Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Non-linear cyclic regimes of short-term climate variability Sih, Perry 2003

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2003-0535.pdf [ 2.45MB ]
Metadata
JSON: 831-1.0103715.json
JSON-LD: 831-1.0103715-ld.json
RDF/XML (Pretty): 831-1.0103715-rdf.xml
RDF/JSON: 831-1.0103715-rdf.json
Turtle: 831-1.0103715-turtle.txt
N-Triples: 831-1.0103715-rdf-ntriples.txt
Original Record: 831-1.0103715-source.json
Full Text
831-1.0103715-fulltext.txt
Citation
831-1.0103715.ris

Full Text

NON-LINEAR CYCLIC REGIMES OF SHORT-TERM CLIMATE VARIABILITY by PERRY SIH B.Sc, The University of British Columbia, 2001 A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES (Department of Physics and Astronomy) We accept this thesis as conforming to the re^uireci-$tandard THE UNIVERSITY OF BRITISH COLUMBIA October 2003 ©Perry Sih, 2003 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department The University of British Columbia Vancouver, Canada Date DE-6 (2788) Abstract The Circular Non-linear Pr incipal Component Analysis ( C N L P C A ) , a variation of the non-linear version of the traditional Pr inc ipa l Component Analysis ( P C A ) , is introduced. It is then applied to monthly-averaged geopo-tential heights of the N A S A Goddard Institute for Space Studies SI2000 Global Circulat ion Model ( G C M ) . It is shown that height variabil i ty in the model troposphere and stratosphere is essentially linear, even wi th different aerosol forcings. The daily-averaged model output show weak non-linearity. When C N L P C A is applied to observed geopotential height data, cyclic be-haviour appears. The preferred states of the climate system can be seen. This cyclic behaviour can be tracked by recording the phase angle, a unique feature of the C N L P C A . B y doing so, the preferred direction, as well as the frequency, of the cyclic behaviour can be found. i i Contents Abstract ii Contents iii List of Figures iv 1 Introduction 1 2 The Circular Non-linear Principal Component Analysis 3 2.1 Principal Component Analysis 3 2.2 Non-linear PCA 5 2.3 Circular NLPCA 10 2.4 Non-linearity of a Neural Network 11 2.5 Search for Suitable Parameters 16 3 Monthly-averaged Model Data 20 4 Daily-averaged Model Data 30 4.1 500mb Geopotential Height 31 4.2 50mb Geopotential Height 38 4.3 Ensemble Averages 44 5 Observed Data 53 5.1 Regime Behaviour 53 5.2 Time Series of Phase Angle 59 6 Conclusions 61 Bibliography 63 iii List of Figures 2.1 The five-layer feed-forward neural network used in NLPCA . . 6 2.2 The five-layer feed-forward neural network used in CNLPCA . 12 2.3 CNLPCA with m = 3 and p = 0 13 2.4 CNLPCA with m = 2 and p = 0.6 14 2.5 CNLPCA mode 1 for 500mb heights Trial A, m = 2, unfiltered 17 2.6 CNLPCA mode 1 for 500mb heights Trial A, m = 3, unfiltered 18 2.7 CNLPCA mode 1 for 500mb heights Trial A, m = 2, filtered . 19 3.1 First two PCs of SI2000/B399 model at 300mb 22 3.2 Scores of the first two PCs of SI2000/B399 model at 300mb . . 22 3.3 First two PCs of SI2000/B399 model at 30mb 23 3.4 Scores of the first two PCs of SI2000/B399 model at 30mb . . 23 3.5 First two PCs of SI2000/B424o model at 300mb 24 3.6 Scores of the first two PCs of SI2000/B424o model at 300mb . 24 3.7 First two PCs of SI2000/B424o model at 30mb 25 3.8 Scores of the first two PCs of SI2000/B424o model at 30mb . . 25 3.9 First two PCs of SI2000/B424 model at 300mb 26 3.10 First two PCs of SI2000/B424 model at 30mb 26 3.11 Scores of first three PCs of SI2000/B424 model at 300mb . . . 27 3.12 CNLPCA results of SI2000/B424 model at 300mb 27 3.13 Scores of first two PCs of SI2000/B424 model at 30mb . . . . 29 3.14 CNLPCA results of SI2000/B424 model at 30mb 29 4.1 First two PCs for the B424 Trial A at 500mb 32 4.2 Scores for first three PCs for the Trial A at 500mb 33 4.3 CNLPCA results, in PC space, for trial A at 500mb 33 4.4 Angular distribution of the results at 500mb 35 4.5 Spatial patterns at each regime at 500mb 36 4.6 Spatial patterns at max and min PCI and PC2 at 500mb . . . 37 iv 4.7 First two PCs for the Trial A at 50mb 39 4.8 Scores for first three PCs for the Trial A at 50mb 39 4.9 CNLPCA results, in PC space, for Trial A at 50mb 40 4.10 Angular distribution of the results at 50mb 41 4.11 Spatial patterns at each regime at 50mb 42 4.12 Spatial patterns at max and min PCI and PC2 at 50mb . . . 43 4.13 First two PCs of the ensemble average at 500mb 45 4.14 Scores of the first three PCs of the ensemble average at 500mb 45 4.15 CNLPCA results, in PC space, for the ensemble average at 500mb 46 4.16 Distribution of the results of the ensemble average at 500mb . 47 4.17 First two PCs of the ensemble average at 50mb 48 4.18 Scores of the first three PCs of the ensemble average at 50mb 48 4.19 CNLPCA results, in PC space, for the ensemble average at 50mb 49 4.20 Angular distribution of the results of the ensemble average at 50mb 51 4.21 Spatial patterns at each regime at 50mb 52 5.1 First two PCs for observed data at 500mb 55 5.2 Scores for first three PCs for observed data at 500mb 56 5.3 CNLPCA results, in PC space, for observed data at 500mb . . 56 5.4 Angular distribution of the results of the observed data at 500mb 57 5.5 The observed regimes 58 5.6 Time series of the phase angle of observed geopotential height data 60 Chapter 1 Introduction Our atmosphere is a non-linear system. Although we can learn a lot from modelling it as a linear system, it is important to investigate its non-linear properties in order to gain more understanding. In 1991, Kramer [1] intro-duced the non-linear version of the Principal Component Analysis (NLPCA) using a five-layer feed-forward neural network. It is successful in modelling non-cyclical non-linear datasets, such as the Lorentz attractor [2]. However, since it is only capable of producing open, U-shaped curve solutions, the NLPCA is not good at modelling cyclical signals within a dataset. In 1996, Kirby and Miranda [3] modified the NLPCA by replacing the single neuron at the bottleneck layer by the circular neurons. By doing so, closed curve (loops) solutions, as well as open curve solutions, can be obtained. When there exists a closed curve solution, the system is considered cyclical. When applying the Circular NLPCA to the geopotential height fields, we will come across the Pacific North America (PNA) teleconnection pattern, the North Atlantic Oscillation (NAO), as well as the Arctic Oscillation (AO). 