AI-Powered Multispectral Imaging System for Precision Agriculture

Growing pressure on global food security, exacerbated by population growth, climate change and environmental degradation, requires innovative solutions to ensure sustainable and efficient agricultural production. This thesis addresses several challenges by integrating advanced optical and Artificial Intelligence (AI) technologies. The main objective is the early identification of biotic and abiotic stresses, the optimisation of resource use, the reduction of waste and the minimisation of adverse environmental impacts. This work starts with an analysis of the theoretical foundations of optics and imaging techniques, including spectrometry, multispectral imaging, and thermographic imaging. The interaction of light with leaves and the generation of spectral signatures for crop health monitoring are explored, highlighting how such signatures can be used to identify biotic and abiotic stresses. Subsequently, non-invasive techniques for disease and pest detection are analysed, with a focus on remote sensing, vegetative indices and applications of imaging technologies. Particular attention is paid to the application of Deep Learning (DL) methodologies in agriculture. A significant contribution of this research is the design and implementation of two innovative multispectral imaging systems, called MSX (with filter wheel) and MS5c (with multiple sensor arrays). Both systems are equipped with environmental sensors to map as much information as possible. These advanced multispectral systems have been optimised thanks to the integration of information from spectrometric data, which has allowed the identification of markers for various applications, such as the determination of the ripening state and the presence of diseases and pests. Experimental validations conducted in the laboratory, greenhouses, and open fields demonstrate the effectiveness of multispectral systems in improving agricultural management. The thesis proposes the use of advanced image processing methods based on DL, which improve and facilitate the automatic detection capabilities of flavescence dorée vectors. Furthermore, an innovative method based on a multi-model pipeline called Multi-Models Ensemble (MME) has been developed, which can optimally exploit the entire multispectral data stack for plant disease detection, even in the early stages of development. The integration of these technologies allows the development of automated and reliable systems for monitoring, reducing the need for manual inspections and increasing the precision of interventions. These systems can promptly alert the farmer, facilitating sustainable management of resources in the context of precision agriculture. In summary, these innovations highlight the potential of optical technologies and artificial intelligence in improving the sustainability and resilience of agricultural practices, opening new research perspectives.