A new automatic pattern recognition approach for the classification of volcanic tremor at Mt. Etna, Italy
21
0
0
Testo completo
(2) Outline. 1. Introduction 2. Data and Methods 3. Results 4. Discussion. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(3) Introduction Data and Methods Results Discussion. Mount Etna Volcanic Tremor Automatic Pattern Recognition Approach. Outline. 1. Introduction 2. Data and Methods 3. Results 4. Discussion. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(4) Introduction Data and Methods Results Discussion. Mount Etna Volcanic Tremor Automatic Pattern Recognition Approach. Mount Etna Mount Etna is the largest active volcano in Europe: Type: Basaltic stratovolcano Location: Sicily, Italy (3350 m a.s.l.) Latest eruptions: 2001, 2002, 2004. Mount Etna’s volcanic monitoring represents a key issue. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(5) Introduction Data and Methods Results Discussion. Mount Etna Volcanic Tremor Automatic Pattern Recognition Approach. Volcanic Tremor For basaltic volcanoes (e.g. Mount Etna)… . Volcanic tremor is a persistent seismic signal marking different states of the volcano’s activity:. Pre—eruptive . Lava fountain. Eruptive. Post—eruptive. Volcanic tremor provides reliable information for alerting governmental authorities during a crisis and permits surveillance even when direct access to the eruptive theatre is not possible Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(6) Introduction Data and Methods Results Discussion. Mount Etna Volcanic Tremor Automatic Pattern Recognition Approach. Automatic Pattern Recognition Approach How to develop an automatic classifier able to recognize different states of the volcano’s activity from the analysis of its volcanic tremor?. Data Collection & Labeling. Feature Extraction. Training of the Classifier. Test of the Classifier. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(7) Introduction Data and Methods Results Discussion. Data Features Classification. Outline. 1. Introduction 2. Data and Methods 3. Results 4. Discussion. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(8) Introduction Data and Methods Results Discussion. Data Features Classification. Data :: Collection The 17 July—08 August, 2001 eruption is considered…. 01 Jul 01. 16 Jul 01. 08 Aug 01. 15 Aug 01. Analysis is performed over 01 July—15 August, 2001 Seismograms are recorded at the 3—component station ESPD: 142 seismograms for the East—West (EW) component 142 seismograms for the North—South (NS) component 142 seismograms for the Vertical (Z) component Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(9) Introduction Data and Methods Results Discussion. Data Features Classification. Data :: Labeling Seismograms are labeled according to their recording date… Lava fountain (FON) seismograms. 01 Jul 01. 16 Jul 01. Pre—eruptive (PRE) seismograms By considering the three components of each seismogram as different patterns… Matteo Masotti, et al.. 08 Aug 01. Eruptive (ERU) seismograms. 15 Aug 01. Post—eruptive (POS) seismograms N. of patterns for each class. Full set. PRE. FON. ERU. POS. 154. 55. 180. 37. April 5, 2006 — EGU General Assembly.
(10) Introduction Data and Methods Results Discussion. Data Features Classification. Features Features are computed by… 1. Calculating the spectrogram of each seismogram (10 min., 0—15 Hz) 2. Averaging the rows of each spectrogram (62—dimensional feature vector). Pre—eruptive. Lava fountain. Matteo Masotti, et al.. Eruptive. Post—eruptive. April 5, 2006 — EGU General Assembly.
(11) Introduction Data and Methods Results Discussion. Data Features Classification. Classification For classification, a Support Vector Machine (SVM) classifier is chosen… Hyperplane w·x +b=0. 1. SVM finds the hyperplane w·x +b=0 maximizing the margin between the two classes in the training set 2. If feature vectors are not linearly separable, the problem is mapped into a higher feature space by means of a kernel function Φ(x) Matteo Masotti, et al.. Margin. Training Φ. Training. Training. April 5, 2006 — EGU General Assembly.
(12) Introduction Data and Methods Results Discussion. Cross—Validation Leave—One—Out. Outline. 1. Introduction 2. Data and Methods 3. Results 4. Discussion. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(13) Introduction Data and Methods Results Discussion. Cross—Validation Leave—One—Out. Cross—Validation :: Data Partitioning First, performances are studied using cross—validation + random subsampling… Total number of patterns. … Test pattern By considering the three components of each seismogram as different patterns, for one fold… Matteo Masotti, et al.. Fold1. Fold2. FoldN N. of patterns for each class PRE. FON. ERU. POS. Full set. 154. 55. 180. 37. Train set. 124. 44. 144. 30. Test set. 30. 11. 36. 7. April 5, 2006 — EGU General Assembly.