1 The PNA pattern is one of the most important modes of low-frequency cli-mate variability, especially during the Northern Hemisphere winter [4]. The NAO is a north to south dipole oscillation of geopotential height anomalies across the North Atlantic Ocean [5]. The AO is a circulation pattern in which the atmospheric pressure over the polar regions varies in opposition with that over middle latitudes on time scales ranging from weeks to decades [6]. The NAO and AO are similar patterns [7], but the AO covers more of the Arctic Ocean. These three main spatial patterns of atmospheric variability will combine in a non-linear system and regime behaviour can be observed. Regime behaviour in our atmosphere has been observed and documented ([8], [9], [10], and [11]). These regimes represent quasi-stationary atmospheric states. Transitions between regimes are fast when compared to the time spent within regimes. The Circular NLPCA is a convenient way to follow the evolution of the atmosphere, and has potential to be used in forecasting. 2 Chapter 2 The Circular Non-linear Pr incipal Component Analysis It is best to break down the introduction to CNLPCA into several steps. Since the CNLPCA uses principal components as input, PCA will be dis-cussed briefly. The NLPCA will be introduced next, and then we will modify it with a circular node to get the CNLPCA. 2.1 Principal Component Analysis For a given dataset that can be expressed as a A;-dimensional spatial vector that evolves in time, x(t„) = (x1(tn),x2{tn), ...,xk(tn)) (2-1) where tn is the nth observation time. The Principal Component Analysis looks for a ui (£„), which is a linear combination of the components of x ( t n ) : 3 Ul{tn) =x(tn) -ex (2.2) so that the approximation to x x = uiei (2.3) is such that the mean square error between x and x J = ( | | x - x | | 2 ) (2.4) is minimized. From the residuals, e2 can be found using the same method. In general, *(*n) = £ [ * ( ' " ) • e i t e ( 2 - 5 ) i= i or if we write it in another way, i x (*n) = X^ X(*«) ' e j ] e J + €n (2-6) where en are the residuals. We can see that this is a type feature extraction problem, in which x(tn) = f(f(x(tn))) + en (2.7) where f is a projection function and f is an expansion function. We can see, by comparison, that the projection and expansion functions in the PCA 4 are both linear. This means that the coordinate axes that come from the orthogonal eigenvectors, e,, are straight lines. Therefore, the P C A is opt imal only i f the feature to be extracted can be characterized by such a set of axes. For example, data clouds that have the shapes of an ellipsoid or a cylinder can be described well b y P C A , but data clouds that have the shapes of rings or bows can hardly be characterized by a linear model. Non-linear P C A can provide a better characterization. 2.2 Non-linear P C A The N L P C A allows the projection function f and the expansion function f to be non-linear. The solution to the feature extraction problem (equation (2.6) subject to (2.4)) can be implemented using a five-layer feed-forward neural network [1]. A feed-forward neural network is made up of a number of parallel layers of processing units, which are called neurons. The output from each neuron in the ith layer w i l l be used as the input in the (i + l ) t h layer. Let y^ be the output of the j t h neuron of the ith layer, then a feed-forward network can be summarized as the following: E<1M° + >irfl) (2.8) where w^1^ is the weight function or synaptic strength and b^+1^ is the bias. <j( l+1) is the transfer function that characterizes the (i + l ) t h layer and it can 5 be linear or non-linear. Cybenko [12] discovered that it is possible to approx-imate to arbitrary accuracy any continuous function from k dimensions to I dimensions by a three-layer neural network with k input neurons, hyperbolic transfer functions in the second layer, and linear transfer functions in the third layer with / neurons. At the third layer, the solution comes from com-pressing the original data to a one-dimensional time series, therefore x can be considered as the optimal one-dimensional approximation to x, embedded in fc-dimensional space. In order to visualize the embedded feature, a second network can be used to map the extracted feature from I dimensions back to k dimensions. In general, we can use one network in place of f to go from k to I dimensions, and another one in place of f to go from I to k dimensions. We shall use the example in figure 2.1 to demonstrate the details. Bottleneck Output LTr ^ layer Encoding Decoding layer layer Figure 2.1: The five-layer feed-forward neural network used in NLPCA. This figure can be found in [13]. In this particular model, the neurons in the three middle layers are called 6 hidden neurons, so called as they are not physically measurable quantities. There are two hidden neurons in the second and fourth layers, and thus we define the number of hidden neurons, m, to be 2 for this example. The first layer is the input layer. In our example of figure 2.1, there are three input neurons. These three neurons, yf \ receive data (a three-component vector time series x(in) of principal components, for instance) presented to the network, therefore its transfer function, is just the iden-tity function. The second layer is the encoding layer with two hidden neurons. The fcth neuron in this layer will have the following as its input: y<? = tanh[„iM> + n & C U + + (2-9) The third layer is the bottleneck layer and it consists of one single neuron. This means that the output of this neuron will be a one-dimensional time series. The transfer function is the identity function, and therefore, for this neuron u, y[V=u = wW-yW+bW (2.10) We can impose a normalization condition such that we get unit variance, (u2) = 1, by modifying the cost function: J = ( l |x -x | | 2 ) + ( ( n 2 ) - l ) 2 (2.11) 7 The fourth layer is the decoding layer. If we treat the bottleneck layer as the input neuron, then the transfer function wi l l be the hyperbolic tangent: y i 4 ) = tanh(44W44)) (2.12) The fifth and final layer is the output layer. The transfer function is the identity function, and thus: y^ = ^M4) + ^ + ^ (2.13) The final output is then: x = _ _ y f )e i (2.14) The cost function (equation (2.11)) is minimized by finding the optimal values of all of the weight and bias functions. This process is called "training the network." One thing of notice is that since the system is non-linear wi th a large number of degrees of freedom (2lm + Am + 1 + 1 without normalization, where / is the number of input neurons, and m is the number of neurons in each of the encoding and decoding layers), there exist many local min ima in the space of J (equation (2.11)). Usually the minimizat ion of J w i l l end up in one of the local min ima instead of the global minimum because the minimizat ion algorithm is designed to move in the direction of decreasing value