Dataset and script used for the publication: From seeding to detachment: leveraging deep learning to quantify the transport of tyre wear microplastics in a wind tunnel

Background This dataset contains data from tyre wear microplastic detachment experiments conducted in a small boundary-layer wind tunnel at the University of Bayreuth. Image data were acquired using a combination of two cameras: a Sony Alpha 7R II full-frame visual camera and a Chronos 1.4 high-speed camera equipped with a K2 DistaMax long-distance microscopy lens. These cameras were used to capture particle positions before (for training data) and during detachment experiments. The images were later processed using image segmentation, particle tracking, and detachment analysis workflows. The experiments were conducted to investigate the detachment behavior of tyre wear particles in an idealized system using glass surface. Dataset description This dataset contains processed data, image data, trained model weights, and Python scripts used for the detachment analysis. The dataset includes training and validation images used for YOLOv8 segmentation, detachment image sequences from eight experimental replicates, processed CSV files, and analysis scripts. The main data records included in this repository are: Detachment_Exp_Images_8_REP.zip This compressed folder contains the detachment image sequences from the eight experimental replicates, labelled from R1 to R8. These images were used for the detachment analysis. merged_all_samples.csv This CSV file contains the processed particle-level data obtained from the eight replicate experiments. It includes the output generated from the particle-tracking workflow and was used for the detachment analysis. weights/ This folder contains the trained YOLOv8 model weights. The best.pt file represents the trained model checkpoint with the best validation performance and was used for image prediction, particle segmentation, and detachment analysis. Training/ This folder contains training-related datasets and outputs, including initial training, fine-tuning, validation, inference, and visualization materials used during YOLOv8 model development. Code description The Python scripts included in the 'script' zip folder are example scripts used for image preparation, YOLOv8 training, image prediction, particle tracking, and detachment analysis. Image_cropping_mask_generation.py This script was used to prepare image and mask datasets for YOLOv8 training. The script reads the original images, generates binary masks using manual thresholding, morphological opening, and image inversion, and then divides each image into 16 non-overlapping sub-images using a 4 × 4 cropping strategy. This was used to increase sample diversity for the training and validation dataset. Yolov8_training.py Used to train and fine-tune a YOLOv8 segmentation model for particle segmentation. The script loads the selected YOLOv8 model weights, defines the training hyperparameters, trains the model using the prepared dataset, and validates the trained model after training. Yolov8_prediction_visualization.py This script was used to load the trained YOLOv8 model weights and run prediction on all images in a selected folder. The script saves the predicted and annotated images to an output folder. This allows visual inspection of the segmentation performance before using the results for particle tracking and quantitative analysis. Particle_tracking.py Script was used to detect, filter, and track particles across image sequences. The script loads the trained YOLOv8 segmentation model, applies it to each image frame, extracts predicted masks, compares predicted masks with generated ground-truth masks using the Dice score, and calculates particle properties. Note that this script is not design for particle motion tracking. detachment_ustar.py Used to analyse particle detachment as a function of friction velocity. The script calculates the detachment fraction from the number of particles remaining in each frame, normalizes the detachment fraction for each sample, bins the normalized detachment data, and fits both linear OLS and logistic models, and the U*50 was computed accordingly. The same general workflow can be adapted for size and shape resolved analysis. Usage notes and disclaimer Users should update the relative paths accordingly before running the code. For subsequent reuse or retraining of the model, the "best.pt" file is reccommended for usage of the trained YoloV8 weight. Mask images should also be generated before running workflows that require comparison between the predicted and reference mask. These scripts is explicitly used for particle detection, segmentation and subsequent particle tracking via matching detected particles across sequential image frames. Therefore, it should not be misinterpreted as a particle motion tracking tool. Users who want to analyze particle trajectories, velocities and motion dynamics might adapt this model as a baseline, reconfigure or write a dedicated tracking algorithm. The user should kindly also note that the output may be affected by image quality, segmentation performance, mask quality, threshold settings, and particle-matching assumptions.