Self-organizing Maps Patterns Research Articles

Intraseasonal and Interdecadal Jet Shifts in the Northern Hemisphere: The Role of Warm Pool Tropical Convection and Sea Ice

Abstract This study uses cluster analysis to investigate the interdecadal poleward shift of the subtropical and eddy-driven jets and its relationship to intraseasonal teleconnections. For this purpose, self-organizing map (SOM) analysis is applied to the ECMWF Interim Re-Analysis (ERA-Interim) zonal-mean zonal wind. The resulting SOM patterns have time scales of 4.8–5.7 days and undergo notable interdecadal trends in their frequency of occurrence. The sum of these trends closely resembles the observed interdecadal trend of the subtropical and eddy-driven jets, indicating that much of the interdecadal climate forcing is manifested through changes in the frequency of intraseasonal teleconnection patterns. Two classes of jet cluster patterns are identified. The first class of SOM pattern is preceded by anomalies in convection over the warm pool followed by changes in the poleward wave activity flux. The second class of patterns is preceded by sea ice and stratospheric polar vortex anomalies; when the Arctic sea ice area is reduced, the subsequent planetary wave anomalies destructively interfere with the climatological stationary waves. This is followed by a decrease in the vertical wave activity flux and a strengthening of the stratospheric polar vortex. An increase in sea ice area leads to the opposite chain of events. Analysis suggests that the positive trend in the Arctic Oscillation (AO) up until the early 1990s might be attributed to increased warm pool tropical convection, while the subsequent reversal in its trend may be due to the influence of tropical convection being overshadowed by the accelerated loss of Arctic sea ice.

Uncovering spatial and temporal patterns of Adriatic Sea colour with self-organizing maps

Fifteen years (1998–2012) of monthly chlorophyll-like pigment concentrations from satellite images (JRC database) were analysed using self-organizing maps (SOMs). The aim was to classify the pigment structures of the Adriatic Sea into a number of patterns present in the data by this novel neural network method. The images are represented by a vector in the multidimensional data-space limited to the Adriatic area. We tried using several predefined patterns, finally selecting an optimal number of 16 spatial structures. Distinct neurons are primarily associated with patterns of different pigment concentration over the whole Adriatic area and with a different extension of the colour front along the west coast. Rare phenomena and variability limited to smaller areas were less represented in the SOM patterns. Although temporal pigment dynamics are synchronous in many areas of the Adriatic Sea, the analysis of their annual course resulted in ten temporal patterns. This representation also demonstrates the ability of the SOM method to connect areas with a particular seasonal variation in chlorophyll.

Spatial Changes of Central Field Loss in Diabetic Retinopathy After Laser

To explore the spatial distribution of central visual field loss in untreated proliferative diabetic retinopathy (PDR) and to quantify the effect of medium-pulse Optos-guided 20-millisecond Pascal laser treatment on the central field. Visual field data (Swedish Interactive Threshold Algorithm 24-2) from 99 eyes (66 patients) with treatment-naive PDR were used to train a self-organizing map (SOM) that classified the defects into nine patterns. Twenty-eight eyes of 23 patients treated with 20-millisecond Pascal retinal laser photocoagulation underwent Optos widefield fundus fluorescein angiography (WF-FFA) at baseline and 3 months after treatment. Postlaser changes in SOM patterns and global indices were analyzed. Visual field defect changes (Total Deviation [TD]) with eccentricity and extent of initial loss were analyzed. Grading of WF-FFA after laser was undertaken by two masked retina specialists. At baseline, 44.4% of PDR eyes showed early visual field loss patterns (1 to 3), with 23.2% classified into the advanced patterns (7 to 9). Mild SOM patterns had more superior hemifield field defects, whereas advanced patterns involved both superior and inferior hemifield field loss. After laser, a significant shift to early SOM patterns were observed (p = 0.02), as well as improvement of Mean Deviation and Pattern Standard Deviation (p = 0.003 and p = 0.06, respectively). Improvement of TD was commonly observed in test locations of 0 to 10, 10 to 20, and 20 to 30 degrees. Greater improvement was observed with deeper baseline TD (p < 0.001). Masked WF-FFA image grading showed 78.6% PDR regression. The SOM method is a promising technique to classify and monitor over time PDR-associated visual field defects. Medium-pulse Optos-guided 20-millisecond Pascal laser treatment improved the spatial patterns and global parameters of central field defects.

