Deep Machine Learning Based Possible Atmospheric and Ionospheric Precursors of the 2021 Mw 7.1 Japan Earthquake

Abstract

Global Navigation Satellite System (GNSS)- and Remote Sensing (RS)-based Earth observations have a significant approach on the monitoring of natural disasters. Since the evolution and appearance of earthquake precursors exhibit complex behavior, the need for different methods on multiple satellite data for earthquake precursors is vital for prior and after the impending main shock. This study provided a new approach of deep machine learning (ML)-based detection of ionosphere and atmosphere precursors. In this study, we investigate multi-parameter precursors of different physical nature defining the states of ionosphere and atmosphere associated with the event in Japan on 13 February 2021 (M_w 7.1). We analyzed possible precursors from surface to ionosphere, including Sea Surface Temperature (SST), Air Temperature (AT), Relative Humidity (RH), Outgoing Longwave Radiation (OLR), and Total Electron Content (TEC). Furthermore, the aim is to find a possible pre-and post-seismic anomaly by implementing standard deviation (STDEV), wavelet transformation, the Nonlinear Autoregressive Network with Exogenous Inputs (NARX) model, and the Long Short-Term Memory Inputs (LSTM) network. Interestingly, every method shows anomalous variations in both atmospheric and ionospheric precursors before and after the earthquake. Moreover, the geomagnetic irregularities are also observed seven days after the main shock during active storm days (Kp > 3.7; Dst < −30 nT). This study demonstrates the significance of ML techniques for detecting earthquake anomalies to support the Lithosphere-Atmosphere-Ionosphere Coupling (LAIC) mechanism for future studies.

1. Introduction

Earthquakes cause intense shaking of the Earth’s surface, with the earthquakes being the result of the brittle failure of the lithosphere. The 21st century has brought an evaluation to the field of satellite measurements, especially for earthquake precursors. Many researchers have observed anomalies prior to major earthquakes over the epicenter at different altitudes with various GNSS and RS satellites during the seismogenic period [1,2,3,4,5,6,7]. These short-term anomalies can be observed from 2 to 20 days before and 10 days after the main event, but still, no one presented a legitimate earthquake anomaly. Interestingly, these anomalies show positive and negative deviations beyond the defined bounds and are not only limited to the surface but these can also be observed in the ionosphere and atmosphere [8]. The lithosphere encounters various geophysical changes and as a result, these changes, precursors occurred over the earthquake epicenter [9]. In general, earthquake precursors are categorized into three different types; surface precursors, such as land surface temperature and SST, atmospheric precursors, such as OLR, RH, AT [5], and ionospheric precursors, such as TEC, electric field disturbance, and high energy particle flux [10,11,12,13]. The propagation of earthquake anomalies from the lithosphere to ionosphere can be defined by the LAIC hypothesis [9]. There is also another theory that correlates the ionosphere and atmosphere precursors with the impending main shock (i.e., Positive holes (p-holes)). During the earthquake preparation period, stressed rocks release p-holes that result in the formation of the electric current and consequently emit electromagnetic radiations that move upward to create vulnerability in the ionosphere [14]. Furthermore, the idea of radon emission from the epicentral region also gained importance for the presentation of earthquake anomalies by air ionization near the Earth’s surface, followed by emanation to the atmosphere and ultimately rising to the ionosphere [9]. In the epicentral region, emission of gases, such as radon, from the deformation of rocks due to tectonic stress can cause atmospheric anomalies. Moreover, ionospheric anomalies related to seismic activity over the epicenter are also reported by GNSS TEC. However, no clear ionospheric anomalies were observed by studying ionospheric anomalies around the 1279 earthquakes of M_w ≥ 6 during the time period of 2000–2014 [15]. Still, there are many well-published studies about the atmospheric and ionospheric anomalies associated with earthquakes. For example, both positive and negative TEC anomalies are detected by utilizing statistical analysis on ground receiver data for the M_w 7.8 Nepal earthquake [16]. Furthermore, significant TEC variations were also observed before and after the 2014 M_w 6.9 Samothrace earthquake [17]. An anomalous variation of more than 30 W/m² was observed in OLR for the Japan earthquakes and other events [5,18,19,20,21]. The reflection, emission, and absorption of OLR are due to a complicated system of aerosols, clouds, ocean surface temperature, and land surface temperature variations during the main shock preparation period [22]. Both the negative and positive anomalies are observed in the SST variable, which is considered to be the most important atmospheric precursor [23]. The emission of radon resulted in anomalous variations of AT and RH over the epicenter, as explained in the LAIC model [9]. Similarly, stress and tectonic blocks’ movement activate the thermodynamics phenomena and ionization process that result in the anomalous variations of AT and RH. Recent earthquakes showed clear enhancement in the atmospheric and ionospheric variations associated with main shocks [24,25,26,27,28,29,30,31,32]. GNSS and RS have played a vital role in understanding and finding the variations in atmospheric and ionospheric parameters from various space observations and ground stations [33,34,35,36,37]. The use of different algorithm on time series data is also very common these days [38,39,40,41,42]. Nevertheless, there is a gigantic gap of applications and knowledge related to the science of earthquake precursors.

