1. A comparison of the ability of artificial neural network and polynomial fitting was carried out in order to model the horizontal deformation field. It is performed by means of the horizontal components of the GPS solutions in the Cascadia Subduction Zone. One set of the data is used to calculate the unknown parameters of the model and the other is used only for testing the accuracy of the model. The problem of overfitting (i.e., the substantial oscillation of the model between the training points) can be avoided by restricting the flexibility of the neural model. This can be done using an independent data set, namely the validation data, which has not been used to determine the parameters of the model. The proposed method is the so-called “stopped search method”, which can be used for obtaining a smooth and precise fitting model. However, when fitting high order polynomial, it is hard to overcome the negative effect of the overfitting problem. The computations are performed with Mathematica software, and the results are given in a symbolic form which can be used in the analysis of crustal deformation, e.g. strain analysis. Crustal velocity field modelling with neural network and polynomials Piroska Zaletnyik 1, Khosro Moghtased-Azar 2 1 Department of Geodesy and Surveying, Budapest University of Technology and Economics – Hungary, piri@agt.bme.hu 2 Institut of Geodesy, University of Stuttgart – Germany, moghtased@gis.uni-stuttgart.de The adaptation of neural networks to the modeling of the deformation field offers geodesists a suitable tool for describing structural deformation. Overfitting problem can occur in higher order polynomials, but neural network overcomes the problem thanks to stopped search method. The greatest advantage of this method is that the solution can be given as an analytical function, which could be use to compute derivation of the velocity vectors for strain analysis. The first author wishes to thank to the Hungarian Eötvös Fellowship for supporting her visit at the Department of Geodesy and Geoinformatics of the University of Stuttgart (Germany), where this work has been accomplished. Abstract Results Fig 1.GPS determined horizontal velocity field by Pacific Northwest Geodetic Array (PANGA), which is plotted relative to North American Plate. Introduction Polynomial fitting When there are more points than the number of parameters, there is a possibility for adjustment calculation, i.e. for polynomial fitting. In this case 62 points were used for the adjustment. The calculations were carried out with Mathematica software. Figure 4 : Northing velocity model by polynomials Figure 5 : Northing velocity model by neural networks Conclusion Acknowledgement Neural network with stopped search method Testing set residuals (mm/year)minmaxmeanStd. Northing velocities-11.136.31.49.0 Easting velocities-9.317.31,86.1 Teaching set residuals (mm/year)minmaxmeanStd. Northing velocities-4.13.90.01.2 Easting velocities-7.98.20.02.8 Table1 :The statistics of the differences between polynomial model and real velocities in the 62 teaching points. Testing set residuals (mm/year)minmaxmeanStd. Northing velocities-6.28.3-0.52.8 Easting velocities-4.68.90.23.8 Teaching set residuals (mm/year)minmaxmeanStd. Northing velocities-4.95.50.01.6 Easting velocities-6.89.50.23.1 The GPS measurements to determine crustal strain rates were initiated in the Cascadia region (US Pacific Northwest and south-western British Columbia, Canada) more than a decade ago, with the first campaign measurements in 1986 and the establishment of permanent stations in 1991. Nowadays, continuous GPS data from the Pacific Northwest Geodetic Array processed by the geodesy laboratory serves as the data analysis facility for the Pacific Northwest Geodetic Array (PANGA). This organization has deployed an extensive network of continuous GPS sites that measure crustal deformation along the CSZ. Fig. 1 illustrates the horizontal velocity field along the Cascadia margin assuming the North American plate to be stable. Table 2. The statistics of the differences between 3D polynomial model and real velocities in the 20 testing points. The overfitting problem means that the error of the teaching set is decreasing while the error of the testing set is growing, in other words the network excessively fits the teaching points which is illustrated by Fig. 3. Overfitting problem Fig. 2. Overfitting problem Comparing the results of Table 1 and Table 2 we recognize a significant difference between the deviations of the teaching and the testing set. The determined model by polynomial, works well only in the teaching points but between them it does not work as well. The testing set, which was not used during the determination of the model, is also needed in order to qualify the results. As a classical approximation model, 3D polynomial fitting technique is used to build continuous velocity field as a function of geodetic coordinates. Displacement vector which can be derived from GPS observations have east, north and up components in topocentric coordinates. For modeling the horizontal displacement field we use only the north and the east elements. Accuracy of modeling is determined by differences between true values and values estimated by 3D polynomial fitting. When we increasee the degree of the polynomial, accuracy is increaseing up to the 6 th degree, but above that started to decrease, because of the deterioration of the conditions of the equations (ill conditioned equations). The 6 th order polynomial was the best fitting model. In this case 28 points are needed, because a two-variable 6 th order polynomial has 28 parameters. A central issue in choosing the most suitable model for a given problem is selecting the right structural complexity. Clearly, a model that contains too few parameters will not be flexible enough to approximate important features in the data. If the model contains too many parameters, it will approximate not only the data but also the noise in the data. Overfitting may be avoided by restricting the flexibility of the neural model in some way. The Neural Networks package in Mathematica offers a few ways to handle the overfitting problem. All solutions rely on the use of a second, independent data set, the so-called validation data, which has not been used to train the model. One way to handle this problem is the stopped search method. Stopped search refers to obtaining the network’s parameters at some intermediate iteration during the training process and not at the final iteration as it is normally done. During the training the values of the parameters are changing to reach the minimum of the mean square error (MSE). Using validation data, it is possible to identify an intermediate iteration where the parameter values yield a minimum MSE. At the end of the training process the parameter values at this minimum are the ones used in the delivered network model. In order to avoid the overfitting problem by means of stopped search method, we will need more data. A learning set and a validation set. Hence, we have to divide the used teaching set (62 points) into two sets, the first will be the learning set with 42 points and the remaining 20 points will be the validation set. In fig. 3 we can see the errors of the learning and the teaching set during the learning procedure of the neural network model for the northing velocities. The errors of the learning set (continuous line) decrease during the whole procedure, but the errors of the validation set (dashed line) are decreasing only until the 262 nd iteration step, from that point are growing. The maximum number of iteration was 500, but the best parameter set is the one calculated at the 262 nd iteration step. In the model in the hidden layer 7 neurons (nodes) were used. The selection of number of neurons is basically depends on the number of known points. In fact, by having more known data we can increase the numbers of neurons. Let us check the statistics of the residuals for the whole teaching set (62 points) in Table 3. Fig.3. Errors of learning (continuous line) and validation set (dashed line) during stopped search method Table 3. The statistics of the differences between neural network model and real velocities in the 62 teaching points. Table 4. The statistics of the differences between neural network model and real velocities in the 20 testing points. Let’s see the differences between the neural network model and the real velocities in the 20 testing points (Table 4.) Using neural network model with stopped search technique we can obtain a smooth and good fitting model, while in the case of high order polynomial model there are substantial oscillations between the teaching points. See fig. 4 and 5.