Research publications and ongoing manuscripts spanning machine learning, intelligent decision support, environmental modelling, and applied AI. Entries are grouped by publication status and include abstracts, full texts, and external records where available.
Irrigation management is a key challenge in modern agriculture due to water scarcity and increasing climate variability. This work proposes a deep reinforcement learning (DRL) framework for irrigation scheduling that addresses data scarcity by training the agent within a digital twin of crop–soil dynamics. The environment combines a KNN-based weather generator that extends a 30-year historical record into 1000 synthetic seasons, an XGBoost model for daily soil water tension estimation, and the AquaCrop simulator for crop biomass modeling. A DRL agent is trained using Proximal Policy Optimization to learn a weather-aware irrigation policy without predefined rules. The framework is evaluated on vineyard field data from the Val d’Adige region (Trentino, Italy) over the 2023–2024 growing seasons. Results show a 19% reduction in seasonal water use and more than a twofold increase in the number of days within the optimal soil tension range, while maintaining comparable crop productivity to observed practices.
DOI:10.1109/MetroInd4.0IoT69397.2026.11653152
@inproceedings{silvestri2026deep,title={Deep Reinforcement Learning for Irrigation Optimization Based on Crop-Soil Dynamics},author={Silvestri, Romeo and Antonini, Mattia and Vecchio, Massimo and Antonelli, Fabio},booktitle={2026 IEEE International Workshop on Metrology for Industry 4.0 \& IoT (MetroInd4. 0 \& IoT)},pages={766--771},year={2026},organization={IEEE},doi={10.1109/MetroInd4.0IoT69397.2026.11653152},}
Efficient water management is a critical challenge for modern agriculture, particularly in the context of increasing climate variability and limited freshwater resources. This study presents a comparative field-based evaluation of two fuzzy-logic-based irrigation decision support systems for vineyard management: a Mamdani-type controller with expert-defined rules and a Takagi–Sugeno system designed to enable automated learning from ultra-local historical field data. Both systems integrate soil moisture sensing, short-term forecasting, and weather predictions to provide optimized irrigation recommendations. The evaluation combines counterfactual simulations with a bootstrap-based statistical analysis to assess water use efficiency, soil moisture control, and robustness to environmental variability. The comparison highlights distinct strengths of the two approaches, revealing trade-offs between water conservation and crop stress mitigation, and offering practical insights for the design and deployment of intelligent irrigation management solutions.
DOI:10.3390/s25237188
@article{silvestri2025smart,title={Smart Irrigation with Fuzzy Decision Support Systems in Trentino Vineyards},author={Silvestri, Romeo and Vecchio, Massimo and Pincheira, Miguel and Antonelli, Fabio},journal={Sensors},volume={25},number={23},pages={7188},year={2025},publisher={MDPI},doi={10.3390/s25237188},}
This paper presents the development and evaluation of a Fuzzy Decision Support System for irrigation management to promote sustainable water use in precision agriculture. A Mamdani-type fuzzy logic model was designed to optimize irrigation scheduling for vineyards in the Val d’Adige region of Trentino, Italy. The system integrates expert knowledge with real-time data from tensiometers and weather stations to generate adaptive, site-specific recommendations. Bayesian optimization was used to fine-tune the membership functions of fuzzy variables, enhancing system performance. Field evaluations conducted in 2023 across multiple sectors assessed total water use, average soil moisture, and days exceeding critical moisture thresholds. Results show that the system reduced total water consumption by over 52% compared to traditional methods while maintaining soil moisture within optimal levels. These findings underscore the potential of combining fuzzy logic and IoT-based sensing to support scalable, adaptive irrigation strategies across various crops and regions.
DOI:10.1109/CODIT66093.2025.11321831
@inproceedings{codit25,author={Silvestri, Romeo and Vecchio, Massimo and Antonelli, Fabio},booktitle={2025 11th International Conference on Control, Decision and Information Technologies (CoDIT)},title={A Fuzzy Decision Support System to Optimize Irrigation Practices in Trentino Region},pages={1-6},volume={1},year={2025},doi={10.1109/CODIT66093.2025.11321831},}
This paper provides valuable insights into the application of spatial interpolation techniques in smart agriculture and highlights the potential for further improvements through the integration of advanced geostatistical models. Specifically, it evaluates and compares two spatial interpolation techniques, Inverse Distance Weighting and Ordinary Kriging, for estimating soil moisture in apple orchards located in the Val di Non region of Trentino, Italy. Data were gathered from 18 tensiometer sensors deployed across the apple orchards, providing continuous soil moisture measurements over a specified time frame in 2023. The accuracy of both interpolation methods was assessed using root mean square error as the primary evaluation metric, with various validation methods employed to ensure robustness. Additionally, statistical analyses were conducted to determine the significance of differences in performance between the methods. The results indicate that Inverse Distance Weighting, despite its computational efficiency, slightly outperforms Ordinary Kriging in terms of accuracy, with statistically significant lower error values, making it a preferable choice for real-time soil moisture mapping and precision irrigation management in the region.