In this paper, we have studied possible ionospheric and atmospheric anomalies associated with the M_w 7.1 Namie, Japan earthquake by implementing statistical as well as machine learning procedures (NARX and LSTM) on TEC and other remote sensing indices, including RH, OLR, and SST. The main aim is to find a synchronized window of the atmospheric and ionospheric anomalies within the same occurrence days. Moreover, another aim is to find the presence of pre-and/or post-seismic anomalies. The structure of this paper is as follows: Study area and brief description of earthquake data are described in Section 2 and Section 3, respectively. Methodology is explained in Section 4. Results are described in Section 5. Discussion is explained in Section 6. Section 7 is dedicated to the conclusion.

2. Study Area

A strong earthquake of M_w 7.1 struck the east coast of Honshu, Japan on 13 February 2021, at 14:07:49 UTC (LT = UTC + 09:00 = 23:07:49). The epicenter was located at 73 km of Namie, Japan (37.7°N, 141.7°E). The United States Geological Survey (USGS) announced the maximum felt intensity of IX (heavy) and shaking intensity of VIII (severe) on a modified Mercalli scale. One person died in Fukushima and around 187 people were injured. There was a massive destruction to property with an approximate value of the damages being 9137 houses and 311 schools. The earthquake occurred as a result of thrust faulting near the subduction zone interface plate boundary between the North American and Pacific plates (Figure 1). Subduction was beneath the Japan Trench. The Pacific plate moves westward relative to the North America plate, at a velocity of 70 mm/yr. The several micro plates in this region describe the relative motions among Pacific, North American, and Eurasian plates. The February 13, 2021 earthquake struck in the vicinity of the rapture area of the March 11, 2011 great Tohoku earthquake, which was widely felt along many islands of Japan and killed almost 16,000 people. The epicenter of the February 13 earthquake was located approximately 74 km from the epicenter of the Tohoku earthquake. After the March 2011 M_w 9.1 Tohoku earthquake, six earthquakes of M_w > 7 have occurred within 250 km of the February 13 earthquake. The USGS website contains more tectonic and technical information about this earthquake (https://earthquake.usgs.gov/earthquakes/eventpage/us6000dher/executive; accessed on 11 November 2022).

3. Materials and Datasets

3.1. Outgoing Longwave Radiation

OLR is one of the most vital parameters used to define the Earth’s radiation emitted from the earthquake regions. The OLR is a combination of emissions from clouds, the lower atmosphere, ground, which have been used to investigate Earth’s radiative climate [43,44,45,46]. The OLR (W/m²) data for both the day and night are acquired at spatial resolution of 1° of AIRS/Aqua from Goddard Earth Sciences Data and Information Services Center (GES DISC). The area averaged data of OLR time series covering the earthquake epicenter within a seismogenic area (139.4°E–143.9°E, 35.6°N–39.5°N) was retrieved from GIOVANNI with a web-based application developed by (GES DISC).

3.2. Relative Humidity

The RH is the ratio of the actual measured water vapor in the air to the saturation vapor pressure at a specific temperature. The main purpose of RH is to study the global energy budget, atmospheric dynamics, climate response related to earthquakes, and troposphere-stratosphere exchange over the epicenter. We study the RH data for both daytime and nighttime from AIRS/Aqua (GES DISC), GIOVANNI.

