5:15 PM - 7:15 PM
[MIS15-P13] Evaluating object detection for counting fossil diatoms: the effects of fossil preservation and morphological variation on detection accuracy
Keywords:fossil diatoms, object detection, Deep learning, the Southern Ocean, Paleoenvironmental indicator
The assemblage and morphological analyses of fossil diatoms serve as crucial indicators for biostratigraphy and paleoenvironmental reconstruction. Traditional methods rely on manually classifying and counting fossils mounted on permanent slides, a time-consuming and labor-intensive process. This process is not unique to fossil diatoms but extends to other microorganism and microfossil studies, leading to increasing interest in the application of machine learning to automate microfossil classification. In particular, the deep-learning-based object detection approach has demonstrated potential in extracting intended fossils from sediment samples. However, studies on object detection applied to fossil diatoms in sediment samples remain limited. Fossil diatom assemblages in sediments often exhibit intraspecific morphological diversity and varying degrees of valve fragmentation due to biogeographic and sedimentological conditions. Automating the manual counting of these complex assemblages using object detection requires an assessment of whether the trained model can accurately count diatom fossils, including fragmented countable valves, as well as an evaluation of effective training approaches. Here, we evaluated the impact of fossil preservation and intraspecific morphological variation on the accuracy of an object detection model as an experimental application of an automated detection method for Eucampia antarctica, an endemic species of the Southern Ocean, from marine sediments.
We selected four training sites based on fossil preservation (good/ moderate) and the proportion of large E. antarctica valves (moderate/ rare). Detection models (YOLOv5-x) were trained using datasets from each of the four training sites, as well as combinations of two sites. Detection accuracy for all trained models was evaluated across fourteen test sites in the Southern Ocean that were not included in the training datasets. All microscope images were obtained by automatically capturing permanent diatom slides using high-resolution slide scanners.
Detection of E. antarctica valves at the test sites using models trained on each of the four training sites resulted in mean recall and precision values exceeding 0.85. Intraspecific morphological variation of E. antarctica across the four training sites had minimal influence on detection accuracy in the test sites. Among the four training sites, the dataset with well-preserved and moderately abundant large E. antarctica valves demonstrated lower recall values, indicating that excluding images with non-target sediment particles from the training dataset may lower detection accuracy. High detection accuracy was shown even at test sites distant from the training sites; however, low accuracy was observed at certain test sites regardless of the model used. These results support that differences in detection accuracy are influenced by variations in particle density on the slides and differences in species composition other than the target species. Furthermore, models trained with combinations of two training sites improved detection accuracy compared to those trained with a single training site, achieving performance comparable to manual annotations. Our results indicated that even models trained on samples from a limited number of sites can achieve high detection accuracy for E. antarctica, including valves with intraspecific morphological variation and fragmentation, on previously uninvestigated sediment samples.
We selected four training sites based on fossil preservation (good/ moderate) and the proportion of large E. antarctica valves (moderate/ rare). Detection models (YOLOv5-x) were trained using datasets from each of the four training sites, as well as combinations of two sites. Detection accuracy for all trained models was evaluated across fourteen test sites in the Southern Ocean that were not included in the training datasets. All microscope images were obtained by automatically capturing permanent diatom slides using high-resolution slide scanners.
Detection of E. antarctica valves at the test sites using models trained on each of the four training sites resulted in mean recall and precision values exceeding 0.85. Intraspecific morphological variation of E. antarctica across the four training sites had minimal influence on detection accuracy in the test sites. Among the four training sites, the dataset with well-preserved and moderately abundant large E. antarctica valves demonstrated lower recall values, indicating that excluding images with non-target sediment particles from the training dataset may lower detection accuracy. High detection accuracy was shown even at test sites distant from the training sites; however, low accuracy was observed at certain test sites regardless of the model used. These results support that differences in detection accuracy are influenced by variations in particle density on the slides and differences in species composition other than the target species. Furthermore, models trained with combinations of two training sites improved detection accuracy compared to those trained with a single training site, achieving performance comparable to manual annotations. Our results indicated that even models trained on samples from a limited number of sites can achieve high detection accuracy for E. antarctica, including valves with intraspecific morphological variation and fragmentation, on previously uninvestigated sediment samples.