">To circumvent the need for extensive destructive sampling, the efficacy of mobile device light detection and ranging (LiDAR) sensors to estimate aboveground vegetative structural biomass and structure numbers was evaluated. This was accomplished by capturing point cloud data across the 2023 agricultural season for tomato and watermelon plants, while simultaneously measuring the structural characteristics via destructive sampling. Regression of plant point cloud volume with structural characteristics yielded mixed success, with low error rates among watermelon plants and higher error rates among tomato plants (as assessed using relative root mean square error, RRMSE). Mobile device LiDAR was thus deemed an appropriate data collection tool under only specific circumstances.</p><p dir="ltr">Finally, to assess the efficacy of ML yield prediction models trained on VT data, a set of ML algorithms were trained on a dataset of tomato raw VT data: gated recurrent units (GRU) and long short-term memory (LSTM) neural networks and extreme gradient boosting (XGBoost) regression and random forest (RF) regression decision tree algorithms. An RO growing degree day-based Bayesian hierarchical model was simultaneously trained for comparison. Supplementing the VT data was weather data gathered from NOAA weather stations and the NASA POWER gridded weather dataset. The models were evaluated over a range of prediction windows, from 0 to 60 days in advance, as well as a full season of hypothetical weather data. The XGBoost model performed well across prediction windows (as evaluated by R2 and RRMSE), capable of predicting plot-level yield at a cultivar-level. It was concluded that ML models such as XGBoost can be reliably trained on VT data and may thus be utilized as the basis for developing crop model-based DSTs.</p>