The Master SOM

The SOM algorithm identified the representative patterns of 500 hPa geopotential height anomalies displayed in Fig. 1. Patterns in diagonal corners tend to display nearly opposite features. The percentage of days belonging in each node is displayed above each plot. Central nodes are the least frequent while corner patterns tend to occur most often.

Figure 1

Matrix of representative patterns in 500-hPa geopotential height anomalies created using the self-organizing map algorithm. Daily data from the NCEP/NCAR reanalysis span 1948 to 2018, 30°N to 80°N, and 30°E to 180° longitude. Shading displays height anomalies in meters, and percentages over each pattern denote the occurrence of days in that node. Large bold numbers are for node reference.

Nodes in the upper left section of the SOM exhibit positive (negative) height anomalies in high (middle) latitudes, indicative of warm Arctic conditions accompanied by cool anomalies particularly in East Asia. Nodes in the lower right display negative height anomalies in high latitudes along with high heights over East Asia. Anomalous wind fields can be inferred from height patterns. In nodes #1 and #5, for example, anomalous northeast winds would be expected in central Asia. Cold Arctic winds from the north blow through central Asia in #9 and #10, while warm southerlies would be expected in #3 and #4.

We can also assess the seasonal occurrence of each pattern based on the days that reside in it. Figure 2 illustrates that patterns #1 and #12, which exhibit the most extreme opposite features, occur most often during cold months, while patterns with weak features (#2, #6, #7, and #11) are more likely in the warm season. This result is consistent with observations of strong (weak) baroclinicity and large airmass differences during winter (summer). Four of the patterns occur nearly equally in all months.

Figure 2

Monthly distribution of days residing in each node of the master SOM (Fig. 1). Units are months (x-axis) and days (y-axis), indicated on bottom left node.

Extreme temperatures and precipitation

Next we investigate the frequency and spatial distribution of extreme weather conditions associated with each large-scale pattern. Anomalies in air temperatures at 925 hPa (chosen to avoid small-scale surface effects) and total precipitation (relative to mean over entire record) are calculated from daily reanalysis output. The number of extreme days occurring at a gridpoint—i.e., days exhibiting anomalies above or below 1.5 σ—are mapped to the matrix and displayed in Fig. 3. Areas with a high frequency of extremely warm days (Fig. 3a) correspond fairly closely with positive height anomalies in the master SOM (Fig. 1), and also with regions where anomalous winds with a southerly component would be expected based on the location and orientation of height anomalies. A few interesting cases are evident, such as the large region of frequent warm extremes in node #12 across most of central and eastern Asia as well as the western Pacific, even though positive height anomalies are weak in those areas. Extremely high temperatures in these regions correspond with low heights in the Arctic. The same is true for #10, where extreme warmth is common across southeastern Asia where height anomalies are nearly neutral but low heights exist in the Siberian Arctic. Also noteworthy is that the most widespread warm extremes are associated with patterns that exhibit strong and meridionally oriented height anomalies, not the weak patterns that occur most often in summer months (Fig. 2).

Figure 3

Number of days that (a) 925 hPa air temperatures exceed 1.5 σ, (b) 925 hPa air temperatures fall below -1.5 σ, and (c) total precipitation exceeds 1.5 σ.

Similarly, cold temperature extremes (925 hPa air temperatures < − 1.5 σ) are mapped to the SOM in Fig. 3b. While patterns look similar to but opposite those in Fig. 3a, interesting exceptions are evident. A large region of frequent cold extremes spans much of Asia in node #1, a pattern characterized by an anomalously warm Arctic (positive height anomalies) and only weakly negative values over eastern Asia. Nodes #9 and #10 also exhibit regions of high heights in central and western Siberia, along with a broad band of cold extremes extending from southwestern Asia to northeastern Siberia. Interestingly, height anomalies in node #5 look similar but cold extremes are absent, likely because the height gradient is weaker and has a more zonal orientation.

