Artificial Intelligence applied to the automated detection and identification of Devonian miospores

This repository provides the dataset and code for miospore detection and classification. Detection is performed with YOLOv11, and classification is implemented with VGG16, ResNet-18, and EfficientNet-B0. The dataset includes microscope images and corresponding annotations for three Devonian miospore species: Teichertospora torquata, Geminospora lemurata, and Samarisporites triangulatus, from the Ertych section (Central Armenia). Annotations were created using the Roboflow platform (https://roboflow.com/) and are provided in YOLOv11 PyTorch format. These data were used to train a YOLOv11 detection model that achieves mAP@0.5 = 0.995 (mean average precision at an Intersection-over-Union threshold of 0.5). The same images were also used to train classification models that identify the three miospore species listed above. Table 1. Summary of classification results on the image + area (size proxy) dataset using EfficientNet-B0, ResNet-18, and VGG16. Values are reported as mean ± standard deviation over 10 runs. Accuracy and macro-F1 are given in percent; MCC (Matthews correlation coefficient) is unitless. Architecture EfficientNet-B0 ResNet-18 VGG16 Inputs Images + area size Number of runs 10 Epochs per run 30 Best validation accuracy (mean, %) 74.6 76.2 81.2 Average test accuracy (%) ± std 79.6 ± 0.1048 87.9 ± 0.0797 87.5 ± 0.0665 Average test macro-F1 (%) ± std 76.4 ± 0.1305 86.9 ± 0.1033 87.3 ± 0.0698 Average test MCC ± std 0.75 ± 0.1315 0.85 ± 0.0920 0.84 ± 0.0829 Average Per-Class Accuracy (%) 86.25 84.5 85.5 Requirements and installation We recommend using Python 3.10+ and a GPU-enabled environment (e.g. Google Colab or a local machine with CUDA). Install the required packages with: pip install -r requirements.txt The requirements.txt file includes (among others): - torch, torchvision- ultralytics- efficientnet_pytorch- pandas, numpy, scikit-learn- matplotlib, Pillow