VISIONE STEREO PER L’IRRORAZIONE A DOSE VARIABILE NEI VIGNETI: RIDUZIONE DELL’USO DI PRODOTTI CHIMICI NELLA VITICOLTURA ITALIANA

As one of the most important grape-growing regions in Europe, Italy’s Veneto region is renowned for its production of high-quality wines. However, the area is currently facing significant challenges in balancing crop protection with environmental sustainability. In viticulture, the adequate application of plant protection products (PPPs) is essential to secure both yield and wine quality, yet their excessive use has resulted in ecosystem degradation and biodiversity loss. To address this dual challenge, variable-rate spraying (VRS) technology has emerged as a highly promising solution, offering the potential to ensure effective crop protection while simultaneously reducing PPP consumption. Focusing on frontier methods and innovations, this dissertation, presented as a collection of papers, aims to advance the application of VRS in vineyards through state-of-the-art stereo vision technology. Around this central theme, three main research questions were addressed. First, to evaluate spray application, methods for analysing water-sensitive paper (WSP) were investigated. Several classical convolutional neural networks were benchmarked for droplet stain segmentation, and the best-performing model, U-Net, was extended into a multi-task learning model capable of both droplet segmentation and adhesive droplet splitting. Results show that the proposed model outperforms traditional WSP analysis approaches, reducing errors in droplet deposition assessment while enabling splitting of adhesive droplets. Second, a vineyard VRS system was developed by integrating a commercial stereo vision camera with an edge computing unit equipped with a GPU. The system combines YOLO-based canopy detection with GPU-accelerated 3D perception. Canopy volume was estimated in real time using trellis posts as dynamic references, and application rates were adjusted accordingly. Field evaluations demonstrated that the system accurately regulated spray flow rates based on the calculated canopy volume, reducing PPP usage by up to 57% while maintaining adequate spray coverage. Analyses of real-time performance and deployment costs confirmed its suitability for practical vineyard applications. Finally, this work examined the potential of monocular depth estimation (MDE), a cutting-edge 3D perception technique, for vineyard VRS applications. Two representative MDE models were integrated and evaluated for canopy volume measurement. Results indicated that although MDE lags behind stereo vision in geometric accuracy and temporal consistency, its superior semantic accuracy makes it particularly suitable for VRS based on leaf wall area. Through the development of a VRS system and its experimental validation, this dissertation demonstrates the practical potential of stereo vision–based VRS in vineyards and establishes a complete application framework, thereby contributing to the more sustainable and efficient use of PPPs in Italian viticulture.