Towards advancing the earthquake forecasting by machine learning of satellite data.

Earthquakes have become one of the leading causes of death from natural hazards in the last fifty years. Continuous efforts have been made to understand the physical characteristics of earthquakes and the interaction between the physical hazards and the environments so that appropriate warnings may be generated before earthquakes strike. However, earthquake forecasting is not trivial at all. Reliable forecastings should include the analysis and the signals indicating the coming of a significant quake. Unfortunately, these signals are rarely evident before earthquakes occur, and therefore it is challenging to detect such precursors in seismic analysis. Among the available technologies for earthquake research, remote sensing has been commonly used due to its unique features such as fast imaging and wide image-acquisition range. Nevertheless, early studies on pre-earthquake and remote-sensing anomalies are mostly oriented towards anomaly identification and analysis of a single physical parameter. Many analyses are based on singular events, which provide a lack of understanding of this complex natural phenomenon because usually, the earthquake signals are hidden in the environmental noise. The universality of such analysis still is not being demonstrated on a worldwide scale. In this paper, we investigate physical and dynamic changes of seismic data and thereby develop a novel machine learning method, namely Inverse Boosting Pruning Trees (IBPT), to issue short-term forecast based on the satellite data of 1371 earthquakes of magnitude six or above due to their impact on the environment. We have analyzed and compared our proposed framework against several states of the art machine learning methods using ten different infrared and hyperspectral measurements collected between 2006 and 2013. Our proposed method outperforms all the six selected baselines and shows a strong capability in improving the likelihood of earthquake forecasting across different earthquake databases.

The Effect of Catalogue Lead Time on Medium-Term Earthquake Forecasting with Application to New Zealand Data.

'Every Earthquake a Precursor According to Scale' (EEPAS) is a catalogue-based model to forecast earthquakes within the coming months, years and decades, depending on magnitude. EEPAS has been shown to perform well in seismically active regions like New Zealand (NZ). It is based on the observation that seismicity increases prior to major earthquakes. This increase follows predictive scaling relations. For larger target earthquakes, the precursor time is longer and precursory seismicity may have occurred prior to the start of the catalogue. Here, we derive a formula for the completeness of precursory earthquake contributions to a target earthquake as a function of its magnitude and lead time, where the lead time is the length of time from the start of the catalogue to its time of occurrence. We develop two new versions of EEPAS and apply them to NZ data. The Fixed Lead time EEPAS (FLEEPAS) model is used to examine the effect of the lead time on forecasting, and the Fixed Lead time Compensated EEPAS (FLCEEPAS) model compensates for incompleteness of precursory earthquake contributions. FLEEPAS reveals a space-time trade-off of precursory seismicity that requires further investigation. Both models improve forecasting performance at short lead times, although the improvement is achieved in different ways.

Groundwater oxygen anomaly related to the 2016 Kumamoto earthquake in Southwest Japan.

Here, we report the groundwater oxygen isotope anomalies caused by the 2016 Kumamoto earthquake (MJMA7.3) that occurred in Southwest Japan on April 16, 2016. One hundred and seventeen groundwater samples were collected from a deep well located 3 km to the southeast of the epicenter in Mifune Town, Kumamoto Prefecture; they were drinking water packed in PET bottles and distributed in the area between April 2015 and March 2018. Further, the oxygen and hydrogen isotopes were evaluated via cavity ring-down spectroscopy without performing any pretreatment. An anomalous increase was observed with respect to the δ18O value (up to 0.51‰) soon after the earthquake along with a precursory increase of 0.38‰ in January 2016 before the earthquake. During these periods, there was no noticeable change in the hydrogen isotopic ratios. Rapid crustal deformation related to the earthquake may have enhanced the microfracturing of the aquifer rocks and the production of new surfaces, inducing δ18O enrichment via oxygen isotopic exchange between rock and porewater without changing δ2H.

Identification of earthquake precursors in soil radon-222 data of Kutch, Gujarat, India using empirical mode decomposition based Hilbert Huang Transform.

Soil radon (Rn-222) has been continuously monitored at Badargadh station (23.47°N, 70.62°E) in Kutch region of Gujarat to study the pre-seismic anomalies prior to occurrence of local earthquakes. This monitoring site is in close proximity to the South Wagad Fault, a seismically active fault in the study area. The raw data of radon along with meteorological parameters such as temperature, pressure and humidity in soil of this station for the period of January 01 to December 31, 2017 with a sampling interval of 10 min were used in the analysis. The wind speed and rainfall data of the corresponding period were collected from the nearest weather station. From descriptive statistics, we found an average soil radon concentration of 343 Bq.m-3. It is observed that radon has a maximum concentration during the rainy season compared to the other two seasons. We found that radon emission rate is less during mid-nights and early morning, whereas, the radon emission is more during afternoon hours when the sun light intensity is more. In order to identify and extract the periodic oscillations in the radon time series, the Empirical Mode Decomposition (EMD) was applied to the soil radon (Rn-222) time series by decomposing it into different oscillatory modes known as the Intrinsic Mode Function (IMF). Several interesting non-linear features emerged from the analysis after applying Hilbert Huang Transform (HHT) on significant IMFs. The temporal variation of the instantaneous energy is well correlated with four local earthquakes during the study period. Most interestingly, intermittencies in the temporal evolution of the instantaneous energy function have been observed prior to these local earthquakes. We present the results of the seismic and aseismic periods as well as a brief discussion of the analysis of radon data which can be used as a precursor of seismic activity. It is now possible to identify anomalies in radon time series using EMD based HHT method even for small-magnitude earthquakes.

