5:15 PM - 7:15 PM
[AOS12-P03] Long-term variation in vertical temperature profiles in the Northwest Pacific using unsupervised clustering
Keywords:Northwest Pacific, Argo data, unsupervised clustering, temperature profile
We analyzed the northwestern Pacific Ocean's oceanic structure by applying the Profile Classification Model (PCM) to Argo float data. PCM is a type of machine learning that uses unsupervised clustering to classify vertical ocean profiles. It was first developed by Maze et al. (2017) and initially applied to the North Atlantic. This method categorizes ocean profiles based on their vertical structure, Maze et al. showed the distribution of each class was concentrated in specific regions on the map. It means that this classification allows us to identify areas with specific water properties and locate boundaries where oceanic conditions change.
In the northwestern Pacific, PCM was applied by Sambe and Suga (2022), who characterized each class. Based on temperature, salinity, density, and the spatial distribution of each class, they identified the following classifications: Class 1: Subarctic water, Class 2: A mix of subtropical and subarctic water (transition zone), Class 3: Transition regions and the Subarctic Frontal Zone (SAFZ), Class 4: Subtropical Mode Water (STMW), Class 5: The lighter variety of Central Mode Water (L-CMW). However, no studies have examined the long-term changes in profiles within each class. Therefore, this study uses PCM to investigate long-term trends at different depths within the classified profiles.
We used 62,743 Argo profiles collected up to September 2024 from the Global Data Assembly Center (GDAC). The study region was defined as the northwestern Pacific (30°N–60°N, 140°E–180°E). Since the measurement depths vary across profiles, we interpolated the data at 5-meter intervals from 5 m to 1000 m using the Akima method. The analysis assumed that Argo float data are randomly and evenly distributed in time and space. The classification was performed using PCM as mentioned earlier. Additionally, PCM provides a labeling metric that quantifies the certainty of class assignments, where values close to 1 indicate high certainty and values close to 0 indicate low certainty. We also analyzed data using this metric. Long-term trends were calculated as annual averages at each depth using the least squares method.
We focused on the long-term trend of water temperature at each depth in each class. Class 1, class 2, class 3, and class 4 showed interannual and decadal-scale variations, while class 5 showed a significant positive long-term trend (up to 0.09 °C/year) at many depths. The same analysis was performed for profiles with labeling metric values of 1.0-0.99. Comparing the results with the aforementioned results, the regression coefficients were larger at some depths for class 1, class 2, and class 5, but not significantly different for class 3 and class 4. As a next step, we will conduct similar analyses focusing on seasons and specific regions.
In the northwestern Pacific, PCM was applied by Sambe and Suga (2022), who characterized each class. Based on temperature, salinity, density, and the spatial distribution of each class, they identified the following classifications: Class 1: Subarctic water, Class 2: A mix of subtropical and subarctic water (transition zone), Class 3: Transition regions and the Subarctic Frontal Zone (SAFZ), Class 4: Subtropical Mode Water (STMW), Class 5: The lighter variety of Central Mode Water (L-CMW). However, no studies have examined the long-term changes in profiles within each class. Therefore, this study uses PCM to investigate long-term trends at different depths within the classified profiles.
We used 62,743 Argo profiles collected up to September 2024 from the Global Data Assembly Center (GDAC). The study region was defined as the northwestern Pacific (30°N–60°N, 140°E–180°E). Since the measurement depths vary across profiles, we interpolated the data at 5-meter intervals from 5 m to 1000 m using the Akima method. The analysis assumed that Argo float data are randomly and evenly distributed in time and space. The classification was performed using PCM as mentioned earlier. Additionally, PCM provides a labeling metric that quantifies the certainty of class assignments, where values close to 1 indicate high certainty and values close to 0 indicate low certainty. We also analyzed data using this metric. Long-term trends were calculated as annual averages at each depth using the least squares method.
We focused on the long-term trend of water temperature at each depth in each class. Class 1, class 2, class 3, and class 4 showed interannual and decadal-scale variations, while class 5 showed a significant positive long-term trend (up to 0.09 °C/year) at many depths. The same analysis was performed for profiles with labeling metric values of 1.0-0.99. Comparing the results with the aforementioned results, the regression coefficients were larger at some depths for class 1, class 2, and class 5, but not significantly different for class 3 and class 4. As a next step, we will conduct similar analyses focusing on seasons and specific regions.