Application of YOLOv11 deep learning model for classification and counting ice-rafted debris (IRD) in core sediments in the Arctic Ocean

The classification and quantification of ice-rafted debris (IRD) in marine sediments are key to reconstructing glacial-interglacial dynamics and sediment provenance. However, traditional IRD analysis, based on manual grain identification under binocular microscopes, is time-consuming and dependent on expertise, which acted as a barrier to entry for IRD research. Here, we present a deep learning-based framework for the automated detection and lithological classification of IRD from high-resolution microscopic images with grains from natural Arctic sediments. Using the YOLOv11 algorithm, we designed a two-stage model: an instance segmentation model (Model 1) that detects individual IRD from multi-grain images, and a classification model (Model 2) that categorizes each grain into one of four lithological types (i.e., quartz, detrital carbonate, clastic, and crystalline). The dataset comprises 110 images containing 9,642 grains from the Chukchi Sea sediment core ARA14C-ST12. Model 1 achieved the detection performance with precision = 0.95, recall = 0.97, m AP50 = 0.98, m AP50-95 = 0.85, and F1 score = 0.96, demonstrating high model performance to detect complex morphology of grain. The evaluation metric of Model 2, used to identify lithological classes, showed average Top-1 accuracy of 0.87 and 0.75 on the validation and test sets, respectively. The classification model showed reliable recognition for quartz and detrital carbonate, with moderate confusion between clastic and crystalline grains. These results demonstrate that the proposed YOLOv11-based approach enables rapid, reproducible, and objective lithological classification of IRD grains, providing an efficient alternative to conventional manual counting. • •Two-stage YOLOv11 framework detects and classifies. • •High detection performance: Precision = 0.95, Recall = 0.97, mAP50 = 0.98. • •Classifier achieves Top-1 accuracy 0.87 (validation) and 0.75 (test). • •Reduce expert labor by automating multi-grain counting and lithology of IRD.