Statistical Significance of Earth's Electric and Magnetic Field Variations Preceding Earthquakes in Greece and Japan Revisited.

By analyzing the seismicity in a new time domain, termed natural time, we recently found that the change of the entropy under time reversal (Physica A2018, 506, 625-634) and the relevant complexity measures (Entropy2018, 20, 477) exhibit pronounced variations before the occurrence of the M8.2 earthquake in Mexico on 7 September 2017. Here, the statistical significance of precursory phenomena associated with other physical properties and in particular the anomalous variations observed in the Earth's electric and magnetic fields before earthquakes in different regions of the world and in particular in Greece since 1980s and Japan during 2001-2010 are revisited (the latter, i.e., the magnetic field variations are alternatively termed ultra low frequency (ULF) seismo-magnetic phenomena). Along these lines we employ modern statistical tools like the event coincidence analysis and the receiver operating characteristics technique. We find that these precursory variations are far beyond chance and in addition their lead times fully agree with the experimental findings in Greece since the 1980s.

The Behavior of VLF/LF Variations Associated with Geomagnetic Activity, Earthquakes, and the Quiet Condition Using a Neural Network Approach.

The neural network approach is proposed for studying very-low- and low-frequency (VLF and LF) subionospheric radio wave variations in the time vicinities of magnetic storms and earthquakes, with the purpose of recognizing anomalies of different types. We also examined the days with quiet geomagnetic conditions in the absence of seismic activity, in order to distinguish between the disturbed signals and the quiet ones. To this end, we trained the neural network (NN) on the examples of the representative database. The database included both the VLF/LF data that was measured during four-year monitoring at the station in Petropavlovsk-Kamchatsky, and the parameters of seismicity in the Kuril-Kamchatka and Japan regions. It was shown that the neural network can distinguish between the disturbed and undisturbed signals. Furthermore, the prognostic behavior of the VLF/LF variations indicative of magnetic and seismic activity has a different appearance in the time vicinity of the earthquakes and magnetic storms.

Regional Seismic Information Entropy to Detect Earthquake Activation Precursors.

A method is presented to detect earthquake precursors from time series data on earthquakes in a target region. The Regional Entropy of Seismic Information (RESI) is an index that represents the average influence of an earthquake in a target region on the diversity of clusters to which earthquake foci are distributed. Based on a simple qualitative model of the dynamics of land crust, it is hypothesized that the saturation that occurs after an increase in RESI precedes the activation of earthquakes. This hypothesis is validated by the earthquake catalog. This temporal change was found to correlate with the activation of earthquakes in Japanese regions one to two years ahead of the real activation, more reliably than the compared baseline methods.

Multifractality of Pseudo-Velocities and Seismic Quiescence Associated with the Tehuantepec M8.2 EQ.

By using earthquake catalogs, previous studies have reported evidence that some changes in the spatial and temporal organization of earthquake activity are observed before and after of a main shock. These previous studies have used different approaches for detecting clustering behavior and distance-events density in order to point out the asymmetric behavior of foreshocks and aftershocks. Here, we present a statistical analysis of the seismic activity related to the M w = 8.2 earthquake that occurred on 7 September 2017 in Mexico. First, we calculated the inter-event time and distance between successive events for the period 1 January 1998 until 20 October 2017 in a circular region centered at the epicenter of the M w = 8.2 EQ. Next, we introduced the concept of pseudo-velocity as the ratio between the inter-event distance and inter-event time. A sliding window is considered to estimate some statistical features of the pseudo-velocity sequence before the main shock. Specifically, we applied the multifractal method to detect changes in the spectrum of singularities for the period before the main event on 7 September. Our results point out that the multifractality associated with the pseudo-velocities exhibits noticeable changes in the characteristics of the spectra (more narrower) for approximately three years, from 2013 until 2016, which is preceded and followed by periods with wider spectra. On the other hand, we present an analysis of patterns of seismic quiescence before the M w = 8.2 earthquake based on the Schreider algorithm over a period of 27 years. We report the existence of an important period of seismic quietude, for six to seven years, from 2008 to 2015 approximately, known as the alpha stage, and a beta stage of resumption of seismic activity, with a duration of approximately three years until the occurrence of the great earthquake of magnitude M w = 8.2 . Our results are in general concordance with previous results reported for statistics based on magnitude temporal sequences.

Statistical Evaluations of Variations in Dairy Cows' Milk Yields as a Precursor of Earthquakes.