of the cost function. Therefore, one must inspect the minimizat ion results from an 8 ensemble of feed-forward neural networks with different in i t ia l weights and choose from it the case where the mean square error is the smallest. Also, the encoding layer and the decoding layer are not required to have the same number of neurons. But generally it is fixed in order to reduce the number of free parameters in the model architecture. It is possible to obtain'better results if the input variables are appropri-ately scaled. This is because if the values of the variables vary by several orders of magnitudes, some of the weight functions wi l l tend to be negligible when minimiz ing J. A good way to scale each variable is to first remove its mean and then divide by its standard deviation. If the leading P C s are used, then there is no need to scale the variables because the P C s are already scaled. If the data cloud is concentrated but of irregular shape, there wi l l be a greater chance of obtaining an overfitted solution. In order to avoid that, a certain fraction (for example, 20%) of the original data can be randomly cho-sen and excluded from the minimizat ion. After the minimizat ion is complete, the mean square error of the two portions can be compared. A n overfitted solution is one that has a mean-square error (MSE) from the training por-t ion lower than the MSE from the randomly selected test samples. Overfitted solutions can be rejected when determining the best solution. If the underlying feature to be extracted resides in a two-dimensional space instead of a one-dimensional space, a feed-forward neural network wi th two bottleneck neurons can be used [2]. In general, the number of bottleneck 9 neurons corresponds to the number of dimensions in which the feature being sought resides. 2.3 Circular N L P C A A cyclic underlying feature containing three or more regimes is best char-acterized by a closed curve solution to the minimizat ion problem. Th is is because each regime need to be connected to two regimes in order for the feature to be considered cyclic. In an open curve solution, the regimes at either end of the curve are connected to only one other regime. The system is forced to move towards one regime, and therefore such a feature is not con-sidered cyclic. The N L P C A is not capable of obtaining closed curve solutions because the bottleneck neuron, u, is not an angular variable. It is possible to make the bottleneck variable an angular variable. K i r by and M i randa [3] introduced the circular neurons in the bottleneck layer. It is accomplished by using two bottleneck neurons, p and q, while adding the restriction that they are normalized: p2 + q2 = l ( 2 . 1 5 ) This condit ion constrains the network to have only one degree of freedom wi th two neurons. To bui ld the C N L P C A network, we create the following two unnormalized 10 neurons: p0 = Wrjy{2) + 43Jy^ + b^ (2.16) a0 = w%yV+w$yP + bW (2.17) Using the normalization restriction (equation 2.15), we replace u in equation (2.10) by the following normalized bottleneck neurons: Po 2^ \JPI + Qo q = - ^ L = (2-19) \JPI + % Since the bottleneck layer has changed, the decoding layer (equation (2.12)) must change accordingly: yP = tanh(w$p + w<$q + bP) (2.20) The rest (equations (2.9) and (2.13)) remains unchanged. See figure 2.2 for the schematics of the CNLPCA. 2.4 Non-linearity of a Neural Network There are two ways to adjust the non-linearity of a neural network. One is to introduce a penalty factor, and the other is to adjust the number of hidden neurons. The non-linearity of CNLPCA, as well as NLPCA, comes from the non-linear transfer function, the hyperbolic tangent. It has the property that 11 Figure 2.2: The five-layer feed-forward neural network used in CNLPCA. This figure can be found in [13]. for a large weight, w, tanh (wx) approaches the step function. On the other hand, if w is small, tanh (wx) WX, which is approximately linear. So, if the optimization arrives at a minimum where w is large, we may obtain a highly non-linear result. This can cause overfitting, in which noise is mis-taken as signal. Therefore, if we can penalize large w and force w to be sufficiently small, then the non-linearity can be decreased. This can be done by modifying the cost funtion: J=(||x-x | | 2 )+pEK 2 ) ) 2 (2-21) ik Generally, to choose the appropriate penalty, we can choose a set of values between 0.1 and 1.0 and obtain a solution for each of them, and then choose the solution that yields the smallest MSE to be the solution of our choice. 12 0.1 0.05 0 -0.05 -0.1 -0.15'— -0.2 0.05 x-0.05 -0.1 -0.15 -0.1 (a) o x i (c) 0.1 0.2 r< ^£rfp -0.05 x-0.05 -0.1 -0.15 -0.2 (b) -0.1 0 x1 (d) 0.1 0.2 -0.15 -0.1 -0.05 0.05 0.1 Figure 2.3: CNLPCA with m = 3, and a penalty factor of 0. The data are taken from the subsurface temperature of the tropical Pacific Ocean at a depth of 120m. Six PCs were used in the input and output layers, but only the first three are shown, (a) mode 2 vs mode 1; (b) mode 3 vs mode 1; (c) mode 3 vs mode 2; (d) a 3D view of the three modes. The original data are represented by the dots. The linear PCA solution is represented by the line. The CNLPCA solution is represented by the overlapping circles. The vertex at the top of (a) is a symptom of overfitting. 13 (a) (b) 0.1 0.05 0 -0.05 -0.1 -0.15 -0.2 ylj; Jj]Si,ft'\k> -• 0.05 X -0 .05 -0.1 -0.15 -0.1 0 X1 (C) 0.1 0.2 • .'. . .' . •• 0.05 -0.05 -0.1 -0.15 -0.2 -0.1 0 x1 (d) 0.1 0.2 -0.15 -0.1 -0.05 0 0.05 0.1 Figure 2.4: C N L P C A with m = 2 and a penalty factor of 0.6. Notice that wi th one fewer hidden neuron and a penalty factor, the solution is smoother than the last case. Th is is a better solution since it is not overfitted. 