关键词: 矢量海岸线, 形态模式判别, 预训练模型, BERT模型

Abstract:

Discriminating the morphological patterns of vector-based coastlines is vital for monitoring coastal evolution, marine disaster forecasting, and coastal zone planning, and it also serves as an important step in coastline cartography. Discriminating methods based on traditional machine learning technique rely on manually defined features and require large amounts of labeled samples with long-term training. To overcome these drawbacks, this study proposes a pre-trained model-based method that decouples generic geometric feature learning of coastlines from downstream morphological pattern discrimination. First, the coastlines are represented as Token sequences suitable for embedding learning using the operations of coordinate system resetting and coordinate normalization. Then, a self-supervised coordinate prediction task based on random masking is designed and integrated into the BERT model to construct a pre-trained model for the embedding learning of coastline geometric features. Finally, the pre-trained BERT model is fine-tuned with labeled dataset and transferred to the morphological pattern discrimination task. Based on open-source coastline data, a pre-trained dataset containing 195 649 samples and a labeled dataset with 1000 samples were collected. The proposed method achieves anF₁ score of 90.72% in a discrimination task involving five types of coastline morphological patterns, outperforming methods based on LSTM and 1D-CNN by 7.31%~9.38%.

Key words: vector-based coastline, morphological pattern discrimination, pre-trained model, BERT model