渤海海域新生代孢粉化石智能识别

Intelligent identification of Cenozoic spore and pollen fossils in Bohai Sea area

通过鉴定古生物化石类别信息和分布情况,可以为地质年代、古沉积环境及油气勘探工作提供重要信息。但传统古生物化石鉴定工作耗时耗力,人工依赖性高,难以满足当前快速勘探评价的需要。鉴于孢粉化石图像数量有限、属种分类多、具有科、属、种的特定分类逻辑等特点,围绕孢粉化石图像处理、化石图像筛选、化石目标检测、化石分类识别等方面,通过利用目标检测深度学习、标签松弛等技术,改进了有效化石筛选和孢粉化石分类识别的智能化水平。以渤海海域浅层新生代孢粉化石鉴定为例,采用YOLOv5和DenseNet等神经网络开发了一套孢粉化石智能识别方法,其平均识别准确率达94%,基本满足了孢粉化石鉴定实际生产准确性要求,可以辅助人工开展古生物化石鉴定工作。该方法将各种深度学习技术与古生物领域专业知识有效结合,并从数据和模型2个角度相结合,提高了识别模型的泛化能力与识别精度,并得以实际应用,使得能够在减少时间人力成本的前提下提供准确的鉴定结果,证实了人工智能技术在传统古生物鉴定领域的可行性。

The identification of paleontological fossil types and their distribution provides important information for geochronological,paleoenvironmental studies,and oil and gas exploration.However,traditional fossil identifi-cation methods are time-consuming,labor-intensive,and highly dependent on manual efforts,making it difficult to meet the current demand for rapid exploration and evaluation.Given the limited number of spore and pollen fossil images,the complex classification of taxa,and the specific taxonomy of family,genus,and species,this research focused on improving the automation for fossil image processing,image screening,object detection,and classification.By utilizing techniques such as deep learning for object detection and label smoothing,the efficiency of fossil screening and spore and pollen fossil classification was significantly enhanced.Taking the identification of the Cenozoic spore and pollen fossils from the Bohai Sea shallow area as a case study,a set of intelligent identification methods was developed using neural networks such as YOLOv5 and DenseNet,with an average identification accuracy of 94%,basically meeting the practical accuracy requirements for fossil identification in production.The system could assist in the manual identification of paleontological fossils.By effectively combining various deep learning techniques with specialized knowledge in paleontology,the generalization ability and recognition accuracy of the identification model were improved from both data and model perspectives.Its successful application demonstrates the feasibility of artificial intelligence in the traditional field of paleontological fossil identification,reducing time and labor costs while providing accurate results.