ML assisted drumlin inventory: YOLOv11 segmentation, polygon reconstruction, and accuracy assessment

Drumlins are key indicators of subglacial processes and former ice-flow patterns, but regional drumlin inventories are often compiled manually, limiting scalability and reproducibility. This contribution presents a workflow for drumlin segmentation, prediction, and GIS-ready analysis using deep learning applied to digital elevation models (DEM) and an existing, manually mapped, drumlin reference shapefile from Latvian drumlin fields. The DEM is tiled into 640×640 image patches, and reference polygons are converted into You Only Look Once (YOLO)(Ultralitics S.a.) segmentation labels. A YOLOv11 segmentation model is trained for drumlin delineation and then used to generate predictions across the tiled DEM. Predicted outputs (TXT masks) are converted back into georeferenced polygons and exported as GeoPackages, enabling immediate integration with standard GIS-based morphometric analysis and mapping methods.

Model performance is evaluated using both standard YOLO metrics and an inventory comparison against the control dataset. The statistics for bounding boxes, precision and recall reach 0.808 and 0.652, with mAP50 of 0.755 (mAP50–95: 0.514). For segmentation masks, precision and recall are 0.757 and 0.601, with mAP50 of 0.672 (mAP50–95: 0.275). Inventory comparison yields 1190 predicted vs 1146 control drumlins, with 906 true positives, 284 false positives, and 247 false negatives, corresponding to precision 0.761, recall 0.786, and F1 0.773 (nMCC: 0.387).

The results demonstrate that YOLO-based segmentation can produce georeferenced drumlin polygons at scale with quantified uncertainty, providing a practical route toward repeatable drumlin inventories and downstream geomorphological analyses.

This research was funded by the Latvian Council of Science, project "Reconstruction of ice stream dynamics and deglaciation of the SE sector of the Scandinavian Ice Sheet in Latvia", project No.lzp-2024/1-0193