(14) Introduction Data and Methods Results Discussion. Cross—Validation Leave—One—Out. Cross—Validation :: Performances By repeating 100 times train and test… 1. Global average classification performances: (94.7 ± 2.4)% 2. Single—class average classification performances: (%) Predicted Class PRE. PRE. FON. ERU. POS. 94.2 ± 5.2. 5.4 ± 4.8. 0.4 ± 1.3. 0.0 ± 0.0. 76.4 ± 13.7. 3.4 ± 5.1. 0.0 ± 0.0. 0.3 ± 1.3. 99.6 ± 0.4. 0.1 ± 0.6. 0.0 ± 0.0. 0.0 ± 0.0. 100.0 ± 0.0. Actual FON 20.2 ± 12.6 Class ERU 0.0 ± 0.3 POS. 0.0 ± 0.0. 3. Similar results when EW, NS, and Z are taken into account separately Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(15) Introduction Data and Methods Results Discussion. Cross—Validation Leave—One—Out. Leave—One—Out :: Data Partitioning Second, performances are studied using leave—one—out… Total number of patterns. … Test pattern By considering the three components of each seismogram as different patterns, for one (example) fold… Matteo Masotti, et al.. Fold1. Fold2. FoldN N. of patterns for each class PRE. FON. ERU. POS. Full set. 154. 55. 180. 37. Train set. 153. 55. 180. 37. Test set. 1. 0. 0. 0. April 5, 2006 — EGU General Assembly.
(16) Introduction Data and Methods Results Discussion. Cross—Validation Leave—One—Out. Leave—One—Out :: Performances By repeating 142 times (on each single component) train and test… EW. NS. Z. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(17) Introduction Data and Methods Results Discussion. Misclassified Events Intra—class Variability. Outline. 1. Introduction 2. Data and Methods 3. Results 4. Discussion. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(18) Introduction Data and Methods Results Discussion. Misclassified Events Intra—class Variability. Misclassified Events Focusing on the Z component for brevity… (but analogous considerations can be drawn for EW and NS) . Misclassifications are mostly concentrated near class transitions. . Reasonably because of: 1. Intrinsic fuzziness in the transition from one volcanic state (i.e. class) to the other 2. Human imprecisions in labeling. Intra—class variability. Z. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(19) Introduction Data and Methods Results Discussion. Misclassified Events Intra—class Variability. Intra—Class Variability :: PRE and FON Focusing on the Z component for brevity… (but analogous considerations can be drawn for EW and NS) PRE variability: quite high some PRE events are similar to FON events. FON variability: high many FON events are similar to PRE or ERU events. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(20) Introduction Data and Methods Results Discussion. Misclassified Events Intra—class Variability. Intra—Class Variability :: ERU and POS Focusing on the Z component for brevity… (but analogous considerations can be drawn for EW and NS) ERU variability: quite low few ERU events are similar to FON events. POS variability: low very few POS events are similar to PRE events. Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(21) Introduction Data and Methods Results Discussion. Summary and Conclusions Summarizing… Volcanic tremor recorded at Mount Etna is automatically classified Data: 01 July—15 August, 2001 Features: Spectrogram—based Classifier: Support Vector Machine (SVM) Classification error: < 6% Concluding… Practical utility: on—line classification Practical/Scientific utility: off—line classification of huge (past) databases Practical/Scientific utility: the SVM classifier is a mathematical tool linking volcanic tremor to different states of the volcano’s activity in a reproducible way Matteo Masotti, et al.. April 5, 2006 — EGU General Assembly.
(22)
Documenti correlati
Emphasis is thus laid on chronology, or what Sternberg calls the ontological obscurity of the future (Telling II 531). We can draw up an incomplete inventory of the different
Tornando ora alla questione della narratività dei primi libri manganelliani, sono convinto che uno dei fattori che a lungo ha contribuito ad un’iscrizione quasi automatica dei
Tra gli elementi più antichi, tendenti al Bronzo Finale, citiamo i rari frammenti di probabili olle a colletto basso verticale o rientrante (tav. XX: 17-18), l’olla con orlo
Il paesaggio urbano del Lungomare di Marciana Marina nelle sue essenziali componenti ambientali sia connota in modo preponderante per le forti relazioni ambientali instaurate sia
BR has been obtained by normalizing the number of signal events to the number of the observed normal muon decays, taking into account the correction factors related to the differences
Cattaio, villa degli Obizi illustrata dal Betussi a 85. Cherubino da Spoleto, suo trattato spirituale, e supposta edizione del Giolito fatta nel 1563, da Remigio Fiorentino b 191.
This being the case, it is evident that there is a question of transgenerational justice to be answered, in relation to decisions and actions that have been made by a group of people