3.3. Air Temperature

AT indicates response of seismic activity in atmosphere and is used to understand the atmospheric phenomena related to main shock. AT (K) data is acquired from the National Oceanic and Atmospheric Administration (NOAA) at the spatial resolution of 2.5° × 2.5° along latitude × longitude. Moreover, some AT data was also retrieved from the National Centers for Environmental Prediction (NCEP)/National Center for Atmospheric Research (NCAR) reanalysis project.

3.4. Sea Surface Temperature

SST is one of the most significant variables used to figure out air-sea interaction and climate scenario over the epicenter. SST (°C) data was obtained from the Modern Resolution Imaging Spectroradiometer (MODIS) AQUA/ASMR-E Satellite. MODIS SST data have a very remarkable correlation with previous research for air-surface temperature information associated with future earthquakes.

3.5. Total Electron Content

In this paper, we also retrieved TEC from three IGS stations (USUD, MTKA, and HYDE). Two stations, USUD and MTKA (Japan), are operating in the earthquake breed zone, while the third station, HYDE (India), is outside of the earthquake-affected zone. We analyzed TEC data from the HYDE station to confirm the ionospheric anomalies of the geomagnetic storm and distinguish them from the earthquake ionospheric anomalies. The GPS TEC from these stations is retrieved via (http://www.ionolab.org/index.php?page=index&language=en; accessed on 21 November 2022). The GPS TEC values before and after the earthquake were examined in the earthquake preparation zone in the form of vertical TEC (VTEC). Moreover, VTEC values were obtained from Slant TEC (STEC) and examined in the TEC unit (1 TECU = 10¹⁶ el/m²). The STEC values were used to calculate the VTEC values, as shown below [47].

$$STE{C}_a^h=\frac{-(f_1^2{f}_2^2)}{40.3(f_1^2-f_2^2)}\left(P_{\left(4,a\right)}^h-c.DC{B}_a-c.DC{B}^h\right)$$

(1)

$$VTEC=STEC\times \cos \left[\mathit{arcsin}\left(\frac{Rsinz}{R+H}\right)\right]$$

(2)

where ($f_1^2,f_2^2)$ are the dual frequencies received at the GPS stations, $P_{\left(4,a\right)}^h$ is the difference between the smoothed coded measurements, $c$ is the speed of light, $DC{B}^h$ and $DC{B}_a$are differential code biases for GPS and satellite receiver, respectively. Similarly, $R$ is the Earth’s radius, $z$ is the zenith angle of satellite, and $H$ is the ionospheric height.

4. Methodology

We examined ionospheric and atmospheric data of 1 month (i.e., 20 days before and 10 days after) over the epicenter in earthquake breeding zone within the Dobrovolsky region [48]. For any earthquake, the stress radius (km) can be calculated by the following equation:

$$R={10}^{0.43M}$$

(3)

where, $R$ and M are the preparation zone and magnitude of the earthquake, respectively. Furthermore, the earthquake magnitude value affects the radius of the breeding zone; i.e., a large magnitude earthquake has larger stress radius and vice versa. The radius of the Japan earthquake is approximately 1130 km, as calculated by the Dobrovolsky formula.

4.1. Anomaly Detection Using Statistical Method

We calculated the confidence bounds based on the mean (M) and standard deviation (STDEV) to check and validate the any deviations in the atmospheric datasets for seismic influence. These confidence bounds (lower and upper) were obtained from the mean and standard deviation of all the values before and after the main shock by the below equations [10]:

$$UB=M+1.6\ast (STDEV)$$

(4)

$$LB=M-1.6\ast (STDEV)$$

(5)

The above equations declared that the calculated values falling outside these bounds will be considered as abnormal seismic values. The * is for multiplication.

4.2. Anomaly Detection Using Wavelet Transformation

To further validate the anomalous variations of the surface and atmospheric parameters, the wavelet transformation has been implemented on the OLR, AT, RH, and SST time series data of earthquakes.

$$\left({\omega }_{\mathsf{\Psi }}\right)\left(y,z\right)=\frac{1}{\sqrt{y}}{\int }_{-\infty }^{+\infty }f(x){\mathsf{\Psi }}^{\ast }\left[\frac{x-z}{y}\right]dx$$

(6)

where, $f(x)$ is the time series under analysis and ${\mathsf{\Psi }}^{\ast }$ is the complex conjugate of continuous wavelet function. Similarly, $y$ and $z$ are the scaling factor and location parameter, respectively. In this study, M ± 1.6(STDEV) has been selected as the optimum threshold value to detect the unusual values of the wavelet coefficients for the Japan earthquake. The anomalies outside the confidence bounds can be related to main shock.