14 Notice that the MSE is always small when the penalty is close to zero because these are the overfitted solutions we are trying to avoid. Also notice that with a large penalty factor, the weights will go to zero in order to minimize the cost function. The neural network will thus become linear. The number of hidden neurons also affects the non-linearity of the system greatly. In a neural network with I input neurons and m hidden nuerons, adding one extra hidden neuron will introduce 21 + 2 extra variables to the minimization problem. Typically, more hidden neurons yield smaller mean square errors. But this is usually considered a case of overfitting. Usually, m = 2 or m = 3 will be sufficient. An example can be seen in figures 2.3 and 2.4. A small data set containing monthly averaged subsurface sea temperature at 120m below sea surface from the Scripps Institution of Oceanography is analyzed using CNLPCA. In figure 2.3, three hidden neurons in each of the encoding and decoding layers and no penalty factor were used, thus allowing high degree of non-linearity. The result is a closed curve that overfit the data. The vertex at the top of figure 2.3a) is a symptom of overfitting. It is because of noise that moves the curve away from the centre such that it needs to take a sharp turn to fit the next point. The CNLPCA approximation using only two hidden neurons can be seen in figure 2.4. The resulting curve goes through a ring-shaped path defined by a dense upper branch and a scarce lower branch, in which the data points on either side of the curve are about the same distance away from the curve, figure 2.4 is a better fit because the curve is smoother and 15 it takes the shape and the density of the data cloud. The short solid lines offer a comparison between the C N L P C A and the linear P C A . The linear solution simply goes through where the density of the data cloud is the highest, treating all the outliers as noise, whereas the C N L P C A takes some of the outliers into account. But, by how much should we take these outliers into account? 2.5 Search for Suitable Parameters Since we are trying to fit a loop in multi-dimensional space, there are many local minima in the cost function (equation (2.21)) in which our solutions may be trapped. Therefore, it is important to find suitable parameters that would lead to the best possible solution. Aside from the penalty factor and a suitable number of hidden neurons, a suitable filter is also useful in finding the best solution. We shall use the daily-averaged output of 500mb geopotential heights from the N A S A Goddard Institute for Space Studies Global Climate Model (details in Chapter 4) as an example. In figure 2.5, no filter is used. As a result, the solution with the lowest mean square error is a linear one. However, if the number of hidden neurons is increased to three, as in figure 2.6, to allow for more non-linearity due to the lack of filtering, then the cyclical behaviour can be retrieved. However, a larger number of hidden neurons increase the risk of overfitting. Therefore, a filter is still recommended. This is illustrated by figure 2.7, in which two hidden neurons are used together with filtering. Since some high-16 o CL (b) -5000 P C , (c) 5000 4000 2000 2 0 -2000 -4000 -4000 -2000 0 2000 4000 P C 2 o 0. 4000 2000 -2000 -4000 -5000 P C . 5000 Penalty = 0.1, M S E = 1.8888 o 0. -4000 Figure 2.5: CNLPCA mode 1 for 500mb heights Trial A with no filtering. Two hidden neurons are used, and the result yields a linear solution even with a small penalty. frequency noise is removed here using a 10-day low-pass filter, the averaging effect that would otherwise lead to a linear solution is also removed. A loop solution exists. In general, non-linear solutions have lower mean square error. Although it is almost impossible to reach the global minimum during the minimization, a loop solution can bring insights into cyclical behaviour, as shown in the Chapter 5 using observed data. 17 (a) o Q. 4000 2000 0 -2000 -4000 -5000 P C 1 (c) 5000 O 0. 4000 2000 -2000 -4000 -4000 -2000 0 PC„ 2000 4000 O D -4000 2000 -2000 -4000 (b) -5000 P C . 5000 Penalty = 1, M S E = 1.6309 o 0. Figure 2.6: CNLPCA mode 1 for 500mb heights Trial A with no filtering. Three hidden neurons are used, and the best results come with penalty of 1.0. A larger penalty is needed to compensate for the non-linearity introduced by adding a third hidden neuron. Since the mean square error is lower than the previous case, figure 2.5, this is a superior result. In general, cyclical results yield lower MSE. 18 Figure 2.7: CNLPCA mode 1 for 500mb heights Trial A. The heights were filtered using a 10-day low-pass filter before being processed through the CNLPCA neural network. Two hidden neurons are used, but this time a loop solution is obtained. Removing noise reduces the averaging effect, and hence increase the observable non-linearity. 19 Chapter 3 Monthly-averaged Mode l Data The CNLPCA is used to investigate if there exists any cyclical behaviour in a set of monthly-averaged model data. The model chosen is the Global Cir-culation Model version SI2000 from the NASA Goddard Institute for Space Studies. In particular, three sets of data with different atmospheric forcings were selected. The B399 experiment includes no time:varying radiative forcings besides a climatological annual cycle that repeats itself. Boundary forcing comes from the observed sea surface temperature and sea ice, based on the Hadley data set HadlSST 1 [14]. The B424 experiment takes into account six types of time-varying radiative forcings. These forcings are caused by variations in the concentrations of greenhouse gases, ozone, solar heating, stratospheric water vapour, stratospheric and tropospheric aerosols, and by variability in the solar heating. The B424o includes the same radiative forcings as B424, but its ocean representation has a specific "q-flux" horizontal heat transport, which adds diffuse mixing with the deep ocean for transient experiments. 20 All three data sets have a horizontal resolution of 4' x 5', and a vertical resolution of twelve layers (nine in the troposphere and three in the strato-sphere). Each data set uses the same set of data processing techniques and param-eters. Height value at each grid point is first multiplied by the squareroot of the cosine of its latitude in order to compensate for the increase in spa-tial density of the points with increasing latitude. The climatological annual cycle is calculated by averaging each month of the year, which is then sub-tracted from each grid point. Three-month winters are then extracted out of each year to obtain the winter anomalies. For the B399 model at 300mb (first two PCs in figure 3.1 and their score in figure 3.2), the first PC is similar to the Arctic Oscillation (AO). The second PC has a ring-like positive structure surrounding an intense negative peak in North America. This is associated with the PNA pattern. At 30mb (first two PCs in figure 3.3 and their score in figure 3.4), the monopole oscillation is the dominant mode and the dipole oscillation is seen in the second PC. The first PC of the B424o model at 300mb (first two PCs in figure 3.5 and their score in figure 3.6) shows a large positive peak around the Arctic Circle. This is associated with the AO. The second PC has three negative peaks, located on the Pacific and Atlantic Oceans, as well as the Northern Eurasian continent, surrounding a positive peak in the arctic portion of North America. This can be regarded as a cold ocean, warm land pattern. The PCs at 30mb (first two PCs in figure 3.7 and their score in