DOI:10.1109/MetroAgriFor63043.2024.10948750
@inproceedings{silvestri2024comparative,title={Comparative Analysis of Soil Moisture Interpolation Techniques in Apple Orchards of Trentino Region},author={Silvestri, Romeo and Vecchio, Massimo and Pincheira, Miguel and Antonelli, Fabio},booktitle={2024 IEEE International Workshop on Metrology for Agriculture and Forestry (MetroAgriFor)},pages={557--562},year={2024},organization={IEEE},doi={10.1109/MetroAgriFor63043.2024.10948750},}
Accepted & forthcoming
Accepted manuscripts awaiting their final publication record.
Systematic reviews and meta-analyses are essential for advancing scientific research, but they are time-consuming, labor-intensive, and prone to human error during literature selection and data extraction. This paper proposes a framework that integrates Large Language Models (LLMs) and retrieval-augmented generation (RAG) into the PRISMA 2020 workflow to assist and standardize key review tasks. Our contribution is methodological rather than algorithmic: we operationalize its stages as a single reproducible, auditable pipeline built from established, openly available components, prioritizing transparency and re-execution over customized elements. The framework uses structured query templates to ensure reproducible bibliographic searches, an LLM-driven pipeline for title and abstract screening, and a RAG-based full-text eligibility assessment. Both automated stages rely on structured prompt templates that encode inclusion and exclusion criteria as machine-interpretable instructions. Beyond selecting relevant papers, it extracts specific information from documents, such as application details or datasets, and harmonizes it into a standardized metadata schema. To evaluate our proposal, we performed an experimental validation using an existing systematic literature review on Human Action Quality Assessment (AQA) as a reference. The framework achieved an extraction recall of 85.11% for domain-specific datasets and identified 25 additional data sources not reported in the original review. The pipeline achieved an overall article detection rate of 71.28%. This reduction was primarily a consequence of the strict, dataset-oriented Boolean query adopted in the identification phase, rather than failures in the automated LLM or RAG components. Supplementary analyses confirmed near-deterministic behavior across repeated runs and successful transfer to a different LLM family at negligible computational cost. These results suggest that the framework can effectively assist in identifying relevant studies and extracting information while maintaining alignment with PRISMA 2020 guidelines. The proposed approach is structured to facilitate adaptation to other research areas by declaring new query and prompt templates.
Submitted & under review
Manuscripts currently in the editorial and peer-review process.
Regression models deployed in production are vulnerable to concept drift, which degrades performance over time and requires timely monitoring. We introduce the Wasserstein Drift Detector, a model-agnostic online detector that monitors a deployed regression model’s squared residual stream. The detector compares an older reference window and a recent window within a sliding residual buffer using the empirical 1-Wasserstein distance, and signals drift when this distance exceeds a robust threshold calibrated from a rolling interquartile-range estimate; a hysteresis mechanism reduces redundant alarms. We evaluate the detector against ADWIN, KSWIN, and Page-Hinkley on nine synthetic drift scenarios under a shared automatic calibration protocol, covering abrupt, gradual, local, and global input-target shifts plus two shape-only residual-distribution shifts preserving the first two moments. At the calibrated operating points, the detector achieves a detection rate of at least 0.92 in every scenario, whereas each baseline fails, with a detection rate of 0.00 in one scenario, while its flicker rate never exceeds 0.02 redundant alarms per 1,000 post-drift samples. Holm-corrected paired tests and a Friedman–Nemenyi ranking confirm these differences; an ablation study attributes alarm stability to the hysteresis mechanism and robust threshold; a sensitivity analysis identifies the threshold multiplier as the main hyperparameter-sensitivity source on local and shape-only drifts. Two real-world case studies complement the benchmark: on a Toulouse bike-sharing residual stream with repeatedly non-unimodal residuals, the detector raises one to two orders of magnitude fewer alarms than the baselines, and a preliminary soil-moisture forecasting probe shows the same pattern.
In preparation
Ongoing manuscripts that are being prepared for submission.