4.3. Anomaly Detection Using Artificial Neural Network (ANN)

ANN has offered a promising technique for the time series forecasting. Various successful applications of this method proved that ANN can be very effective in the forecasting and modeling of non-linear time series [49,50]. ANN uses a complicated connection between the input and output parameters to show the anomalous pattern in the data. In this study, we input ANN as obtained values, mean values, deviation from mean values, upper and lower bounds values, and deviation from bounds values. Moreover, the anomalies in atmospheric parameters were observed based on the comparison between the observed and ANN predicted values.

4.4. Nonlinear Autoregressive Network with Exogenous Inputs (NARX)

The NARX model is widely used in time series modeling with feedback connections enclosing various layers of the network. The dependent output signal $y(t)$ estimation is regressed against both the independent input signal and the output signal in the NARX network architecture used to represent the time series. The NARX model-defining equation is below.

$$y\left(t\right)=f(y\left(t-1\right),y\left(t-2\right),..,y\left(t-n_y\right),x\left(t-1\right),x\left(t-2\right),..,x\left(t-n_x\right))$$

(7)

The network architecture and training estimation error were established repeatedly by acquiring the output, input delays, hidden layer count, related neurons, activation functions, and learning method. The function $f$ is a nonlinear function. To implement the prediction process, N observations$(y_1$, $y_2$, …, $y_N)$ were selected as the training set and the remaining ones $(y_{N+1}$, $y_{N+2}$, …, $y_{N+m}$) were considered to be the test set. In this study, a projected network was constructed consisting of six nodes in the input layer, 10 in the first and second hidden layers, and one in the output layer. The observed time series data, respective time, mean value, and deviation of observed time series data from the mean distribution are considered to be inputs. A time series of daily values are displayed within the Dobrovolsky region as the output layer. The training patterns in the proposed network are as follows:

$$y_4=f(y_1,y_2,y_3,t_1,t_2,t_3)$$

(8)

$$y_5=f(y_2,y_3,y_4,t_2,t_3,t_4)$$

(9)

$$y_N=f(y_{N-3},y_{N-2},y_{N-1},t_{N-3},t_{N-2},t_{N-1})$$

(10)

Finding optimum weights that reduce prediction error (PE) is often how prediction performance is assessed. The PE equation appears as follows:

$$PE={\sum }_{k-0}^N(\widehat{y}\left(t-k\right)-y(t-k))$$

(11)

where $\widehat{y}$ represents the output of the network. The testing patterns are:

$$y_{N+4}=f(y_{N+1},y_{N+2},y_{N+3},t_{N+1},t_{N+2},t_{N+3})$$

(12)

$$y_{N+5}=f(y_{N+2},y_{N+3},y_{N+4},t_{N+2},t_{N+3},t_{N+4})$$

(13)

$$y_{N+m}=f(y_{N+m-3},y_{N+m-2},y_{N+m-1},t_{N+m-3},t_{N+m-2},t_{N+m-1})$$

(14)

Finally, the deviations were calculated by comparing the variations between the actual and the predicted values crossing the pre-defined bounds.

4.5. Long Short-Term Memory (LSTM)

The LSTM model can analyze the dynamic temporal pattern in time series data while running through time. A closed layer of the LSTM has three units (output, input, and nine forget) to conduct the flow of information to the memory cell. The mathematical functions of the three units are as follows:

$$i_u=\sigma (W_i{x}_t+W_{Hi}{H}_{t-1}+b_i)$$

(15)

$$f_u=\sigma (W_f{x}_t+W_{Hf}{H}_{t-1}+b_f)$$

(16)

$$o_u=\sigma (W_o{x}_t+W_{Ho}{H}_{t-1}+b_o)$$

(17)

$$c_t=c_{t-1}\otimes {\left(f_u\right)}_t+{\left(i_u\right)}_t\otimes \tanh \left(W_c{x}_t+W_{Hc}{H}_{t-1}+b_c\right)$$

(18)

$$H_t=o_t\otimes \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}(c_{t-1})$$

(19)

$$y_t=H_t$$

(20)

The above equations illustrate input unit, forget unit, output unit, current memory cell unit, hidden unit, and output cell, respectively, at time $t$. The kernel functions are represented by $\sigma$ and $\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}$, whereas $H_{t-1}$ denotes the previously hidden unit and $c_t$ shows the current cell unit. Moreover, $b$ and $W$ indicate bias variables and weight matrices of three units and memory cell units, respectively. Similarly, $\otimes$ indicates the element-wise multiplication between the output, input, and cell units of closed layers.

