Development of postharvest life prediction model and smart quality management technologies for cut roses

다국어 초록 (Multilingual Abstract)

Recent expansion of online distribution and e-commerce has increased the demand for systems that can guarantee the quality and vase life (VL) of cut roses. Conventional quality evaluation of cut roses relies on visual inspection and the experience of evaluators, making it difficult to provide objective and consistent assessments. In addition, VL is determined by nonlinear interactions among multiple physiological factors, such as water balance, disease development, and ethylene response; therefore, simple statistical approaches are insufficient for accurate prediction. Artificial intelligence (AI) based smart quality management technologies have consequently attracted attention. However, the development of high-performance models is constrained by the lack of standardized physiological indicators for AI training, limited availability of large-scale labeled datasets, and a shortage of experts capable of integrating postharvest physiology with engineering design of imaging and sensing systems. Thus, high-performance AI models for VL prediction require integrated research on postharvest physiology, non-destructive quality assessment, AI based VL prediction techniques, and AI training methods that account for horticultural crop characteristics. This dissertation aimed to develop VL prediction models and smart quality management technologies that reflect characteristics of cut roses and to propose a design pipeline for an AI based VL guarantee system.
In Chapter 1, consumer and industry perceptions of an AI prediction based VL guarantee system were investigated through surveys conducted in Osaka and Fukuoka, Japan. A total of 65.9% of general consumers expressed a preference for VL guaranteed cut flowers, whereas only 27.4% of flower industry professionals indicated purchase intention. Regarding the preferred guarantee period, 48.3% of consumers favored at least 7 d, while 44.7% of professionals preferred at least 10 d, revealing group differences in required VL guarantees. In Osaka, 61.7% of consumers reported low trust in AI prediction based VL guarantee, whereas 52.9% of consumers in Fukuoka considered it reliable, indicating regional differences in AI acceptance. These results suggest that AI prediction based VL guarantee systems should be designed with differentiated service and pricing strategies according to customer type, and that the credibility of AI prediction must be supported by physiological and imaging evidence.
In Chapter 2, the effects of light sources in a hyperspectral imaging (HSI) system on postharvest quality of cut roses and object detection performance were evaluated using four light sources including halogen (HAL), incandescent (INC), fluorescent (FLU), and light emitting diode (LED). HAL produced high quality spectra across 480–900 nm and yielded the highest object detection performance of YOLO11x models; however, it increased petal temperature by 1-2 °C and shortened VL. FLU and LED had smaller impacts on petal temperature and VL, but both generated transient peaks between 480 and 620 nm that reduced detection performance. These findings indicate that hardware design and plant physiological responses must be considered together in automated sorting systems and support a two phase illumination strategy in which HAL is used for exploratory identification of diagnostic wavelengths and LED is used for routine monitoring.
In Chapter 3, the usefulness of thermography for early detection of gray mold disease (GMD) in cut roses. Flowers infected with Botrytis cinerea (B.cinerea) showed reduced opening, disrupted water balance, and decreased VL. One day before visible symptoms appeared, petal temperature of infected flowers was 1.1 °C higher than that of non-inoculated flowers, and this temperature increase was associated with changed mRNA levels of ethylene, reactive oxygen species, and aquaporin related genes. Multiple regression analysis showed that disease incidence on petals was positively correlated with petal temperature measured one day before symptom appearance, demonstrating that thermal imaging can be used as a non-destructive tool to monitor petal temperature and detect GMD at an early stage.
In Chapter 4, a VL prediction model for cut roses by combining HSI with a convolutional neural network (CNN). Dry and wet transport treatments, ethylene exposure, and B.cinerea inoculation were applied to induce water stress and disease, and their effects on VL, water relations, and spectral reflectance were evaluated. Reflectance in the 470–680 nm regions was closely associated with GMD symptoms, whereas reflectance in the 700–900 nm regions was strongly related to petal wilting. Based on these results, VL was classified into two categories, longer than 5 d and shorter than 5 d. Using composite HSI images and a YOLOv5 based CNN model, VL prediction of the two cultivars were predicted with R² values of 0.86 and 0.83, respectively, and GMD lesions were detected with high precision. These results indicate that HSI combined with deep learning provides an effective non-contact method for simultaneous disease diagnosis and VL evaluation.
In Chapter 5, six standard and spray rose cultivars were grouped into three sensitivity classes and a symptom-based scoring system was developed for VL prediction model. HSI images were analyzed using YOLOv5x, YOLOv8x, YOLOv11x, SSD, and Faster R-CNN. YOLOv11x showed the highest accuracy and fastest inference time for detecting flower organs and various senescence and disease symptoms. Detected symptoms were converted into four severity levels based on their frequency at the end of VL within each sensitivity group and restructured into weighted scores for water stress sensitive, GMD sensitive, and ethylene sensitive groups. When these scores were used as a time-series training dataset, the LSTM model achieved an MAE of about 1.4 d, a maximum R² of 0.88, and an RMSE of about 1.0 d, outperforming LSTM models that used only object detection results. This demonstrates that cultivar sensitivity based scoring system that reflects physiological responses and crop traits is an effective approach for VL prediction.
Overall, this dissertation integrates survey-based market analysis defining VL guarantee requirements, HSI and thermal imaging based non-destructive quality assessment, CNN based detection of disease and senescence symptoms, cultivar specific scoring systems, and LSTM time series models into an AI development pipeline for VL prediction in cut roses. The proposed pipeline can be extended to other horticultural crops and provides a framework for designing AI based VL prediction and quality guarantee systems that account for cultivar sensitivity and distribution environments. Such systems offer guidelines for smart APC and automated grading facilities and can contribute to waste reduction and improved sustainability in the horticultural industry through AI-based quality management.

목차 (Table of Contents)

• ABSTRACT 5
• CONTENTS 10
• LIST OF TABLES 15
• LIST OF FIGURES 16
• LIST OF SUPPLEMENTAL TABLES 19
• LIST OF SUPPLEMENTAL FIGURES 20
• LITERATURE REVIEW 21
• CHAPTER 1. A survey of consumer perceptions of vase life guarantees for cut flowers based on artificial intelligence predictions 32
• CHAPTER 2. Light sources in hyperspectral imaging simultaneously in-fluence object detection performance and vase life of cut roses 51
• CHAPTER 3. Early detection of Botrytis cinerea infection in cut roses using thermal imaging 92
• CHAPTER 4. Development of a longevity prediction model for cut roses using hyperspectral imaging and a convolutional neural network 132
• CHAPTER 5. Development of a vase life prediction system for cut roses based on cultivar specific scoring system 194
• CONCLUSION 246
• ABSTRACT IN KOREAN 249