figure 3.8) are 21 PC#1 [Exp=37.9%] PC #2 [Exp=8.9%] Figure 3.1: First two PCs of SI2000/B399 model at 300mb 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 3.2: Scores of the first two PCs of SI2000/B399 model at 300mb 22 PC#1 [Exp=62.5%] PC #2 [Exp=11.2%] Figure 3.3: First two PCs of SI2000/B399 model at 30mb 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 3.4: Scores of the first two PCs of SI2000/B399 model at 30mb 23 PC#1 [Exp=31.0%] PC #2 [Exp=11.2%] 1950 1955 1960 1965 1970' 1975 1980 1985 1990 1995 2000 Time Figure 3.6: Scores of the first two PCs of SI2000/B424.O model at 300mb similar to that in the B399 model. The first two PCs of the B424 model at 300mb (first two PCs in figure 3.9 and their score in figure 3.11) are similar to those of the B399 model. The PCs at 30mb (first two PCs in figure 3.10 and their score in figure 3.13) again show monopole and dipole oscillations, which are also similar to those of the B399 model. One thing of notice is that the percentage of variance accounted for by the first PC in the stratosphere in all the models are around 60%. This shows that the model stratosphere may be oversimplified. 24 PC#1 [Exp=59.0%] PC #2 [Exp=11.1%] Figure 3.7: First two PCs of SI2000/B424o model at 30mb 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 3.8: Scores of the first two PCs of SI2000/B424o model at 30mb 25 PC#1 [Exp=39.0%] PC #2 [Exp=8.4%] Figure 3.9: First two P C s of SI2000/B424 model at 300mb PC#1 [Exp=63.8%] PC #2 [Exp=10.9%] Figure 3.10: First two P C s of SI2000/B424 model at 30mb 26 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 3.11: Scores of the first three PCs of SI2000/B424 model at 300mb (a) (b) o CL O Q . 1000 500 0 -500 -1000 -1500 1000 500 0 -500 -1000 b 4, -2000 -1000 0 1000 2000 P C 1 (c) • 4BE"»-o Q . 1000 500 -500 -1000 -2000 -1000 0 1000 2000 P C 1 (d) 1000 o Q . -1500 -1000 -500 0 500 1000 P C 2 2000 Figure 3.12: CNLPCA results of SI2000/B424 model at 300mb, three hidden neurons, penalty 1.2. The broken lines represent the linear PCA solutions. Slight non-linearity and no cyclic behaviour can be seen. 27 The scores of the first eight PCA modes (I = 8) are used as inputs to the CNLPCA neural network. Three hidden neurons (m = 3) are used in the hidden layers. A series of runs are conducted with different penalty factors, and the optimal ones are found for each model. We find that even though the PCs look different in some cases, the CNLPCA results are similar. Therefore, we shall use the B424 run to illustrate the results. The scores of the first three PCA modes for the B424 runs can be seen in figures 3.11 and 3.13. The respective results can be see in figures 3.12 and 3.14. It can be seen that the CNLPCA solutions are very close to the PCA solutions. The PC2 magnitude in the 300mb run is only 10% of the PCI magintude, and 20% in the 30mb run. Therefore we conclude that these models are essentially linear. Also, since the CNLPCA solutions do not yield closed loops, we conclude that these monthly-averaged models do not show cyclic behaviour. If monthly-averaged model output data are essentially linear, then per-haps there is some non-linear behaviour when we shorten the averaging time interval [15]. So, next we look at the daily-averaged models. 28 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 3.13: Scores of the first three P C s of SI2000/B424 model at 30mb Figure 3.14: C N L P C A results of SI2000/B424 model at 30mb, three hidden neurons, penalty 1.1. The broken lines represent the linear P C A solutions. There is slightly more non-linearity than at 300mb, but there is s t i l l no cyclic behaviour. 29 Chapter 4 Daily-averaged Model Data The SI2000/B424 model is used to obtain daily-averaged geopotential height data at 500mb and 50mb. An ensemble of five runs is produced. Each model integration is started from slightly different initial conditions. Initially, each run is analyzed separately. Similar results are obtained for each run at the same pressure level, therefore the Trials A for each level are used in illustration. The data processing and parameters used in individual data sets are the same. For each grid point, height value is first multiplied by the square-root of the cosine of its latitude in order to compensate for the increase in spatial density of the points with increasing latitude. Then, the clima-tological annual cycle is calculated by averaging similar days of successive years. Anomalies are created by subtracting the climatological annual cycle from each grid point. A ten-day low-pass filter is applied to the anomalies to remove high-frequency noise, such as the synoptic-scale storms. Finally, three-month winters (December, January and February) are extracted out of 30 each year. The principal components are found for the processed data using the Principal Component Analysis (PCA). The first eight PCs become the input to the C N L P C A neural network. Three hidden neurons are used in each of the hidden layers to allow sufficient non-linearity. A variety of penalty factors are used before the one that yields the lowest mean square error is determined the appropriate result. 4.1 500mb Geopotential Height Figure 4.1 shows the first two principal components at 500mb. The first P C is similar to the Arctic Oscillation, where there is a maximum near the polar area. The pattern of the anomaly shows a different sign in the polar area and the area around it, with respect to the Pacific Ocean. The second P C is similar to the North Atlantic Oscillation, characterized by the strong maximum in the Atlantic Ocean. The positive anomaly extends around that latitude, enclosing a minimum centered in the northern part of Canada. Figure 4.2 shows the time series of the first three principal components. The first eight principal components, accounting for 68.8% of the total vari-ance, form the input to the C N L P C A neural network, and the result can be seen in figure 4.3. The dots that represent the input data follow a path. This is because a ten-day low-pass filter is applied in order to remove high-frequency noise, hence the more obvious trajectory. The loop solution sug-gests cyclical behaviour. It is interesting to investigate the time spent in 31 PC#1 [Exp=22.3%] PC #2 [Exp=10.7%] Figure 4.1: First two PCs for the B424 Trial A at 500mb between 20°N and 90°N. "Exp" is the percentage of variance accounted for by the corresponding P C . each area of the loop. A histogram of the angular distribution can be seen in figure 4.4. The angle is calculated by: 9 = tan _! PC2 PCI (4.1) The system spends the most time in the positive P C I direction. The negative P C I direction is the second most frequent. The system also spends a large proportion of time at the —85° and the 110° directions. These are close to the positive and negative PC2 directions, but there is a small P C I component mixed in them. The spatial patterns at each point within the central bin, as well as the two bins on each side, are averaged to obtain spatial patterns for each regime. Figure 4.5 shows the spatial patterns of the high frequency regimes mentioned 32 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 4.2: The scores of the first three PCs for the Trial A at 500mb. (a) Model 500a 1500 1000 500 -500 1000 -2000 -1000 0 1000 2000 P C 1 (c) 1500 1000 500 -1000 -500 0 500 1000 1500 PC 2 -2000 -1000 2000 Penalty = 1, M S E = 1.8945 2000 Figure 4.3: CNLPCA results for trial A at 500mb in PC space. Three hidden neurons are used. The dots are the input data into the CNLPCA neural network. The dashed line represents the linear PCA results. 