Varlese, Rosaria (2025) Functional artificial intelligence and robotics for agro-eco system services and food production assurance (fAIR). [Tesi di dottorato]

Documento PDF TESI_ROSARIA_VARLESE.pdf Visibile a [TBR] Amministratori dell'archivio Download (5MB) | Richiedi una copia

Abstract

Entomophilous pollination is a fundamental process for the reproduction of over 75% of plant species and constitutes an essential ecosystem service for both biodiversity and agricultural production (Aizen et al., 2009; Layek et al., 2023). However, pollinators are increasingly threatened by environmental and anthropogenic pressures, including climate change, the reduction of floral resources, intensive pesticide use, and habitat loss (D. Goulson & Hughes, 2015). These factors compromise their health and, consequently, the stability of natural and agricultural ecosystems. To address this challenge, advanced tools are needed to monitor colony health in real time and identify early warning signs of stress. In this context, the application of the Internet of Things (IoT) and Artificial Intelligence (AI) offers new opportunities to automatically and continuously collect and analyze data, improving pollinator management and contributing to their conservation (Talaviya et al., 2020a). This study, developed within the framework of the PhD project Functional Artificial Intelligence and Robotics for Agro-Ecosystem Services and Food Production Assurance (fAIR), aims to establish an innovative system for monitoring pollinator health by leveraging IoT sensors and Machine Learning models to collect, interpret and model variations in colony weight as an indicator of health status. Through the use of honey bees and bumblebees, employed as sensitive bioindicators, the project analysed two Mediterranean agro-environmental systems characterised by distinct ecological and management conditions. Observations were carried out in an open-field citrus orchard and a greenhouse tomato crop, allowing the identification of site-specific dynamics and highlighting the role of pollinators in detecting potential environmental stress signals. In Metaponto, Basilicata, attention was directed to honey bees (Apis mellifera) in open-field conditions, where colonies are exposed to significant climate variations and rely on the availability of nectar resources. In Torre del Greco, Campania, the research focused on bumblebees (Bombus terrestris) used for tomato pollination in greenhouses, a more stable environment characterised by limited and artificially managed floral resources. One of the most innovative aspects of this work was the development of a structured dataset for the integrated monitoring of honeybee and bumblebee colony health in operational agricultural contexts. The dataset combines biological and microclimatic parameters acquired through high-resolution IoT sensors and represents the first known field-validated example of synchronous and high-frequency data collection on two pollinator species. The system collected data on key parameters such as colony weight, internal and external temperature, humidity, bee activity, sound intensity and external climatic variables, transmitting all information to a scalable, secure cloud platform with real-time access. The central element of the monitoring system was the weight of hives and nests, selected as the primary indicator. Weight variations reflect physiological, trophic, and stress-related dynamics (Childers et al., 2011; Flores et al., 2019). Continuous monitoring enabled by IoT allowed for precise, real-time data collection, providing a reliable proxy for pollinator health assessment. The analysis revealed significant differences in weight dynamics in the two sites: In Metaponto, hive weights showed stable distributions over time, with relatively limited variation. This stability suggests that structured beekeeping management and consistent nectar availability play a key role in maintaining colony balance (Fróna et al., 2019). Observed weight fluctuations primarily followed resource intake and consumption cycles, indicating that bloom availability and nectar flow are crucial in regulating colony biomass. In Torre del Greco, nest weights showed greater variability over time, with wider and more irregular fluctuations. This heterogeneity reflects the annual life cycle of bumblebees, which includes phases of growth and expansion followed by a natural colony decline. Additionally, unlike honeybees, bumblebees proved to be more sensitive to temperature fluctuations, typical of the greenhouse environment, which directly affect their metabolism and activity (Khouri et al., 2011; Sundström et al., 2014). These factors result in more pronounced variations in nest weight, making their patterns less predictable than those of honeybee hives. To model weight variations, five Machine Learning models were applied and compared: Linear Regression, Ridge, LASSO, Decision Tree, and Random Forest. The models were implemented in a supervised, non-temporal (a-temporal) mode, estimating colony weight based on environmental and biological conditions observed at the same moment, without modeling the sequential time structure of the data. Performance was evaluated using Mean Squared Error (MSE) and the coefficient of determination (R²): • Mean Squared Error (MSE), to quantify prediction error magnitude • Coefficient of Determination (R²), to assess how well the model explained variability in hive weight compared to actual data. Linear Regression provided stable results in structured contexts, while Ridge and LASSO Regression handled multicollinearity without significantly improving performance. Decision Trees demonstrated better ability to model non-linear relationships but with a risk of overfitting. Finally, Random Forest emerged as the most accurate model, with R² = 0.97 in Metaponto and R² = 0.76 in Torre del Greco, due to its capacity to capture complex patterns and reduce overfitting (Breiman, 2001). The high predictive accuracy of this model enhanced the understanding of the relationships between environmental variables and colony health, suggesting its potential for early stress detection. This opens the way to targeted, data-driven interventions that optimize pollinator management and support their conservation. Future developments of this research will focus on key areas to further enhance the predictive system and the quality of pollinator colony surveillance. In particular, it will be essential to expand the dataset by including previously unconsidered ecological and management variables, such as the availability of nectar and pollen resources, floral composition, and the intensity of agricultural practices. An additional level of insight will be achieved through the integration of molecular data aimed at investigating in the laboratory the physiopathological conditions potentially relevant to pollinator health; the virological approach tested on a pilot scale in bumblebees represents a concrete first step in this direction. The adoption of advanced temporal predictive models, such as Recurrent Neural Networks (RNNs), will enable the modeling of sequential dependencies and the prediction of colony health evolution. The development of intuitive interfaces for beekeepers, equipped with dashboards and automatic alerts, will make the system practical and accessible for daily monitoring. Furthermore, the integration of melissopalynological analyses will allow the correlation of pollen quality and origin with variations in colony biomass, broadening opportunities for targeted management interventions. This multidimensional approach will make the system increasingly effective, adaptable, and ready to support innovative strategies for pollinator conservation and the sustainable management of agroecosystems.

Downloads

Actions (login required)

Modifica documento