Silvia Scarpetta

Automatic Classification of Seismic Signals at Mt. Vesuvius Volcano, Italy, Using Neural Networks

We present a new strategy for reliable automatic classification of local seismic signals and volcano-tectonic earthquakes (vt). The method is based on a supervised neural network in which a new approach for feature extraction from short period seismic signals is applied. To reduce the number of records required for the analysis we set up a specialized neural classifier, able to distinguish two classes of signals, for each of the selected stations. The neural network architecture is a multilayer perceptron (mlp) with a single hidden layer. Spectral features of the signals and the parameterized attributes of their waveform have been used as input for this network. Feature extraction is done by using both the linear predictor coding technique for computing the spectrograms, and a function of the amplitude for characterizing waveforms. Compared to strategies that use only spectral signatures, the inclusion of properly normalized amplitude features improves the performance of the classifiers, and allows the network to better generalize. To train the mlp network we compared the performance of the quasi-Newton algorithm with the scaled conjugate gradient method. We found that the scaled conjugate gradient approach is the faster of the two, with quite equally good performance. Our method was tested on a dataset recorded by four selected stations of the Mt. Vesuvius monitoring network, for the discrimination of low magnitude vt events and transient signals caused by either artificial (quarry blasts, underwater explosions) and natural (thunder) sources. In this test application we obtained 100% correct classification for one of the possible pairs of signal types (vt versus quarry blasts). Because this method was developed independently of this particular discrimination task, it can be applied to a broad range of other applications.

Approximation of continuous and discontinuous mappings by a growing neural RBF-based algorithm

In this paper a neural network for approximating continuous and discontinuous mappings is described. The activation functions of the hidden nodes are the Radial Basis Functions (RBF) whose variances are learnt by means of an evolutionary optimization strategy. A new incremental learning strategy is used in order to improve the net performances. The learning strategy is able to save computational time because of the selective growing of the net structure and the capability of the learning algorithm to keep the effects of the activation functions local. Further, it does not require high order derivatives. An analysis of the learning capabilities and a comparison of the net performances with other approaches reported in literature have been performed. It is shown that the resulting network improves the approximation results reported for continuous mappings and for those exhibiting a finite number of discontinuities.

Discrimination of Earthquakes and Underwater Explosions Using Neural Networks

We report on the implementation of an automatic system able to discriminate between explosion-generated artificial seismic events and local earthquakes in the Phlegraean Fields (Italy). The explosions are fired weekly at the sea bottom (tens of meters below sea level) by fishermen in Pozzuoli bay; earthquakes are volcano-tectonic quakes with depths shallower than 4 km. The discrimination system is based on an artificial neural network and is composed of two modules. The first is devoted to the extraction of the seismogram signatures and the second to the classification of the seismic events into two classes. For the features extraction (preprocessing stage), instead of the conventional Fourier spectral analysis, we use a Linear Prediction Coding (LPC) algorithm. This approach compresses the data from 256 samples to only 7 parameters and can extract robust features for the spectrogram representation. The classification is performed using a supervised neural algorithm based on a Multilayer Neural Network (MLP) architecture. We applied the method to a set of 30 seismic events recorded by the stations of the local seismic network, 15 of which were generated by the fishermens explosions and 15 were volcano-tectonic earthquakes. We dealt with a total of 280 records from different stations, 121 relating to explosions and 159 to earthquakes. Data were divided in a training set containing 120 traces for earthquakes and 90 for explosions, and a test set containing 70 traces corresponding to 39 records for earthquakes and 31 records for explosions. On the test set the neural net gave a classification performance of 92%, indicating a good ability of the net to generalize. Manuscript received 10 January 2002.

Unsupervised Neural Analysis of Very-Long-Period Events at Stromboli Volcano Using the Self-Organizing Maps

We have implemented a method based on an unsupervised neural network to cluster the waveforms of very-long-period (VLP) events associated with explosive activity at the Stromboli volcano (southern Italy). Stromboli has several active vents in the summit area producing together more than 200 explosions/day. We applied this method to investigate the relationship between each vent and its associated VLP explosive waveform. We selected 147 VLP events recorded between November and December 2005, when digital infrared camera recordings were available. From a visual inspection of the infrared camera images, we classified the VLPs on the basis of which vent produced each explosion. We then applied the self-organizing map (SOM), an unsupervised neural technique widely applied in data exploratory analysis, to cluster the VLPs on the basis of their waveform similarity. Our analysis demonstrates that the most recurrent VLP waveforms are usually generated by the same vent. Some exceptions occurred, however, in which different waveforms are associated with the same vent, as well as different vents generating similar waveforms. This suggests that the geometry of the upper conduit-vent system plays a role in shaping the recurring VLP events, whereas occasional modest changes in the source process dynamics produce the observed exceptions.

Automatic Discrimination among Landslide, Explosion-Quake, and Microtremor Seismic Signals at Stromboli Volcano Using Neural Networks

In this article we report on the implementation of an automatic system for discriminating landslide seismic signals on Stromboli island (southern Italy). This is a critical point for monitoring the evolution of this volcanic island, where at the end of 2002 a violent tsunami occurred, triggered by a big landslide. We have devised a supervised neural system to discriminate among landslide, explosion-quake, and volcanic microtremor signals. We first preprocess the data to obtain a compact representation of the seismic records. Both spectral features and amplitude-versus-time information have been extracted from the data to characterize the different types of events. As a second step, we have set up a supervised classification system, trained using a subset of data (the training set) and tested on another data set (the test set) not used during the training stage. The automatic system that we have realized is able to correctly classify 99% of the events in the test set for both explosion-quake/ landslide and explosion-quake/microtremor couples of classes, 96% for landslide/ microtremor discrimination, and 97% for three-class discrimination (landslides/ explosion-quakes/microtremor). Finally, to determine the intrinsic structure of the data and to test the efficiency of our parametrization strategy, we have analyzed the preprocessed data using an unsupervised neural method. We apply this method to the entire dataset composed of landslide, microtremor, and explosion-quake signals. The unsupervised method is able to distinguish three clusters corresponding to the three classes of signals classified by the analysts, demonstrating that the parametrization technique characterizes the different classes of data appropriately.