33 above. The two patterns at 2° and —172° are similar to the PCI and -PCI patterns (the AO) respectively, as expected. Also, the two patterns at 110° and -85° are similar to the PC2 and -PC2 patterns (the NAO) respectively. Another interesting thing to investigate is the extreme cases, as seen in figure 4.6. Because of the tilt of the loop, the maxima and minima PCI and PC2 values of the CNLPC do not occur on the axes, but rather at an angle. Notice that the maximum and minimum for PCI occur close to the PCl-axis where the frequency is at the highest, but the maximum and the minimum for PC2 occur at a direction off the PC2-axis where the frequency is near the lowest. The PCs do not capture the directional variabilities. The greater the tilt, the more different the spatial patterns look. 34 PDF histogram (60 bins) 2001 1 ! 1 angle Figure 4.4: Angular distribution of the results at 500mb. The horizontal axis is the angle obtained from taking arctan(p^|) for each of the 4320 points on the CNLPCA output. Each bin is 6° in size, for 60 bins in total. 35 Average pattern with angles centered at -172 Average pattern with angles centered at -85 Average pattern with angles centered at 2 Average pattern with angles centered at 110 Figure 4.5: The average spatial patterns for each regime at 500mb, obtained from original processed data. The angles are determined by the peaks in the angular distribution, figure 4.4. 36 Max PC1 at 3.9 degrees Min PC1 at -179.2 degrees Figure 4.6: The spatial patterns at the points on figure 4.3 with (top left) Maximum PCI, (top right) Minimum PCI, (lower left) Maximum PC2, and (lower right) Minimum PC2. 37 4.2 50mb Geopotential Height Figure 4.7 shows the first two principal components of Trial A at the 50mb geopotential height. The structure is simpler than that at the 500mb geopo-tential height. The first PC is an oscillation similar to the tropospheric Arctic Oscillation. This is a symptom of an oversimplified representation of the stratosphere. The second PC is the dipole oscillation. Figure 4.8 shows the score of the first three PCs. The CNLPCA output of the scores (figure 4.9) reveals a loop that lies on the PC1-PC2 plane. This is similar to the tropospheric case, which supports the idea of an oversimplified stratosphere [16]. The penalty factor used to obtain an optimal solution is only 0.1, compared to 1.0 at the 500mb level. This shows that there is less non-linearity involved in the stratosphere of the model. The angular distribution, shown in figure 4.10, has a similar structure as that in the 500mb geopotential height. This shows that even though the spatial patterns are oversimplified, the model successfully captured the evolution of the score. The spatial patterns of the regimes, figure (4.11), show that the strato-sphere changes between oscillating as a monopole (top left and bottom left) and a dipole (top right and bottom right). The patterns at the extrema (figure 4.12), however, show more wavelike structures than the regimes. 38 PC#1 [Exp=33.6%] PC #2 [Exp=15.3%] Figure 4.7: First two PCs for the Trial A at 50mb between 20°N and 90°N. "Exp" is the percentage of variance. 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 Time Figure 4.8: The scores of the first three PCs for the Trial A at 50mb. 39 (a) Model 50a (b) Figure 4.9: CNLPCA results for Trial A at 50mb in PC space are the overlap-ping circles. The dots are the input data into the CNLPCA neural network. The dashed line represents the linear PCA results. 40 PDF histogram (60 bins) 200 I j ! ! angle Figure 4.10: Angular distribution of the results at 50mb. The horizontal axis is the angle obtained from taking arctan(fgf) for each point on the CNLPCA output. Each bin is 6° in size, for 60 bins in total. 41 Average pattern with angles centered at -150 Average pattern with angles centered at -70 Average pattern with angles centered at -2 Average pattern with angles centered at 160 Figure 4.11: The average spatial patterns around each regime at 50mb. The angles are determined by the well-separated peaks in the angular distribution, figure 4.10. 42 Max PC1 at -4.1 degrees Min PC1 at -172.1 degrees Max PC2 at 81.0 degrees Min PC2 at -90.4 degrees Figure 4.12: The spatial patterns at the points on figure 4.9 with (top left) Maximum PCI, (top right) Minimum PCI, (lower left) Maximum PC2, and (lower right) Minimum PC2. 43 4.3 Ensemble Averages When taking an ensemble average to average out the chaotic dynamics and amplify the forced signal, some non-linearity is lost. Therefore, some param-eters need to be adjusted for this change. When analyzing the five-member ensemble average, two hidden neurons are used in each of the encoding and decoding layers, and the first ten PCs are used. No low-pass filter is used be-cause taking an average acts like a filter, but the data are taken in seven-day segments to remove noise with synoptic timescales. Figure 4.13 shows the first two PCs of the five-member ensemble average at 500mb geopotential height. The first PC has the same features as the first PC in Trial A, seen in figure 4.1. The second PC is the PNA teleconnection pattern. The scores of the first three PCs are shown in figure 4.14. Figure 4.15 shows an open curve CNLPCA approximation as a result of averaging. The U-shaped curve suggests that although there is no cyclic behaviour, there is still non-linearity. However, an inspection of the point distribution along the curve and its cumulative distribution function, shown in figure 4.16, tells us that the system spends most of the time between 0.4 and 0.8 of the curve, normalized to unit length. This translates to the area close to the PCI axis. Thus, the system is mostly linear after taking an ensemble average. Figure 4.17 shows the first two PCs of the five-member ensemble average at 50mb geopotential height. The first PC is very similar to the first PC in 44 PC#1 [Exp=14.0%] PC #2 [Exp=8.7%] 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 T i m e Figure 4.14: Scores of the first three PCs of the ensemble average at 500mb. 45 (a) (b) Figure 4.15: CNLPCA results for ensemble average at 500mb in PC space. The dots are the input data into the CNLPCA neural network. The dashed line represents the linear PCA results. An open curve solution is obtained even with a small penalty of 0.05. 