| 摘要: |
| 针对印太交汇区岛屿密布、地形调制、弱梯度锋面识别难的问题，提出一种融合多日时序信息的自适应锋面探测神经网络模型，实现了温度锋与叶绿素锋的有效识别。该模型采用不依赖人工标注的自适应学习策略，通过时序信息和空间结构约束增强锋面识别的稳定性。船测走航数据验证表明，该模型在温度锋识别中总体命中率为70.00%，略低于经过全局优化的改进型直方图探测算法（Cayula and Cornillon Algorithm coupled with IDW and Mathematical morphology operators, CCAIM）的82.59%，但精度（51.63%）和F1指数（59.43%）均显著优于CCAIM（33.07%、47.23%）；在弱锋面条件下，模型命中率为70.77%（略低于CCAIM的77.40%），精度为34.38%（高于CCAIM的30.85%），F1指数达46.28%（优于CCAIM的44.12%），表明模型在弱梯度背景下具有更好的稳定性与可信度。基于该模型，构建了1998-2022年印太交汇区温度锋与叶绿素锋数据集，系统分析了两类锋面的时空分布特征。高频锋区主要分布于岛屿陆坡、主要海峡出口及西边界流汇合区域，复杂地形与强动力过程对锋面形成具有重要作用。以苏拉威西海为典型涡旋活跃区，进一步补充了该区温度锋和叶绿素锋与中尺度涡旋背景之间的统计联系，温度锋频率与相对涡度呈约4个月滞后的显著正相关（r=0.783），叶绿素锋频率与相对涡度则呈极强的同步相关（r=0.937）。两类锋面虽均受中尺度涡旋活动调制，但响应时间尺度存在明显差异：温度锋反映涡旋对热结构的逐步重塑，叶绿素锋则更直接响应于涡旋诱导的垂向输送过程。 |
| 关键词: 海洋锋 印太交汇区 自适应探测 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| ADAPTIVE INTELLIGENT DETECTION AND ANALYSIS OF OCEAN FRONTS IN THE INDO-PACIFIC CONVERGENCE ZONE |
|
Liunan, Gaole
|
|
Institute of Oceanology, Chinese Academy of Sciences
|
| Abstract: |
| To address the challenges of densely distributed islands, topographic modulation, and weak-gradient front detection in the Indo-Pacific Convergence Zone, this study proposes an adaptive neural network model for ocean front detection that integrates multi-day temporal information, enabling effective identification of both thermal and chlorophyll fronts. The model adopts an adaptive learning strategy that does not rely on manually labeled samples, and enhances the stability of front detection through temporal information and spatial-structure constraints. Validation using shipborne underway observations shows that, for thermal front detection, the proposed model achieves an overall hit rate of 70.00%, slightly lower than that of the globally optimized improved histogram-based detection algorithm, the Cayula and Cornillon Algorithm coupled with IDW and Mathematical morphology operators (CCAIM; 82.59%). However, its precision (51.63%) and F1 score (59.43%) are substantially higher than those of CCAIM (33.07% and 47.23%, respectively). Under weak-front conditions, the model achieves a hit rate of 70.77%, slightly lower than that of CCAIM (77.40%), but its precision reaches 34.38%, higher than that of CCAIM (30.85%), and its F1 score reaches 46.28%, exceeding that of CCAIM (44.12%). These results indicate that the proposed model provides greater stability and reliability under weak-gradient conditions.Based on this model, a dataset of thermal and chlorophyll fronts in the Indo-Pacific Convergence Zone from 1998 to 2022 was constructed, and the spatiotemporal characteristics of the two types of fronts were systematically analyzed. High-frequency frontal zones are mainly distributed along island slopes, major strait exits, and western boundary current convergence regions, indicating that complex topography and strong dynamic processes play important roles in front formation. Taking the Sulawesi Sea as a representative eddy-active region, this study further examines the statistical relationship between thermal and chlorophyll fronts and the mesoscale eddy background. Thermal front frequency shows a significant positive correlation with relative vorticity at a lag of approximately four months (r = 0.783), whereas chlorophyll front frequency exhibits a very strong synchronous correlation with relative vorticity (r = 0.937). Although both types of fronts are influenced by mesoscale eddy activity, their response timescales differ markedly: thermal fronts reflect the gradual reshaping of upper-ocean thermal structure by eddies, whereas chlorophyll fronts respond more directly to eddy-induced vertical transport processes. |
| Key words: oceanic front ? Indo-Pacific Convergence Zone ? adaptive detection |