Previous studies have provided quantitative data regarding unusual animal behavior prior to earthquakes; however, few studies include long-term, observational data. Our previous study revealed that the milk yields of dairy cows decreased prior to an extremely large earthquake. To clarify whether the milk yields decrease prior to earthquakes, we examined the relationship between earthquakes of various magnitudes and daily milk yields. The observation period was one year. In the results, cross-correlation analyses revealed a significant negative correlation between earthquake occurrence and milk yields approximately three weeks beforehand. Approximately a week and a half beforehand, a positive correlation was revealed, and the correlation gradually receded to zero as the day of the earthquake approached. Future studies that use data from a longer observation period are needed because this study only considered ten earthquakes and therefore does not have strong statistical power. Additionally, we compared the milk yields with the subionospheric very low frequency/low frequency (VLF/LF) propagation data indicating ionospheric perturbations. The results showed that anomalies of VLF/LF propagation data emerged prior to all of the earthquakes following decreases in milk yields; the milk yields decreased earlier than propagation anomalies. We mention how ultralow frequency magnetic fields are a stimulus that could reduce milk yields. This study suggests that dairy cow milk yields decrease prior to earthquakes, and that they might respond to stimuli emerging earlier than ionospheric perturbations.

The LVD signals during the early-mid stages of the L'Aquila seismic sequence and the radon signature of some aftershocks of moderate magnitude.

The L'Aquila seismic swarm culminated with the mainshock of April 6, 2009 (ML = 5.9). Here, we report and analyze the Large Volume Detector (LVD, used in neutrinos research) low energy traces (∼0.8 MeV), collected during the early-mid stages of the seismic sequence, together with the data of a radon monitoring experiment. The peaks of LVD traces do not correlate with the evolution and magnitude of earthquakes, including major aftershocks. Conversely, our radon measurements obtained by utilizing three automatic stations deployed along the regional NW-SE faulting system, seem to be, in one case, more efficient. In fact, the timeseries collected on the NW-SE Paganica fracture recorded marked variations and peaks that occurred during and prior moderate aftershocks (with ML > 3). The Paganica monitoring station (PGN) seems to better responds to active seismicity due to the fact that the radon detector was placed directly within the bedrock of an active fault. It is suggested that future networks for radon monitoring of active seismicity should preferentially implement this setting.

Unusual Animal Behavior Preceding the 2011 Earthquake off the Pacific Coast of Tohoku, Japan: A Way to Predict the Approach of Large Earthquakes.

Unusual animal behaviors (UABs) have been observed before large earthquakes (EQs), however, their mechanisms are unclear. While information on UABs has been gathered after many EQs, few studies have focused on the ratio of emerged UABs or specific behaviors prior to EQs. On 11 March 2011, an EQ (Mw 9.0) occurred in Japan, which took about twenty thousand lives together with missing and killed persons. We surveyed UABs of pets preceding this EQ using a questionnaire. Additionally, we explored whether dairy cow milk yields varied before this EQ in particular locations. In the results, 236 of 1,259 dog owners and 115 of 703 cat owners observed UABs in their pets, with restless behavior being the most prominent change in both species. Most UABs occurred within one day of the EQ. The UABs showed a precursory relationship with epicentral distance. Interestingly, cow milk yields in a milking facility within 340 km of the epicenter decreased significantly about one week before the EQ. However, cows in facilities farther away showed no significant decreases. Since both the pets' behavior and the dairy cows' milk yields were affected prior to the EQ, with careful observation they could contribute to EQ predictions.

Ground water chemistry changes before major earthquakes and possible effects on animals.

Prior to major earthquakes many changes in the environment have been documented. Though often subtle and fleeting, these changes are noticeable at the land surface, in water, in the air, and in the ionosphere. Key to understanding these diverse pre-earthquake phenomena has been the discovery that, when tectonic stresses build up in the Earth's crust, highly mobile electronic charge carriers are activated. These charge carriers are defect electrons on the oxygen anion sublattice of silicate minerals, known as positive holes, chemically equivalent to O- in a matrix of O2-. They are remarkable inasmuch as they can flow out of the stressed rock volume and spread into the surrounding unstressed rocks. Travelling fast and far the positive holes cause a range of follow-on reactions when they arrive at the Earth's surface, where they cause air ionization, injecting massive amounts of primarily positive air ions into the lower atmosphere. When they arrive at the rock-water interface, they act as •O radicals, oxidizing water to hydrogen peroxide. Other reactions at the rock-water interface include the oxidation or partial oxidation of dissolved organic compounds, leading to changes of their fluorescence spectra. Some compounds thus formed may be irritants or toxins to certain species of animals. Common toads, Bufo bufo, were observed to exhibit a highly unusual behavior prior to a M6.3 earthquake that hit L'Aquila, Italy, on April 06, 2009: a few days before the seismic event the toads suddenly disappeared from their breeding site in a small lake about 75 km from the epicenter and did not return until after the aftershock series. In this paper we discuss potential changes in groundwater chemistry prior to seismic events and their possible effects on animals.