A similar analysis is performed using extreme precipitation values, i.e., exceeding 1.5 σ (Fig. 3c). Widespread precipitation extremes are associated with nodes #1 and #12, which are most prevalent in the cold months (Fig. 2). In the “warm Arctic” pattern of node #1, days with heavy precipitation occur frequently over northwest and southeast Asia, as well as along the east coast. The “cold Arctic” node #12 pattern is associated with precipitation extremes across much of northern Asia, extending southward into western Asia. Nodes #3 and #4 feature anomalous flow from the south into northwest/northcentral Asia, producing heavy precipitation there, while nodes #9 and #10 tend to bring extreme precipitation into central and eastern regions associated with anomalous flow around the negative height anomaly in the northeast part of the continent. The nodes with maximum occurrence during summer months (#2, #6, #7, and #11) exhibit a small number of days with extremes in either temperatures and precipitation.

Societally important questions receiving a great deal of attention recently are how and why extreme weather conditions have changed as the globe has warmed, and what does the future hold? We can shed light on this issue through SOM analysis by assessing changes in the frequency of large-scale atmospheric patterns over time, and relating them to the extreme conditions associated with those patterns. Figure 4 presents time series of days/year that belong in each node of the SOM. Trends significant at 90% (95%) confidence are indicated with dashed (solid) bold lines, determined with the f-test. Nodes that exhibit positive height anomalies in high latitudes are generally occurring more frequently over time, especially since 1995 coincident with rapid Arctic warming. This is particularly evident in node #5 (> 95% confidence), with increases of fairly high confidence in node #1 (> 85%). Patterns with negative height anomalies in high latitudes, in contrast, are generally decreasing, especially in recent decades. This finding is consistent with observations of amplified Arctic warming2,22.

Figure 4

Change in the frequency of occurrence of each node from 1948 to 2018; units in days per year. Dashed (solid) lines indicate trends significant at the 90th (95th) confidence level from 1996 to 2018.

These trends in particular patterns help explain the recent increased heat extremes in the Arctic, western Asia, and eastern Asia (Fig. 3a, nodes #1, #3, and #4), as well as more frequent extreme cold days in central and eastern Asia (Fig. 3b, nodes #1 and #3) and decreased cold events in western areas (nodes #10 and #12). Changes in pattern frequency also contribute to increased days with heavy precipitation over western, southcentral, and eastern Asia (Fig. 3c, nodes #1, #3, and #4) along with decreased precipitation extremes over northern Asia (node #12).

Persistence assessed via frequency of long-duration events (LDEs)

As described above, we identify LDEs by tracking events in which four or more consecutive days occur in each node of the SOM. This threshold was determined subjectively by assessing the distribution of LDEs of varying lengths, as illustrated in Fig. 5. As expected, long-duration LDEs occur less frequently than short ones, with a dearth of events longer than three or more days duration. This threshold was selected to provide sufficient cases to detect statistically significant changes over time.

Figure 5

Distribution of LDEs of varying duration in each node. The horizontal axis is length of LDEs in days, and the vertical axis is total number of LDEs of each duration.

Time series of LDEs ≥ 4 days per year from 1948 to 2018 for each node are presented in Fig. 6. Large interannual variability is evident, and there are no significant trends over the 71-year period, although significant positive (negative) trends are evident in nodes #4 and #5 (#10 and #12) since the mid-1990s. It appears that none of these recent trends are consistently associated with changes in occurrence of short-duration (2 or 3 consecutive days) events (Fig. S1). We also find that the relative probability of an LDE occurring—especially longer LDEs—has increased significantly in recent decades for node #1 (Fig. 7; warm Arctic pattern), thus the change in frequency of LDEs results from the combined effects of the atmospheric pattern occurring more frequently (thus increasing the chance of multiple consecutive days) plus an increased probability of an LDE occurring per day.

Figure 6

Timeseries of the occurrence of LDEs (defined as 4 or more consecutive days in a node) per year based on NCEP/NCAR Reanalysis output. Bold dashed (solid) lines indicate trends from 1996-2018 significant at the 90th (95th) confidence level.

Figure 7

Change in the probability (%) of LDE occurrence (number of LDEs divided by number of days in a node) from 1962-1989 to 1991-2018 for LDEs lasting 7 or more days. Star indicates significance at 95% confidence.