Surface current patterns in the northern Adriatic extracted from high-frequency radar data using self-organizing map analysis

[1] A network of high-frequency (HF) radars was installed in the northern Adriatic in the second half of 2007, aimed to measure surface currents in the framework of the North Adriatic Surface Current Mapping (NASCUM) project. This study includes a detailed analysis of current measurements from February to August 2008, a period in which three radars were simultaneously operational. Current patterns and temporal evolutions of different physical processes were extracted by using self-organizing map (SOM) analysis. The analysis focused on subtidal frequency band and extracted 12 different circulation patterns on a 4 × 3 rectangular SOM grid. The SOM was also applied on a joint data set that included contemporaneous surface wind data obtained from the operational hydrostatic mesoscale meteorological model ALADIN/HR. The strongest currents were recorded during energetic bora episodes, being recognized by several current patterns and having the characteristic downwind flow with magnitudes exceeding 35 cm/s at some grid points. Another characteristic wind, the sirocco, was represented by three current patterns, while the remaining current structures were attributed to weak winds and the residual thermohaline circulation. A strong resemblance has been found between SOM patterns extracted from HF radar data only and from combined HF radar and wind data sets, revealing the predominant wind influence to the surface circulation structures and their temporal changes in the northern Adriatic. These results show the SOM analysis being a valuable tool for extracting characteristic surface current patterns and forcing functions.

VARIABILITY OF SATELLITE-OBSERVED SEA SURFACE HEIGHT IN THE TROPICAL INDIAN OCEAN: COMPARISON OF EOF AND SOM ANALYSIS

Weekly sea surface height (SSH) in the tropical Indian Ocean (20°S - 20°N) was analyzed for the period of January 1993 - December 2007 using an empirical orthogonal function (EOF) and a self-organizing maps (SOM) analysis. The EOF analysis identifies four patterns and three of them are contained in the SOM patterns. The SOM, on the other hand, characterizes the sea level variability, which shows twenty-five patterns. The patterns with low (high) sea surface height anomaly (SSHA) in the southern (northern) hemisphere associated with the monsoonal winds dominate the variation in both two methods. The SOM is also able to separate typical patterns associated with the ENSO and or the Indian Ocean Dipole (IOD) events. Low SSHA occupied the western half of the basin while high SSHA loaded in the eastern basin when the La Nina event is taking place. The El Nino event is characterized by low SSHA in the northern hemisphere, along the equator and along the eastern boundary, while high SSHA in the southwestern part of the basin. The IOD event shows a dipole like pattern with low SSHA in the east and high SSHA in the west. When the IOD co-occurred with El Nino, a distinct dipole pattern is clearly observed.

Temporal variability of vertical nontidal circulation pattern in a partially mixed estuary: Comparison of self‐organizing map and empirical orthogonal functions

On the basis of a 233‐day moored acoustic Doppler current profiler data set in an estuarine strait, temporal variability of vertical estuarine circulation is examined using self‐organizing map (SOM) and empirical orthogonal functions (EOF). The SOM identifies a continuum of patterns of which eight have been selected. The best matching unit (BMU) time series shows distinct seasonal variation which separates the data set into three seasonal segments. Also, monthly variation can be discerned by BMU sequences, and the data set is further separated into seven monthly segments. The EOF analysis of the whole record identifies only four patterns which are contained in the SOM patterns. However, EOF analysis of the seven monthly segments shows distinct monthly variation of the first two EOF modes. Moreover, those EOF modes are compatible with the SOM patterns showing consistency between the two techniques. The synoptic variability is described by four selected events which correspond to particular BMU sequences and represent filling and draining processes. The pattern‐forcing relationships are explored with EOF principle components. Long‐term variations of circulation patterns are dominated by river discharge and winds. Short‐term variations explored within seasonal and monthly segments indicate that the flow patterns are results of combined effect of forcing. Officer's (1976) analytical solutions of estuarine circulation are applied, and the results support the flow patterns and pattern‐forcing relationship revealed by SOM and EOF analysis.

Self-similarity patterns of precipitation in the Iberian Peninsula

Summary An objective classification of the precipitation field over the Iberian Peninsula and the Balearic Islands is obtained. Data are derived from a high-resolution daily precipitation dataset obtained from in-situ measurements. The dataset, Iberian monthly Precipitation Dataset (IPD), consists of monthly precipitation data over a 25 km � 25 km grid from 1 st January 1961 to 31 st December 2003. Therefore, 960 data series over the Iberian Peninsula and the Balearic Islands are disposed over the grid for 43-year period. Multi-resolution wavelet analysis is used to extract similar information in the precipitation field at different timescales. An objective classification of the obtained wavelet coefficient series is carried out by means of the Kohonen’s neural network, also named Self-Organizing Map (SOM). SOM is formed by an unsupervised learning algorithm that may be used to find clusters of similar events in the input data and is able to identify some underlying dynamic structures of the multi-dimensional datasets. SOM is applied to the wavelet coefficients for intramonthly, intermonthly and interannual oscillations, obtaining self-organised maps which objectively identify similar zones of precipitation behaviour over the Iberian Peninsula. The homogeneity of the patterns is also studied by means of non-parametric correlations, energy scalograms and tests of significance. The intramonthly, intermonthly and interannual waves resulted in seven, five and three SOM patterns, respectively. As timescale increases, the wavelet series coefficients tend to be highly clustered. The results indicate that as the oscillation frequencies decrease, the Iberian precipitation behaves more linearly.