Abstract
In this paper we present a Deep Learning-based methodology to automatically classify 12 of the most representative grapevarieties existing in the Douro Demarked region, Portugal. The dataset used consisted of images of leaves at different stages of development, collected on their natural environment. The development of such methodologies becomes particularly important, in a scenario in which ampeleographers are disappearing, creating a gap in the task of inspection of grape varieties. Our approach was based on the transfer learning of the Xcepetion model, using Focal Loss, adaptive learning rate decay and SGD. The model obtained a F1 score of 0.93. To clearly understand the predictions of the model, and realize which regions of the image contributed the most to the classification, the LIME library was used. This way it was possible to identify the parts of the images that were considered for and against each prediction.

Abstract
Wine is one the most important products from Portugal, being the grapevine variety very important to ensure uniqueness, authenticity and classification. In the Douro Demarcated Region, only certain grapevine varieties are allowed, implying the need for an identification mechanism. The ampelographs, professionals that use visual analysis to classify grapevines, are disappearing. In this situation, one possible replacement for ampelographs can be deep learning models. In previous experiments, we successfully classified 12 grapevines varieties, fine-tuning the Xception model, achieving ~0.9 in F1 score, raising the question, "what is the impact of the fine-tuning layers' configuration in our results?". This paper presents an analysis of the impact of different layers' configuration in fine-tuning Xception model to classify 12 grapevine varieties with images acquired in a natural environment. Despite the model achieved F1-score of 0.92 in all configurations, using the Grad-CAM approach, we show that layers' configuration in fine-tuning implies the quality of the models' prediction. As analysis' result, we can see that the model acting as feature extractor and fully fine-tuned obtains similar results in terms of metrics and pixel contribution, and fine-tuning only the last two blocks lead the model to look at more features in the image.

Abstract
Wine is the most important product from the Douro Region, in Portugal. Ampelographs are disappearing, and farmers need new solutions to identify grapevine varieties to ensure high-quality standards. The development of methodology capable of automatically identify grapevine are in need. In the scenario, deep learning based methods are emerging as the state-of-art in grapevines classification tasks. In previous work, we verify the deep learning models would benefit from focus classification patches in leaves images areas. Deep learning segmentation methods can be used to find grapevine leaves areas. This paper presents a methodology to segment grapevines images automatically based on the U-net model. A private dataset was used, composed of 733 grapevines images frames extracted from 236 videos collected in a natural environment. The trained model obtained a Dice of 95.6% and an Intersection over Union of 91.6%, results that fully satisfy the need of localise grapevine leaves.

Abstract
The grapevine variety plays an important role in wine chain production, thus identifying it is crucial for control activities. However, the specialists responsible for identifying the different varieties, mainly through visual analysis, are disappearing. In this scenario, Deep Learning (DL) classification techniques become a possible solution to handle professionals' scarcity. Nevertheless, previous experiments show that trained classification models use the background information to make decisions, which should be avoided. In this paper, we present a study allowing the assessment of removing background regions from the grapevine images in the improvement classification using DL models. The Xception model is trained with a normal dataset and its segmented version. The Local Interpretable Model-Agnostic Explanations (LIME), Grad-CAM, and Grad-CAM++ approaches are used to visualize the segmentation impact in classification decisions. F1-score of 0.92 and 0.94 were achieved, respectively, for segmented-dataset and normal-dataset trained models. Despite the model trained with the segmented-dataset to achieve a worse performance, the Explainable Artificial Intelligence (XAI) approaches showed that it looks into more reliable regions when making decisions.

Abstract
The grape variety plays an important role in the wine production chain, thus identifying it is crucial for production control. Ampelographers, professionals who identify grape varieties through plant visual analysis, are scarce, and molecular markers are expansive to identify grape varieties on a large scale. In this context, Deep Learning models become an effective way to handle ampelographers scarcity. In this work, we explore the benefit of using deep learning vision transformers architecture relative to conventional CNN to identify 12 grapevine varieties using leaf-centred RGB images acquired in the field. We train an Xception model as a baseline and four different configurations of the ViT_B model. The best model achieved 0.96 of F1-score, outperforming the state-of-the-art convolutional-based model in the used dataset.

Abstract
Public lighting is crucial for maintaining the safety and well-being of communities. Current inspection methods involve examining the luminaires during the day, but this approach has drawbacks, including energy consumption, delay in detecting issues, and high costs and time investment. Utilising deep learning based automatic detection is an advanced method that can be used for identifying and locating issues in this field. This study aims to use deep learning to automatically detect burnt-out street lights, using Seville (Spain) as a case study. The study uses high-resolution night time imagery from the JL1-3B satellite to create a dataset called NLight, which is then divided into three subsets: NL1, NL2, and NT. The NL1 and NL2 datasets are used to train and evaluate YOLOv5 and YOLOv7 segmentation models for instance segmentation of streets. And then, distance outliers were detected to find the lights off. Finally, the NT dataset is used to evaluate the effectiveness of the proposed methodology. The study finds that YOLOv5 achieved a mask mAP of 57.7%, and the proposed methodology had a precision of 30.8% and a recall of 28.3%. The main goal of this work is accomplished, but there is still space for future work to improve the methodology.