46 C D F 50001 1 1 1 1 1 r 4000 -x •g 3000 -1000 -0 -0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 distance P D F histogram 3001 1 1 1 1 i 1 1 1 1 1 250 h 200 -u) § 150 -o o 100 -50 h 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 distance Figure 4.16: Top: cumulative distribution function (CDF) of the results of the ensemble average at 500mb. Bottom: distribution of points along the curve, normalized to unit length. The curve starts at 0 on the left side, and ends at 1 on the right side. Each bin is 9° in size, for 40 bins in total. 47 figure 4.7. Also, the positive regions over the Pacific and Asia are connected in the ensemble. The second PCs shows a dipole oscillation similar to the second PC in figure 4.7. In the ensemble, the positive region extends accross the North Pole to cover all of the Arctic Ocean. Figure 4.19 shows a loop solution. A penalty factor of only 0.05 is needed because the ensemble averaging has reduced the level of non-linearity. A larger factor would yield non-cyclical results similar to the previous case at 48 (a) (b) -2000 -1000 0 1000 2000 -2000 -1000 0 1000 2000 Penalty = 0.05, M S E = 2.0162 2000 Figure 4.19: CNLPCA results for ensemble average at 50mb in PC space. The dots are the input data into the CNLPCA neural network. The dashed line represents the linear PCA results. 500mb. The angular distribution of the loop in figure 4.19 is shown in figure 4.20. There are three distinct peaks, one at around —175°, one at around —20° and one at around 80°. The system seldom stays in the region between —50° and -150° . Figure 4.21 shows similar spatial patterns to the first two PCs. However, there are some differences. For example, at the —20° direction, the negative 49 is very weak. In fact, most of that pattern is positive. Also, in the —75° and 80° directions, the Greenland region remain in the negative, instead of oscillating between positive and negative in a linear fashion. These non-linear oscillations are what the PCA cannot capture. 50 PDF histogram (60 bins) 200 I ! ; ! angle Figure 4.20: Angular distribution of the results of the ensemble average 50mb. Each bin is 6° in size, for 60 bins in total. 51 Average pattern with angles centered at -175 Average pattern with angles centered at - 75 Average pattern with angles centered at -20 Average pattern with angles centered at 80 Figure 4.21: The average spatial patterns around each regime of the ensemble average at 50mb. The angles —175°, 20° and 80° are determined by the peaks in the angular distribution, figure 4.20. The trough of the distribution at —75° is also shown. 52 Chapter 5 Observed Data When loop solutions are obtained, it is interesting to know which direc-tion, clockwise or counterclockwise in two-dimensional projection, the sys-tem moves. If a trend is found, it may provide useful insight on our climate system. In this chapter, the National Centers for Environmental Prediction-National Center for Atmospheric Reserach (NCEP-NCAR) reanalysis [17] for the 500mb geopotential height will be used for illustration. The data, which have a 2.5° x 2.5° horizontal resolution, are pre-processed the same way as the monthly-averaged and daily-averaged model data. A 10-day low-pass filter is applied. The first ten principal components are used as the input to a CNLPCA neural network (figure 2.2) with two hidden neurons (m = 2). 5.1 Regime Behaviour Figure 5.1 shows the first two PCs of the observed data. PCI has a positive peak in the Pacific coast of North America and a strong negative peak in the 53 Pacific Ocean. This pattern resembles the positive phase of the Pacific/North America oscillation (PNA) pattern. PC2 has a strong negative peak over Greenland, and the negative anomaly spreads into the Pacific Ocean. There is also a strong positive peak over the Atlantic Ocean. This meridional dipole in the North Atlantic is identified as the NAO. Figure 5.2 shows the score of the first three PCs, which are plotted in dashed lines in figure 5 .3 . A penalty factor of 0.1 is found to give a mean square error smaller than those given by nearby penalty factors. A loop solution is obtained in the PC1-PC2 plane. The probability density function around the loop is plotted in figure 5.4 and several well-separated peaks can be seen. The spatial patterns of these frequently visited regimes are shown in figure 5 .5 . The most frequently visited regime is located near the negative PCI axis. Its spatial pattern, top left in figure 5 .5 is the negative phase of the PNA. This is consistent with the regime A identified in Monahan et al. [10] using NLPCA. Cheng and Wallace [18], as well as Corti et al. [11], also found this pattern using cluster analysis (cluster B in figure 3 of [11]). The spatial pattern of a weak peak near the negative PC2 axis is shown in the top right of figure 5 .5 . It is associated with the negative phase of the NAO, and it is named the "G" pattern after the peak over Greenland. This is consistent with the NLPCA results, as well as that of the cluster analysis (negative of cluster A). This peak is chosen over the apparent peak at —126° because the bin next to the latter is one of the shortest. 5 4 PC#1 [Exp=14.8%] PC #2 [Exp=11.5%] Figure 5.1: First two PCs for observed data at 500mb between 20°N and 90°N. The next peak occurs at —32°. Its spatial pattern (bottom left of figure 5.5 is positive in the high lattitudes and mostly negative in lower lattitudes. This resembles the Arctic Oscillation (AO), which is not captured by the three-regime NLPCA, but is consistent with cluster analysis (cluster D). The last peak is located at 50°. Its spatial pattern (bottom right of figure 5.5 contains three positive peaks and three negative peaks, positioned in a wavelike fashion. Comparing with the top left pattern, this pattern resembles the PNA pattern. It is captured in the NLPCA (regime "R"), as well as the cluster analysis (negative of cluster C). 55 Figure 5.2: The scores of the first three PCs for observed data at 500mb. (a) (b) -2000 -1000 0 1000 2000 -2000 -1000 0 1000 2000 (c) Penalty = 0.1, M S E = 3.5909 Figure 5.3: CNLPCA results for observed data at 500mb in PC space are the overlapping circles. The dots are the input data into the CNLPCA neural network. The dashed lines represent the linear PCA results. 