The North Pacific High (NPH) is a fundamental meteorological feature present during the boreal warm season. Marine boundary layer (MBL) clouds, which are persistent in this oceanic region, are influenced directly by the NPH. In this study, we combine 11 years of reanalysis and an unsupervised machine learning technique to examine the gamut of 850
Low, stratiform clouds that develop in the marine boundary layer (MBL) are of significant interest to the atmospheric science community because they impact meteorological forecasts and, ultimately, a host of human activities
During boreal summer, the northeast Pacific Ocean (NEP) is home to one of the largest MBL stratiform cloud decks
Often called coastally trapped disturbances (CTDs), these mesoscale phenomena develop in response to the reversed pressure gradient and are characterized by southerly MBL flow and a redevelopment of the stratiform cloud deck
To diagnose the various NPH circulations, we first use the National Centers for Environmental Prediction (NCEP) North American Regional Reanalysis (NARR) to develop a self-organizing map (SOM) covering the western United States and the NEP. Our study domain is shown in Fig. 1. We then examine measurements from the Aqua Moderate Resolution Imaging Spectroradiometer (MODIS). Two important variables –
Study region covering the western United States and the NEP for the SOM analysis. United States abbreviations WA, OR, and CA represent Washington, Oregon, and California, respectively, while MX represents Mexico. The red and yellow stars denote the locations of buoys 46013 (Bodega Bay) and 46022 (Eel River), respectively.
The SOM is a type of neural network that uses a competitive, unsupervised machine learning technique
We choose to employ the SOM technique due to its ability to group similar patterns and therefore reveal dissimilar patterns that may be hidden in the large NARR data set considered here. To this end, we use the MATLAB SOM Toolbox (version 2.1) to generate the SOM using the batch algorithm. This algorithm follows the well-known Kohonen technique Define the number of nodes and iterations (one iteration is defined as a single pass through all of the input data vectors), in addition to the neighborhood radius. From the data set, determine the two eigenvectors that have the largest eigenvalues; initialize the SOM node weights linearly along these eigenvectors to provide a first approximation of the input data set. Present all vectors from the input data and calculate the Euclidean distance between each input vector and each node, where the Euclidean distance Update the neighborhood radius. Determine the node that most closely matches each input vector; the winning node is characterized by the minimum Euclidean distance. Update the weight of each node – where the new weight is equal to the weighted average of each input data vector to which that node or any nodes in its neighborhood responded – after a single iteration. Repeat steps 3–6 for
Choosing the number of nodes is critical because a map with too few nodes yields larger sample sizes but insufficient detail, while one with too many nodes yields greater detail but insufficient sample sizes. For the present study, a series of sensitivity tests is conducted using different node map sizes to determine an optimal number of nodes (Fig. 2). Quantization and topographic errors (QE and TE, respectively) for each map are calculated. The QE, which is a measure of map resolution, is equal to the average distance between each input vector and the best matching node, while the TE indicates map topology preservation by determining the percentage of input vectors whose first and second best matching nodes are not adjacent. As the number of nodes increases, the QE decreases, typically at the cost of sacrificing node topology. This trade-off is shown quite well in Fig. 2: the QE decrease is most pronounced as the number of nodes increases from approximately 9 to 20, and the TE increase is most notable above approximately 30 nodes. Moreover, using a nonuniform (rectangular) map appears to reduce the TE, which supports previous work showing the superiority of rectangular maps over square maps
The quantization (red circles; left axis) and topographic (blue diamonds; right axis) errors for each SOM configuration tested in this study. SOM node topologies (rows
Similar to previous work
Synoptic-scale 850
In this study, we consider afternoon satellite measurements from Aqua MODIS because we use 00:00 UTC NARR grids to generate the SOM. The satellite images, which are typically retrieved between 20:30 and 23:30 UTC, are paired with the NARR grid for the next day. For instance, we link the MODIS retrieval from 22:00 UTC on 5 July 2010 to the NARR grid from 00:00 UTC on 6 July 2010. Even in the instance where the time difference between a MODIS image and NARR grid is a maximum (approximately 3.5
For the MODIS retrievals, values of
Percentage of southerly flow hours recorded at each buoy site along the California coastline for each node in the
We now use the SOM output to investigate the various NARR 850
Summary statistics for SOM node frequency. Total and monthly frequencies over the 11-year period are shown.