5. Results

In this study, we analyzed multiple atmospheric and ionospheric datasets for possible earthquake precursors by various statistical as well as ML methods before and after the Japan earthquake. The resulting anomalies are discussed in detail.

5.1. Outgoing Longwave Radiation

We studied the variations in both daytime and nighttime OLR values over the epicenter of the Japan earthquake (Figure 2). The time series analysis of OLR daytime showed immense deviations over the epicenter with considerable anomalies prior to the Japan earthquake. There was an obvious positive anomaly of 36 W/m² on six days before the earthquake (Figure 2a).

Similarly, a positive anomaly of 11.4 W/m² was also observed on the fifth day before the main shock. The time series analysis of OLR nighttime also showed a positive anomaly of 16 W/m² on the sixth day before the earthquake (Figure 2b). On the other hand, the wavelet transformation method implemented on OLR values for daytime and nighttime has demonstrated more evidence about the earthquake-induced anomalies. A clear OLR daytime anomaly of high magnitude occurred on the sixth and fifth days before the main shock (Figure 2e) and a significant nighttime OLR on the fifth day before the main shock (Figure 2f). Moreover, the deviations of NARX-predicted OLR daytime values showed clear anomalies within the five-day window before the seismic event (Figure 3c). Figure 3g showed a clear variation in LSTM-predicted OLR daytime values within the five-day window before the main event. NARX- and LSTM-predicted OLR nighttime values also endorse the anomalies within the 10-day window of the pre-seismic event (Figure 3d–h).

5.2. Relative Humidity

The RH data for both daytime and nighttime are evaluated to observe the variations prior to the Japan event. Time series analysis of the RH daytime values showed a negative anomaly of −8% on the sixth day before the main earthquake day (Figure 4a). Analysis of RH nighttime values also showed a negative anomaly below the bound of Mean−1.6*STDEV of −6% on the fifth day prior to the main shock (Figure 4b). We also noticed sharp variation in the continuous wavelet transformation in both daytime and nighttime RH values of high magnitude within the 10-day window prior to the main shock (Figure 4e,f). NARX- and LSTM-predicted RH daytime values showed a clear negative anomaly on the sixth day before the event (Figure 5c–g). Moreover, Figure 5 showed clear deviations on the sixth and fifth day before the earthquake for both NARX- and LSTM-predicted values in nighttime RH.

5.3. Air Temperature

In this paper, we also studied the AT values of the Japan earthquake, which showed a clear positive anomaly on the fifth day before the main shock (Figure 6a). Moreover, an obvious positive anomaly also occurred beyond the Mean+1.6*STDEV bound on 8th February with a deviation of 2K. The continuous wavelet transformation showed a variation of high magnitude within the five-day period before the main earthquake (Figure 6c). NARX- and LSTM-predicted AT values also showed clear anomalies within five days before the main shock day (Figure 7a–c).

5.4. Sea Surface Temperature

We also studied the time series data of SST, which showed a clear negative anomaly of −4.5 °C on 7th February; i.e., six days before the earthquake (Figure 8a). This sudden decrease is observed on the sixth day before the main shock of the Japan event. Similarly, continuous wavelet transformation showed variations of high magnitude within −8 to −5 days before the earthquake, which is synchronized with the time series anomaly (Figure 8c). The comparison of NARX- and LSTM-predicted SST values with observed values also showed clear deviations −7 and −6 days prior to the main shock (Figure 9b–d).

