High-throughput phenotyping applications in control and outdoor environments for stress tolerance analysis of kale and blueberries production

Abstract

Abiotic stressors such as drought and salinity significantly impact plant growth, development, and productivity. Traditional phenotyping methods for assessing plant responses to these stressors are labor-intensive, time-consuming, and unsuitable for large-scale plant populations. This thesis explores the application of advanced high-throughput phenotyping techniques by utilizing cutting-edge imaging technologies combined with machine learning and deep learning models to assess plant stress tolerance and predict physiological traits. These application-driven approaches provide scalable, non-destructive, and precise phenotyping capabilities, improving resource management and accelerating breeding programs for stress-resilient crops. To evaluate salt stress tolerance in kale plants, a high-throughput phenotyping system utilizing RGB imaging was developed. This system automated the extraction of morphological traits, such as canopy area and axes lengths, using YOLOv8 instance segmentation models trained on images from GoPro and Raspberry Pi cameras. Results showed that the model achieved mAP values between 0.897–0.952, and plants with split-root systems demonstrated superior growth under high salt stress compared to single-root systems. An ARIMA model was also used to forecast plant growth, achieving low MAPE values, providing growers with a new method to optimize resource allocation in controlled environments. Assessing drought stress in blueberries is crucial for supporting breeding programs aimed at developing drought-tolerant varieties. To facilitate laborious and time consumption task, a custom hyperspectral imaging platform was designed to capture high-resolution spectral data. A Transformer-based model, LWC-former, was introduced to predict leaf water content (LWC) by transforming spectral reflectance into patch representations, effectively addressing multicollinearity issues in hyperspectral data. The results showed that our model achieved a coefficient of determination (R²) of 0.81 on the test dataset. The performance of the proposed model was also compared with a multilayer perceptron (MLP), partial least squares regression (PLSR), support vector regression (SVR), and random forest (RF), achieving R² values of 0.71, 0.47, and 0.58, respectively. The results demonstrated that LWC-former outperformed other deep learning and statistical-based models, demonstrating its effectiveness for large-scale drought tolerance phenotyping. To obtain more comprehensive insights into plant responses under drought stress, it is critical to predict additional physiological traits beyond LWC. To this end, a Graph Convolutional Network (GCN)-based model, Plant-GCN, was developed to predict multiple physiological traits of blueberries along with LWC, such as stomatal conductance (gs), electron transport rate (ETR), photosystem II efficiency (φPSII), and photosynthesis (A) using hyperspectral imaging data. The GCN model transformed spectral data into a graph-based representation, capturing complex spectral interactions among plants. The model achieved R² values ranging from 0.89 to 0.94 for different traits and consistently outperformed other statistical and deep learning-based models, demonstrating its ability to accurately and efficiently predict a wide range of physiological traits under drought stress conditions. This unified approach provides a more complete picture of plant responses, offering a scalable solution for phenotyping and improving drought resilience strategies.