Large-scale regimes associated with both offshore continental flow driven by the NPH (e.g., node 5) and onshore continental flow driven by a land-falling cyclone (e.g., node 16) at 850
MODIS estimation of
As in Fig. 5 except for the MODIS retrieval of
Measurements from buoy 46013 (Bodega Bay), which is located just northwest of Point Reyes, California, suggest that southerly flow is present for a substantial number of hours (
Table 1 lists the total and monthly frequencies of occurrence for each node. In general, the majority of days that are represented by the land-falling cyclone regime (nodes 1, 6, 11, and 16) are in early summer (June) and early fall (September). This is not surprising because these systems are more common during transition seasons than during summer
Due to the nature of the SOM, adjacent synoptic-scale patterns are similar to one another, and there is a gradual transition between the nodes as one moves across the SOM. The SOM patterns farther left on the map are associated with generally strong westerly flow offshore and divergent flow near the coastline due to a dominant cold-core land-falling cyclone. Conversely, those patterns toward the right feature northerly, and even northeasterly, flow offshore due to a dominant warm-core NPH. Moreover, several of the nodes (3, 4, and 5) feature a noticeably weak 850
Figures 5 and 6 show the mean
Although the MODIS retrievals are not used directly to generate the SOM, and instead are simply associated with the corresponding days in each node, there is an apparent connection between the various synoptic-scale patterns in the 850
Frequency distributions of
Evident in all of the SOM nodes is a region of high
Frequency distributions reveal that between the various SOM nodes, cloud micro- and macrophysical properties exhibit a broad range that is dependent on the prevailing synoptic-scale pattern (Fig. 7; see Table 2 for median values). The distributions confirm that node 5, in addition to nodes 3, 4, and 10, represents the scenarios where MBL clouds are characterized by relatively high
Summary statistics for SOM node meteorological and cloud properties. We tabulate median values of the frequency distributions of
To explore the potential impact of the regional meteorology associated with each of the synoptic-scale patterns – compared to simply the abundance of aerosol – on the observed cloud properties, we also examine low cloud fraction (LCF), lower tropospheric stability (LTS;
Susceptibility parameters calculated for each node. Here we use the same data that are used to calculate the values in Table 2; that is, the susceptibility for each node is based on the linear regression of all spatial points located within the oceanic area shown in Figs. 3, 5, and 6 and of all MODIS retrieval times associated with that node.
For this susceptibility analysis, we consider the MODIS variables
In Fig. 8, we summarize the relationship between
We emphasize that Fig. 8 elucidates the large range in LWP for the NPH regime, which is in stark contrast to the relatively narrow range in LWP for the land-falling cyclone regime. As a result, for low LWP, the transition from the land-falling cyclone to NPH regime results in little to no change in LWP but a drastic increase in
Scatter plot of LWP versus
Overall, our SOM results elucidate the apparent coupling between NPH dynamics and mesoscale MBL cloud properties through both meteorological and aerosol effects. In terms of the impact of large-scale circulation on cloud physics through aerosol forcing, generally weak flow and/or an enhancement in offshore continental flow at 850
Through the use of a SOM, we show that the location and intensity of the NPH, as well as the presence of land-falling low-pressure systems, play a role in modifying MBL cloud micro- and macrophysical properties offshore of the western United States during boreal summer. The 850
The findings reported here may be of significant interest to atmospheric science communities utilizing climate models (CMs) because the synoptic-scale flow–cloud microphysics relationships from the SOM may be used to test CMs and probe uncertainties in their simulation of aerosol effects. For instance, the SOM results may be used to better understand if CMs are capable of reproducing similar patterns between large-scale circulation and cloud microphysical/radiative effects. One could then quantify the impact of using the radiative effect from the observed SOM relationship with the modeled 850
Moreover, most CMs have difficulty with accurately representing MBL clouds – which are susceptible to aerosol effects – because they often use a horizontal grid spacing that is too large (
While the results presented here are promising, a data set spanning a longer time period is required to develop a robust analysis that evaluates the ability of CMs to reproduce the observed synoptic-scale weather patterns and mesoscale cloud properties. In general, using machine learning techniques to connect large-scale circulation patterns to cloud microphysics, which is challenging using solely observations from field campaigns or modeling case studies, is important for accurate predictions of future atmospheric climate. The results presented here may not be applicable to all marine stratiform cloud decks owing to potential differences in the frequency, strength, and location of the respective high-pressure circulation, as well as differences in, for example, coastal geometry and topography, continental land use, aerosol sources, and sea surface temperature. Future work will explore the application of the methodology outlined herein to the other dominant MBL cloud regions of the world using global reanalysis products and model output.
NARR reanalysis is available from the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Information (NCEI) website (
TWJ designed the study, developed the code, performed the analysis, and wrote the manuscript. ZJL made substantial contributions to the analysis and revised the manuscript.
The authors declare that they have no conflict of interest.
Timothy W. Juliano is grateful for support from the state of Wyoming, the Carlton R. Barkhurst Fellowship, and NCAR through the National Science Foundation. We would also like to acknowledge high-performance computing support from Cheyenne (
This research has been supported by the U.S. Department of Energy, Office of Science (grant no. DE-SC0016354).
This paper was edited by Philip Stier and reviewed by Johannes Mülmenstädt and one anonymous referee.