5.5. Total Electron Content

TEC data retrieved from the two IGS stations (USUD and MTKA) within the breeding zone of the M_w 7.1, Japan earthquake showed clear seismo-ionospheric anomalies. IGS station (HYDE) variation was also observed outside the seismic preparation zone to illustrate the earthquake- and storm-induced TEC anomalies. Figure 10 illustrates the associated VTEC variations recorded at two IGS stations, which operate within the seismic preparation zone. Prominent positive deviations were observed −6 and −5 days before the main shock in both USUD and MTKA stations during quiet storm days. These anomalies are synchronized with the atmospheric anomalies and occurred in the same window as atmospheric anomalies. Similarly, NARX- and LSTM-predicted values also showed clear deviations in VTEC values on the 7th and 8th February for the two IGS stations i.e., USUD and MTKA (Figure 11 and Figure 12). USUD VTEC showed positive anomalies of 7.5 and 1.8 TECU on seven and nine days after the major event (Figure 10c). Figure 10d showed positive variations of 6.6, 1.6, and 6.7 TECU for MTKA VTEC on seven, eight, and nine days after the earthquake day, respectively. Positive TEC deviations were also observed seven and eight days after the seismic event for HYDE station (Figure 10e). During active storm days, VTEC variations for three IGS stations (USUD, MTKA, and HYDE) showed positive anomalies from 7–10 days after the earthquake day. NARX- and LSTM-predicted VTEC also showed clear deviations from 20th to 23rd February for three IGS stations during active storm days (Figure 11d and Figure 12d). All the atmospheric and ionospheric anomalies are listed in

Table 1. The anomalies detected using statistical method.

Parameters	Anomalous Day	Deviations from UB and LB
		Pre-EQ	Post-EQ
OLR (Daytime)	−6, −5	36, 11.4 W/m2	Nil
OLR (Nighttime)	−6	18 W/m2	Nil
RH (Daytime)	−6	−8%	Nil
RH (Nighttime)	−5	−6%	Nil
AT	−5	2 °K	Nil
SST	−6	−4.5 °C	Nil
VTEC (USUD)	−6, −5, 7, 9	3, 1.5 TECU	7.5, 1.8 TECU
VTEC (MTKA)	−6, −5, 7, 8, 9	2.32, 0.5 TECU	6.6, 1.5, 6.7 TECU
VTEC (HYDE)	7, 8	Nil	5.95, 0.8 TECU

, Table 2 and Table 3.

6. Discussion

In this paper, we analyzed OLR, RH, AT, SST, and TEC time series data’s possible deviations within the preparation zone of the Japan earthquake. We found these anomalies due to the immense energy release during the main shock days; Ouzounov [51] showed the earthquake energy outflow towards the atmosphere. Emission of gases from the Earth’s crust caused OLR anomalies, variations in AT and RH, and also resulted in air ionization, which is a notable process in the local area of active faults. There exists a strong correlation between the radon variations during the earthquake preparation period with RH, and AT in the LAI coupling over the epicenter [9,52]. There are various studies on abnormal AT resulting in lowest RH (decrease in RH due to the cooling process of hot gases emitted from the seismic region), which is a possible earthquake precursor. In this study, we also found a significant increase in OLR as a result of sudden earthquake energy release before the main shock (Figure 2), followed by abnormal decrement in RH due to emission of hot gases from the epicentral region (Figure 4). Similarly, OLR is a very important parameter and it can be efficiently used as a possible short-term earthquake precursor [52]. Similarly, rise in surface temperature can also increase OLR, specifically in the earthquake breeding zone. In this paper, daily OLR, RH, and AT deviations over the epicenter of the Japan earthquake showed anomalous behavior beyond the confidence bounds as well as with the machine learning techniques. Previous studies also suggested the decrement in RH and increment in other atmosphere indices before the earthquake [5,6,25,30,31,51,52]. However, we found synchronized and co-located atmospheric and ionospheric anomalies on the fifth and sixth days before the main shock, followed by abnormalities in the data of nearby GPS stations (Table 1, Table 2 and Table 3). The analysis in this study can be further supported by the LAIC hypothesis, which also compromised synchronized anomalies over the epicenter. Variations of OLR, RH, and AT values were also considered as evidence of seismic-induced deformation along the fault lines of seismic event regions [6,53]. SST data also showed strong evidence of disturbance around the earthquake and support of the study of other atmospheric anomalies. Therefore, this finding is consistent and compatible with previous studies [23,54,55]. There are several studies about the LAIC coupling in pre-and post-seismic effects for various events. However, some studies suggested that the earthquake-induced positive holes within the breeding zone could variant the air ionization and consequently rise the electrical conductivity [56,57]. Furthermore, Pulinets and Ouzounov [9] demonstrated that ionospheric VTEC anomalies occurrence from the radon emission of tectonic-stressed rocks during the earthquake preparation period. Additionally, high energy alpha particles, which are ionized by the initial emissions from the earthquake, could raise air conductivity by radioactive decay of the radon [7,58,59]. Moreover, the LAIC phenomenon was explained by the numerical simulations of ionospheric ionization at the lower and upper edge of the atmosphere by the induced earthquake upward electric field [60,61,62,63]. Furthermore, Kuo [60] demonstrated that the increase or decrease in ionospheric perturbations can occur only above the epicenter due to plasma flow from the epicenter. Moreover, many studies have reported several atmospheric and ionospheric anomalies within 2–3 months before and after the earthquake by implementation of the LAIC model [64,65]. We found significant positive and negative anomalies in both atmospheric and ionospheric parameters associated with the M_w 7.1, Japan earthquake. However, there is still a need for observations and analysis for detailed calculations of different earthquake precursors.