Analysis of model simulations

Once the master SOM has been created, fields from other sources can be mapped onto the same master patterns. We take advantage of this analysis tool by mapping output from climate model simulations to the SOM matrix in Fig. 1. The algorithm places daily fields of 500 hPa height anomalies from each of three models into the node whose pattern is most similar to the modeled field. By using the master patterns identified in reanalysis output, any unrealistic patterns (or unrealistic dominance of any pattern) that may exist in model simulations are avoided.

We first compare historical model output with reanalysis results to assess the models’ ability to reproduce monthly distributions of node frequencies (Fig. S2). All three models do a remarkable job of capturing the frequency of patterns and their seasonal distribution. The frequencies and temporal changes in LDEs are also broadly similar (Fig. S3), with generally decreasing frequencies of patterns featuring negative height anomalies at high latitudes, though internal variability clearly dominates over this relatively short time period (1979–2005).

Having demonstrated that model simulations are able to capture realistic distributions of node occurrence and frequencies of LDEs in each node, we now turn to projections for 2006 to 2100 assuming future conditions described by the RCP 8.5 scenario. We further assess the realism of model output by creating two additional entirely independent SOM matrices using historical simulations (not shown) and future projections from the CCSM4 model (Fig. S4). Patterns in this matrix are nearly identical to those created using reanalysis fields (Fig. 1), and the monthly distributions are also very similar. These results provide further confidence in the model’s realism and suggest that the representative atmospheric patterns will not fundamentally change in the future nor shift in their monthly distributions.

Time series for the days/year in each node are presented in Fig. 8. Trends apparently emerging in the recent past (Fig. 4) are much more conspicuous in future projections. Generally, nodes featuring positive height anomalies in high latitudes (upper-left nodes in master SOM, Fig. 1) occur much more frequently in the most aggressive warming scenario of the CMIP5 experiment. From 2006 to 2100, the number of days residing in node #1 is projected to increase by about 150% in CCSM4 simulations, with even larger increases projected by CanESM2 (~ 600%) and GFDL-CM3 (~ 650%). This finding is consistent with expected intensification of disproportionate Arctic warming as greenhouse gases continue to accumulate in the atmosphere23. The AAW-like patterns increase at the expense of those with low height anomalies in the Arctic; in fact, these patterns become exceedingly rare by the end of the twenty-first century in all three models.

Figure 8

Model projections of node frequencies from 2006 to 2100 by the (a) CCSM4, (b) Can-ESM2, and (c) GFDL-CM3. Simulations are for the RCP 8.5 scenario. Dashed (solid) lines indicate significant trends with 90% (95%) confidence.

In terms of extreme weather, assuming the patterns of anomalous temperatures and precipitation associated with each node hold for the future, the projected trends in pattern frequency should have a substantial impact. Many more days residing in node #1, for example, should contribute to an increased frequency of warm extremes in western, northern, and eastern Asia; more extreme cold days in central and eastern Asia; and heavy precipitation in western and southeastern areas. Many fewer days in node #12, in contrast, will favor a decrease in warm extremes over southern Asia, fewer cold days in western and northern regions of the continent, and fewer precipitation extremes in northern and southwestern regions.

Trends in the frequency of LDEs generally follow the changes in node occurrence (Fig. 9). This is to be expected, as a larger number of days belonging in a node will increase the chances of events with multiple consecutive days. As in the analysis based on observations, we find that the probability-per-day of long LDEs generally increases in AAW patterns derived from future projections, as well (Fig. S5). The combination of these changes suggests that the future will bring more frequent occurrences of persistent weather patterns. It should be noted that no negative trends (in node frequency or LDEs) occur for patterns with positive high-latitude height anomalies, and no positive trends occur in nodes with a cold Arctic. Even in the more optimistic RCP scenarios, disproportionate Arctic warming is expected to continue, and these results suggest that it will be accompanied by more frequent persistent weather patterns as well as the specific types of extreme events across the Asian continent associated with those patterns.

Figure 9

Model projections of LDE frequencies per year from 2006 to 2100 by the (a) CCSM4, (b) Can-ESM2, and (c) GFDL-CM3. Simulations are for the RCP 8.5 scenario. Dashed (solid) lines indicate significant trends with 90% (95%) confidence.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *