Titolo della tesi: SMARTFAI: smart farming with artificial intelligence
This thesis, combining the themes of agriculture and AI, aims to provide an innovative solution in the agricultural sector. SMARTFAI (Smart Farming with Artificial Intelligence) is an industrial PhD project that aims to develop advanced and innovative artificial intelligence techniques to embed in monitoring systems. Additionally, the study and development of advanced vision models aim to enhance speed and efficiency of processes. Precision agriculture (PA), also known as smart farming, is an advanced approach that utilizes technology and data to optimize various aspects of crop production. This method involves the application of information technology, data analysis, and automation to enhance efficiency, productivity, and sustainability of the sector. Employing precision agriculture techniques, farmers can make more informed decisions about irrigation, fertilization, pest control, and harvesting. In addition, this approach contributes to resource wastage minimization, and environmental impact reduction, and improves overall yield and profitability in modern farming practices. Based on preliminary investigations, the research project has been developed focusing on grapevine cultivation, which is widely spread both in Italy and abroad and is simultaneously highly profitable. Specifically, the proposed approaches address one of the challenges that cultivation faces in the pre-harvest phase, namely, yield estimation. This is a common issue in precision agriculture for several reasons. Therefore, the thesis explores innovative methodologies aimed at achieving more accurate and faster yield predictions compared to manual methods through two classification approaches. Both methods include pixel-wise segmentation techniques for identifying grape bunches. Experimental results illustrate the efficacy of these methodologies in efficiently identifying grape pixels from RGB images containing yellow and blue bunches. The two automatic AI methods utilize Support Vector Machine and Convolutional Neural Networks as classification
models, relying on visual contrast-based features defined according to grape bunch color visual perception. Extensive experimental results demonstrate that the proposed methods can accurately segment grapes even in uncontrolled acquisition conditions and with a limited computational load. Furthermore, these approaches require a small training set, making them suitable for on-site and real-time applications that are implementable on smart devices in a user-friendly fashion, making them usable and even set up by winemakers.