Introduction The identification of pollen grains provides valuable information for a diversity of fields, such as pollination ecology, paleobotany, melissopalynology or allergology. But counting and identifying pollen grains manually on a light microscope is time consuming and requires expert knowledge. Automatization of these labour-intensive tasks first emerged in the 60’s and has tremendously progressed with the development of computer vision with deep-learning algorithms. Most studies so far have managed to classify images containing one single pollen grain. Prior to classification, detection of pollen grains in images containing several is recent in palynology and has been performed under ideal conditions. For example, Kubera et al. (2022) selected images with pollen grains from three taxa, showing only whole pollen grains without neither pollen fragments nor debris. We test here how well the combination of the detection from YOLOv5 algorithm (Jocher et al. 2022) and the classification from ResNet50 algorithm (Chollet et al. 2015) performs in detecting and classifying pollen grains in images containing numerous Mediterranean pollen taxa and many different types of debris, sampled from annual pollen traps. Evaluation is two-fold, it consists in detecting and classifying (1) two pollen morphs from the European wild and cultivated grapevine, Vitis vinifera, and (2) all pollen taxa, in images taken from pollen traps samples. Vitis vinifera is of high ecological and archaeological value as its domestication has shaped Mediterranean landscapes and social interactions since the Neolithic (Naqinezhad et al. 2018). Its pollen is tricolporate on wild male and cultivated hermaphroditic individuals, while it is inaperturate and sterile on wild female individuals. This inaperturate morph is often unidentified by palynologists, and yet potentially key for inferring stages of the Vitis vinifera domestication (Mercuri et al. 2021). Materials and Methods Annual and monthly pollen traps were located nearby wild and cultivated individuals established in the Mediterranean Massif of Pic-Saint-Loup (Hérault, France). We trained YOLOv5 (Jocher et al. 2022) to detect pollen grains on 1,200 images in which 3,700 pollen grains were manually detected. We then applied the trained YOLOv5 on 18,000 new images and automatically detected ~53,000 pollen grains. We used the Particle Trieur software (Marchant et al. 2020) to manually classify 10,000 of these detected pollen grains, and used them for training the classification ResNet50 algorithm. Performance is evaluated for YOLOv5 and ResNet50 independently, with standard metrics (precision, recall) and from the confusion matrix (false positive and negative). Results and Discussion The detection of pollen grains sampled from pollen traps by YOLOv5 resulted in less than 10% pollen grains left undetected. The classification performance of ResNet50 is currently evaluated on the 10,000 manually classified pollen grains. We will improve those results by increasing our training dataset for detection, and will soon evaluate the performance of the classification method between the two pollen morphs of Vitis.