Artificial Intelligence for Sustainable Vineyard and Forest Management under Climate Change

Forests and agricultural systems play a critical role in carbon sequestration, climate regulation, and sustainable resource management. However, these systems are increasingly threatened by the im-pacts of climate change, including rising temperatures, shifting precipitation patterns, and more frequent extreme weather events. These changes disrupt key ecological processes—diminishing forest carbon-storage potential and altering phenological cycles in crops such as grapevines—thereby complicating sustainable management and undermining climate-mitigation efforts. Accu-rate monitoring of carbon storage and phenological development is essential for climate mitigation and sustainable management, yet traditional field-based methods are labor-intensive, costly, and difficult to scale. This thesis addresses these challenges by integrating remote sensing and advanced machine learn-ing (ML) techniques for environmental and agricultural predictions. Three key areas are explored: (i) forest carbon sequestration and storage estimation, (ii) grapevine phenological stage prediction, and (iii) explainable AI for carbon assessment. For forests, a novel ensemble learning method combining deep learning and traditional models is proposed, leveraging Sentinel-2 imagery, Li-DAR-derived canopy height models (CHM), and climate variables. Results show that CHM signif-icantly improves carbon estimation accuracy, while additional climatic variables offer limited gains. The ensemble approach consistently outperforms individual models, demonstrating its effec-tiveness for large-scale carbon assessment. For viticulture, a semi-supervised pseudo-labeling framework is developed to predict grapevine phenological stages, addressing incomplete annotations in vineyard datasets. Using climatic data, this method improves predictive accuracy while remaining cost-effective and non-intrusive com-pared to traditional image-based monitoring. Growing degree days (GDD) are identified as the most influential feature, and pseudo-labeling reduces prediction errors substantially. Finally, explainable AI techniques, particularly SHAP values, are applied to enhance model inter-pretability. Gray-Level Co-Occurrence Matrix (GLCM) texture analysis of LiDAR data improves organic carbon estimation, identifying key structural and spectral parameters that drive carbon dy-namics. This research, resulting in three peer-reviewed publications, provides a comprehensive framework for sustainable forest management and precision agriculture. Key contributions include: 1. Development of a novel ensemble ML method for forest carbon estimation. 2. Introduction of a semi-supervised pseudo-labeling framework for grapevine phenology pre-diction. 3. Integration of GLCM texture analysis with explainable AI to enhance interpretability. 4. Comparative analysis of multi-source data for environmental and agricultural applications.