| Resumo: | Diseases and pests in agriculture represent a major problem worldwide, severely impacting crop quality and yield. Among them, downy mildew is a particularly devastating disease affecting grapevine. Early detection is crucial for timely intervention, preventing disease spread and reducing chemical treatments. Traditional evaluation relies on experts, which can be laborious, subjective and time-consuming. The integration of artificial intelligence into agricultural practices presents a promising solution for disease management, facilitating the automation of qualitative and quantitative disease assessment. The main objective of the PhD thesis was to develop new artificial intelligence and computer vision-based methods for early assessment of grapevine downy mildew using non-invasive sensing technologies under laboratory and field conditions. In particular, the following objectives were proposed: i) the exploration of artificial intelligence and non-invasive technologies to evaluate downy mildew under laboratory conditions; ii) the development and validation of a method to estimate downy mildew severity under laboratory conditions combining fuzzy logic and computer vision; iii) the use of convolutional neural networks and explainable artificial intelligence to early detect downy mildew under laboratory conditions; iv); the in-field detection and localisation of downy mildew applying explainable deep learning; and v) the use of deep semantic segmentation to assess downy mildew severity in images taken in commercial vineyards. For the first objective, artificial intelligence was used for analysing RGB and hyperspectral images of grapevine leaf discs. Spectral pre-processing, computer vision and machine learning were used to identify downy mildew infection in hyperspectral images. At the same time, classic computer vision was used to locate the symptoms in RGB images. The results demonstrated the potential of artificial intelligence and non-invasive technologies to early detect downy mildew and to estimate its severity accurately and objectively. For the second objective, classic computer vision was used to localise downy mildew symptoms on RGB images of grapevine leaf discs. Then, fuzzy logic was used to evaluate the pixels detected as symptoms with a degree of infection according to their intensity. The results demonstrated that computer vision and fuzzy logic can automatically and accurately estimate the severity of grapevine downy mildew under laboratory conditions. For the third objective, convolutional neural networks were applied to early detect downy mildew and classify disease stages in RGB images of grapevine leaf discs. In addition, Grad-CAM was used to interpret model predictions. The results highlighted the accurate early detection of grapevine downy mildew under laboratory conditions using low-cost techniques. For the fourth objective, a sliding window was used for analysing the grapevine canopy in images captured considering the variability of field conditions. Convolutional neural networks and vision transformers used transfer-learning for detecting regions with downy mildew in the canopy. Predictions were interpreted with explainable artificial intelligence methods. The results highlighted the use of convolutional neural networks for the automatic and explainable detection and localisation of grapevine downy mildew under field conditions. Finally, different semantic segmentation architectures were compared to detect downy mildew symptoms in grapevine canopy images. Imbalance problems due to small symptom size were reduced with data augmentation, MixUp, oversampling and undersampling techniques. Neural networks trained with light-weight encoders and using the Dice loss function allowed accurate and fast assessment of downy mildew severity in grapevine under field conditions. The results of the research presented in this PhD thesis demonstrated the capability of artificial intelligence and computer vision for objective, rapid and accurate early assessment of grapevine downy mildew under both laboratory and field conditions. The potential adaptability of these methods to other crops, diseases and pests offers important advances in precision agriculture. Furthermore, the integration of these methods on mobile platforms, such as tractors, would allow for enhanced disease management over large crop areas, optimising monitoring and intervention directly in the field.
|
|---|
