Fossil Pollen Detection Dataset (LIL-DEEP, Lake Latorița), v1

# Fossil Pollen Detection Dataset (LIL-DEEP, Lake Latorița) — v1 A custom annotated microscopy dataset for object detection of fossil pollen andnon-pollen palynomorphs (NPPs) from the LIL-DEEP sediment core, Lake Latorița,Southern Carpathians, Romania. ## Contents- **3,001 microscope images** (400× brightfield, Olympus CX41)- **8,460 bounding-box annotations** across **22 classes**- Six stratigraphic depths spanning Late Glacial → early Holocene assemblages- YOLO format (normalized `class cx cy w h`), with `data.yaml` ## Split (depth-based, leakage-free)| Split | Depths | Images ||-------|--------|--------|| train | 1101, 1102, 1104, 1148 | 2,523 || valid | 1012 | 120 || test | 1188 | 358 | Images are prefixed by depth (e.g. `1101_0001...`). No depth appears in morethan one split, so train/val/test are biologically independent. ## Classes (22)**Group A — pollen taxa (primary detection target):** Pine, Artemisia, Poaceae,Betula pendula, Chenopodiaceae, Picea, Alnus viridis, Rumex, Ulmus, Asteraceae,Cyperaceae, Apiaceae, Fagus, Salix, Other_pollen. **Group B — non-pollen palynomorphs (reported separately, after van Geel 2001):**Type-128 (HdV-128), Lycopodium, Charcoal, Pediastrum integrum, Pediastrumboryanum, Pinus stomata, Other_NPP. `Other_pollen` and `Other_NPP` aggregate rare taxa (< ~30 instances). Asteraceaesubtypes (Achillea, Aster-type, Cirsium-type, Liguliflorae, Senecio-type) weremerged into a single Asteraceae class. ## Use```pythonfrom ultralytics import YOLOmodel = YOLO("yolov8l.pt")model.train(data="data.yaml", epochs=100, imgsz=640)``` ## LicenseCC BY 4.0