56 PDF histogram (40 bins) angle Figure 5.4: Angular distribution of the results at 500mb. The horizontal axis is the angle obtained from taking arctan(^|) for each of the 3600 points of the CNLPCA output. Each bin is 9° in size, for 40 bins in total. 57 Average pattern with angles centered at -32 Average pattern with angles centered at 50 Figure 5.5: The regime behaviour is determined from the most frequently visited states, seen in figure 5.4. 58 5.2 Time Series of Phase Angle When the arctangent of each point on the CNLPCA solution is taken, the PCI and PC2 scores can be combined into one single time series, shown in figure 5.6. This graph shows the evolution of the system. The long-term behaviour of the system was to move counterclockwise before the mid-1970s, and then clockwise between 1976 and 1983. In 1983-84, it suddenly turned around and moved counterclockwise rapidly, and this happened again in 1987-88. The system turned around again in the mid-1990s, before moving counterclockwise rapidly once again in 1998. Some of these changes in directions coincide with important happenings in our climate system. In the mid-1970s, there was a sudden climate change, resulting in the less frequent visits of regime "G" [10]. A lot of studies involving observed data tend to separate datasets into sub-periods before and after the mid-1970s because of this climate change [19]. The three rapid counterclockwise motions coincide with three of the most intense El Nino Southern Oscillation (ENSO) events ever recorded. On the other hand, La Nina events tend to make the system move in a clockwise direction, although the intense event in 1989 only move the system clockwise by one revolution, compared to the many revolutions during the El Nino events. There was an intense La Nina event in 1999, but the available observed data end just before that. It is expected that the time series will turn around and move in the clockwise direction. 59 Time Series of CNLPCA Phase Angle of Observed Geopotential Height Data 1955 1960 1965 1970 1975 1980 Year 1985 1990 1995 2000 Figure 5.6: Time series of the phase angle of observed geopotential height data. The time series of the phase angle contains several abrupt changes. It changes directions in mid-1970s. There are also three rapid counterclockwise rotations in 1983-84, 1987-88 and 1996-98. 60 Chapter 6 Conclusions The Circular Non-linear Principal Component Analysis is a versatile non-linear data analysis tool. It can be used to determine the level of non-linearity in a given set of data and extract non-linear features. However, the most important value of the CNLPCA is its ability to extract features that contain cyclical behaviour that involve three or more regimes. Further investigations include: • full-year analysis on the effect of discontinuous data between winters, • recent data (1999-current) to investigate the La Nina event in 1999, • usage of the time series of phase angles, for example, in forecasting, • consistency after subdividing a dataset into decadal timescales, and • comparisons with the results obtained from other data analysis methods. The CNLPCA does contain shortcomings. For example, the results are subjective in the sense that the mean square error tolerance level (that con-stitutes overfitting) depends on the researcher. However, this is a universal 61 problem in the field of non-linear data analysis. Also, the cost function mini-mization is very time-consuming when compared to the linear PGA. But with increasing computing power and the need to understand non-linear features, computing time should be reduced greatly in the future. 62 Bibliography [1] M.A. Kramer. Nonlinear principal component analysis using autoasso-ciative neural networks. AIChE Journal, 37:233-243, 1991. [2] A. Monahan. Nonlinear principal component analysis by neural networks: Theory and application to the lorenz system. Journal of Climate, 13:821-835, 2000. [3] M. Kirby and R. Miranda. Circular nodes in neural networks. Neural Computation, 8:391-402, 1996. [4] J.M. Wallace and D.S. Gutzler. Teleconnections in the geopotential height field during the Northern hemisphere Winter. Mon. Wea. Rev., 109:784-812, 1981. [5] J.W. Hurrell and H. van Loon Decadal Variations associated with the North Atlantic Oscillation. Climatic Change, 36:301-326, 1997. [6] D.W. Thompson and J.M. Wallace. The Arctic Oscillation signature in the wintertime geopotential height and temperature fields. Geophys. Res. Lett., 25:1297-1300, 1998. 63 [7] J.M. Wallace. North Atlantic Oscillation/Annular Mode: Two paradigms - One phenomoenon. Quart. J. Roy. Meteor. Soc, 126:791-806, 2000. [8] M. Kimoto and M. Ghil. Multiple flow regimes in the Northern Hemi-sphere winter. Part I: Methodology and hemispheric regimes. J. Atmos. Sci., 50:2625-2643, 1993. [9] P. Smyth, K. Ide, M. Ghil. Multiple regimes in Northern Hemisphere height fields via mixture model clustering. J. Atmos. Sci., 56:3704-3723, 1999. [10] A. Monahan, L. Pandolfo and J. Fyfe. The preferred structure of vari-ability of the Northern Hemisphere atmospheric circulation. Geophysical Research Letters, 28(6):1019-1022, March 2001. [11] S. Corti, F. Molteni and T.N. Palmer. Signature of recent climate change in frequencies of natural atmospheric circulation regimes. Na-ture, 398:799-802, 1999. [12] G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals, and Systems, 2:303-314, 1989 [13] W.W. Hsieh. Nonlinear principal component analysis by neural net-works. Neural Networks, 13:1095-1105, 2000.. [14] N. Rayner. HadlSSTl Sea Ice and Sea Surface Temperature Data Files. Geophysical Research Letters, Hadley Center, Bracknell, U.K., 2000. 64 [15] Yuval, and W.W. Hsieh The impact of time-averaging on the detectabil-ity of nonlinear empirical relations. Quart. J. Roy. Met. Soc, 128: 1609-1622., 2002. [16] A. Monahan, J. Fyfe, L. Pandolfo. The Vertical Structure of Wintertime Climate Regimes of the Northern Hemisphere Extratropical Atmosphere. Journal of Climate, 16(12):2005-2021, 2003. [17] E. Kalnay and Coauthors. The NCEP/NCAR 40-year Reanalysis Project. Bull. Amer. Meteor. Soc, 77:437-471. [18] X. Cheng and J.M. Wallace. Cluster analysis of the Northern Hemi-sphere wintertime 500-hpa height field: Spatial patterns. J. Atmos. Sci., 50:2674-2696, 1993. [19] J.L. Knox, K. Higuchi, A. Shabbar, E. Neil. Secular Variation of Northern Hemisphere 50 kPa Geopotential Height. Journal of Climate, 1(5):500-511, 1988. 65 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0103715/manifest

Comment

Related Items