●YOLOv8 was better than Faster R-CNN in Collembola detection,attaining 97%precision.●Precision varied widely(18.51%-99.17%)with Collembola diversity at genus-level.●Our study shows deep learning’s potential for ra...●YOLOv8 was better than Faster R-CNN in Collembola detection,attaining 97%precision.●Precision varied widely(18.51%-99.17%)with Collembola diversity at genus-level.●Our study shows deep learning’s potential for rapid and precise ecology evaluation.Deep learning methods are increasingly vital for species classification across animal species.However,the arthropod classification and detection of deep learning remain underexplored,especially soil arthropods.In this study,we collected 5020 images of collembolan individual as the origin dataset,spanning 6 families,10 genera,and 51 species in China.By employing the cut-paste method,we created community image datasets of collembolan with diversity gradient to train and evaluate YOLOv8 and Faster R-CNN models for collembolan identification using deep learning methods.Our classification model achieved 83.05%precision and detection models exceeded 97%.YOLOv8 models were the most effective,with a higher top-1 precision all over 80%at the family level,compared to the Faster R-CNN model,which had a minimum precision of 60%.Using community datasets of genera belonging to Isotomidae or Entomobryidae family to train YOLOv8 models,we observed precision ranging from 18.51%to 99.17%,reflecting the changes of detection performance based on Collembola diversity.The YOLOv8 models excelled in identifying genera within Isotomidae family compared to Entomobryidae family.Our study pioneers the application of the YOLOv8 deep learning model for rapid detection and classification of Collembola,while proposing an adaptive training strategy specifically designed for diversity gradient samples.Our research demonstrates the potential of deep learning for rapid,accurate Collembola identification and extends to broader ecological assessments.展开更多
基金supported by the National Natural Science Foundation of China(Grant Nos.41977057,32201421,32361143523)Shenzhen Science and Technology Program(Grant No.JCYJ20240813151036047)research grant of the Sun Yat-sen University Bairen Plan(Grant No.77010-18841290).
文摘●YOLOv8 was better than Faster R-CNN in Collembola detection,attaining 97%precision.●Precision varied widely(18.51%-99.17%)with Collembola diversity at genus-level.●Our study shows deep learning’s potential for rapid and precise ecology evaluation.Deep learning methods are increasingly vital for species classification across animal species.However,the arthropod classification and detection of deep learning remain underexplored,especially soil arthropods.In this study,we collected 5020 images of collembolan individual as the origin dataset,spanning 6 families,10 genera,and 51 species in China.By employing the cut-paste method,we created community image datasets of collembolan with diversity gradient to train and evaluate YOLOv8 and Faster R-CNN models for collembolan identification using deep learning methods.Our classification model achieved 83.05%precision and detection models exceeded 97%.YOLOv8 models were the most effective,with a higher top-1 precision all over 80%at the family level,compared to the Faster R-CNN model,which had a minimum precision of 60%.Using community datasets of genera belonging to Isotomidae or Entomobryidae family to train YOLOv8 models,we observed precision ranging from 18.51%to 99.17%,reflecting the changes of detection performance based on Collembola diversity.The YOLOv8 models excelled in identifying genera within Isotomidae family compared to Entomobryidae family.Our study pioneers the application of the YOLOv8 deep learning model for rapid detection and classification of Collembola,while proposing an adaptive training strategy specifically designed for diversity gradient samples.Our research demonstrates the potential of deep learning for rapid,accurate Collembola identification and extends to broader ecological assessments.
