Sasso, Daniele (2025) Artificial Intelligence for Sustainable Transformation in Bio-Resource Systems: From Smart Sensing to Autonomous Decision-Making. [Tesi di dottorato]

Documento PDF Tesi_Dottorato_Daniele_Sasso.pdf Visibile a [TBR] Amministratori dell'archivio Download (54MB) | Richiedi una copia

Abstract

The exponential growth of the global population and the escalating impacts of climate change are posing unprecedented challenges to food security and environmental sustainability. The agri-food sector must therefore increase productivity while minimizing resource use and ecological impact. In this context, Artificial Intelligence (AI) has emerged as a transformative driver for precision agriculture, sustainable resource management, and data-driven decision making. This PhD thesis investigates the integration of AI methodologies with advanced sensor technologies, edge computing frameworks, and embedded systems to optimize agricultural and environmental practices. The research spans four domains: real-time soil characterization using voltammetric sensors and Machine Learning (ML); microplastic detection in aquatic systems, encompassing both a benchmarking of statistical and ML models on optofluidic signals and their subsequent deployment through TinyML-enabled edge devices; crop mapping from multi-temporal optical and radar remote sensing; and resource management optimization via Reinforcement Learning (RL). A novel soil monitoring system is presented, integrating electrochemical sensor arrays with tailored machine-learning classifiers for in-situ soil characterization. The system achieves accuracies up to 0.99 in controlled laboratory tests and maintains a robust discrimination performance (above 0.90) under varying soil moisture conditions. Two electrode configurations (wire–plate and plate–plate) were benchmarked, demonstrating the influence of sensor geometry and environmental factors on signal quality. Furthermore, the proposed design enables simultaneous energy harvesting through Microbial Fuel Cells, paving the way for self-sustaining soil monitoring networks in precision agriculture. For microplastic detection, a comprehensive benchmarking study on optofluidic signal patterns systematically compared statistical and machine-learning models under varying signal-to-noise ratios and class imbalance conditions. Linear and probabilistic classifiers, including Support Vector Machines, Linear Discriminant Analysis, and Naïve Bayes, achieved error probabilities comparable to the analytical detector (approx 3.5%), confirming the robustness of data-driven approaches across operating regimes and noise levels. Building upon this methodological framework, a TinyML implementation was developed using quantized neural networks compiled into native C code through the NeuralCasting compiler. The resulting edge-deployed models achieved accuracy above 0.98, inference latency down to 210 ms, and a memory footprint reduced by 50 % compared to TensorFlow Lite. This approach ensures deterministic, energy-efficient, and real-time operation on microcontrollers, demonstrating the feasibility of fully embedded AI solutions for environmental microplastic monitoring. Remote sensing experiments on hazelnut orchards demonstrated state-of-the-art performance in crop mapping from multi-temporal optical and radar satellite imagery. The proposed framework integrates feature-level fusion and advanced machine-learning classifiers, achieving up to 0.96 overall accuracy and an F1-score of 0.91. Extensive cross-regional validation confirmed the high generalization capability of the models across heterogeneous agro-ecological areas, highlighting the potential of synergistic optical–SAR data and AI-based methods for scalable orchard monitoring and agricultural resource management. In port energy management, Reinforcement Learning (RL) controllers, particularly the Soft Actor-Critic algorithm, achieved up to 98% reduction in diesel-generator usage and economic performance comparable to or exceeding Model Predictive Control (MPC). The RL-based framework reached approximately 95% of the theoretical optimal cash flow, maintaining high efficiency even under forecast uncertainty. These results confirm the potential of data-driven control strategies to enhance sustainability, resilience, and operational autonomy in smart port infrastructures integrating renewable generation, energy storage, and shore-side electricity supply. Overall, the results provide a unified benchmarking and integration framework for AI applications in the agrifood and environmental domains, demonstrating how the synergy between multi-source data acquisition, edge intelligence, and advanced control strategies can foster sustainable intensification, environmental monitoring, and precision resource management.

Downloads

Actions (login required)

Modifica documento