Uncertainty Analysis for the Classification of Multispectral Satellite Images Using SVMs and SOMs

Classification of multispectral remotely sensed data with textural features is investigated with a special focus on uncertainty analysis in the produced land-cover maps. Much effort has already been directed into the research of satisfactory accuracy-assessment techniques in image classification, but a common approach is not yet universally adopted. We look at the relationship between hard accuracy and the uncertainty on the produced answers, introducing two measures based on maximum probability and α quadratic entropy. Their impact differs depending on the type of classifier. In this paper, we deal with two different classification strategies, based on support vector machines (SVMs) and Kohonens self-organizing maps (SOMs), both suitably modified to give soft answers. Once the multiclass probability answer vector is available for each pixel in the image, we studied the behavior of the overall classification accuracy as a function of the uncertainty associated with each vector, given a hard-labeled test set. The experimental results show that the SVM with one-versus-one architecture and linear kernel clearly outperforms the other supervised approaches in terms of overall accuracy. On the other hand, our analysis reveals that the proposed SOM-based classifier, despite its unsupervised learning procedure, is able to provide soft answers which are the best candidates for a fusion with supervised results.

Hebbian imprinting and retrieval in oscillatory neural networks

We introduce a model of generalized Hebbian learning and retrieval in oscillatory neural networks modeling cortical areas such as hippocampus and olfactory cortex. Recent experiments have shown that synaptic plasticity depends on spike timing, especially on synapses from excitatory pyramidal cells, in hippocampus, and in sensory and cerebellar cortex. Here we study how such plasticity can be used to form memories and input representations when the neural dynamics are oscillatory, as is common in the brain (particularly in the hippocampus and olfactory cortex). Learning is assumed to occur in a phase of neural plasticity, in which the network is clamped to external teaching signals. By suitable manipulation of the nonlinearity of the neurons or the oscillation frequencies during learning, the model can be made, in a retrieval phase, either to categorize new inputs or to map them, in a continuous fashion, onto the space spanned by the imprinted patterns. We identify the first of these possibilities with the function of olfactory cortex and the second with the observed response characteristics of place cells in hippocampus. We investigate both kinds of networks analytically and by computer simulations, and we link the models with experimental findings, exploring, in particular, how the spike timing dependence of the synaptic plasticity constrains the computational function of the network and vice versa.

Neural Avalanches at the Critical Point between Replay and Non-Replay of Spatiotemporal Patterns

We model spontaneous cortical activity with a network of coupled spiking units, in which multiple spatio-temporal patterns are stored as dynamical attractors. We introduce an order parameter, which measures the overlap (similarity) between the activity of the network and the stored patterns. We find that, depending on the excitability of the network, different working regimes are possible. For high excitability, the dynamical attractors are stable, and a collective activity that replays one of the stored patterns emerges spontaneously, while for low excitability, no replay is induced. Between these two regimes, there is a critical region in which the dynamical attractors are unstable, and intermittent short replays are induced by noise. At the critical spiking threshold, the order parameter goes from zero to one, and its fluctuations are maximized, as expected for a phase transition (and as observed in recent experimental results in the brain). Notably, in this critical region, the avalanche size and duration distributions follow power laws. Critical exponents are consistent with a scaling relationship observed recently in neural avalanches measurements. In conclusion, our simple model suggests that avalanche power laws in cortical spontaneous activity may be the effect of a network at the critical point between the replay and non-replay of spatio-temporal patterns.

Mathematical analysis and simulations of the neural circuit for locomotion in lampreys.

We analyze the dynamics of the neural circuit of the lamprey central pattern generator. This analysis provides insight into how neural interactions form oscillators and enable spontaneous oscillations in a network of damped oscillators, which were not apparent in previous simulations or abstract phase oscillator models. We also show how the different behavior regimes (characterized by phase and amplitude relationships between oscillators) of forward or backward swimming, and turning, can be controlled using the neural connection strengths and external inputs.

On-line learning in RBF neural networks: a stochastic approach

The on-line learning of Radial Basis Function neural networks (RBFNs) is analyzed. Our approach makes use of a master equation that describes the dynamics of the weight space probability density. An approximate solution of the master equation is obtained in the limit of a small learning rate. In this limit, the on line learning dynamics is analyzed and it is shown that, since fluctuations are small, dynamics can be well described in terms of evolution of the mean. This allows us to analyze the learning process of RBFNs in which the number of hidden nodes K is larger than the typically small number of input nodes N. The work represents a complementary analysis of on-line RBFNs, with respect to the previous works (Phys. Rev. E 56 (1997a) 907; Neur. Comput. 9 (1997) 1601), in which RBFNs with N >> K have been analyzed. The generalization error equation and the equations of motion of the weights are derived for generic RBF architectures, and numerically integrated in specific cases. Analytical results are then confirmed by numerical simulations. Unlike the case of large N > K we find that the dynamics in the case N < K is not affected by the problems of symmetric phases and subsequent symmetry breaking.