Table 2. Detected anomalies using NARX model.

Parameters	Anomalous Day	Deviations from NARX Predicted Values
		Pre-EQ	Post-EQ
OLR (Daytime)	−6, −5	58, 19.8 W/m2	Nil
OLR (Nighttime)	−6, −5	48, 6 W/m2	Nil
RH (Daytime)	−6	−15%	Nil
RH (Nighttime)	−6, −5	−6, −27%	Nil
AT	−5	8 °K	Nil
SST	−7, −6	−2.9, −9.7 °C	Nil
VTEC (USUD)	−6, −5, 7, 9	3.2, 2.9 TECU	7.6, 1.7 TECU
VTEC (MTKA)	−6, −5, 7, 8, 9	3.3, 1.48 TECU	7.3, 1.8, 6.3 TECU
VTEC (HYDE)	7, 8	Nil	9.7, 6.47 TECU

Table 2. Detected anomalies using NARX model.

Parameters	Anomalous Day	Deviations from NARX Predicted Values
		Pre-EQ	Post-EQ
OLR (Daytime)	−6, −5	58, 19.8 W/m2	Nil
OLR (Nighttime)	−6, −5	48, 6 W/m2	Nil
RH (Daytime)	−6	−15%	Nil
RH (Nighttime)	−6, −5	−6, −27%	Nil
AT	−5	8 °K	Nil
SST	−7, −6	−2.9, −9.7 °C	Nil
VTEC (USUD)	−6, −5, 7, 9	3.2, 2.9 TECU	7.6, 1.7 TECU
VTEC (MTKA)	−6, −5, 7, 8, 9	3.3, 1.48 TECU	7.3, 1.8, 6.3 TECU
VTEC (HYDE)	7, 8	Nil	9.7, 6.47 TECU

Table 3. Detected anomalies using LSTM method.

Parameters	Anomalous Day	Deviations from LSTM Predicted Values
		Pre-EQ	Post-EQ
OLR (Daytime)	−6, −5	53.8, 26 W/m2	Nil
OLR (Nighttime)	−6, −5	54, 9 W/m2	Nil
RH (Daytime)	−6	−15.7%	Nil
RH (Nighttime)	−6, −5	−13.3, −34.5%	Nil
AT	−5	7 °K	Nil
SST	−7, −6	−2, −8 °C	Nil
VTEC (USUD)	−6, −5, 7, 9	3.18, 2 TECU	7.4, 2.1 TECU
VTEC (MTKA)	−6, −5, 7, 8, 9	3.23, 1.39 TECU	6.5, 1.6, 5.8 TECU
VTEC (HYDE)	7, 8	Nil	9.63, 6.32 TECU

Table 3. Detected anomalies using LSTM method.

Parameters	Anomalous Day	Deviations from LSTM Predicted Values
		Pre-EQ	Post-EQ
OLR (Daytime)	−6, −5	53.8, 26 W/m2	Nil
OLR (Nighttime)	−6, −5	54, 9 W/m2	Nil
RH (Daytime)	−6	−15.7%	Nil
RH (Nighttime)	−6, −5	−13.3, −34.5%	Nil
AT	−5	7 °K	Nil
SST	−7, −6	−2, −8 °C	Nil
VTEC (USUD)	−6, −5, 7, 9	3.18, 2 TECU	7.4, 2.1 TECU
VTEC (MTKA)	−6, −5, 7, 8, 9	3.23, 1.39 TECU	6.5, 1.6, 5.8 TECU
VTEC (HYDE)	7, 8	Nil	9.63, 6.32 TECU

7. Conclusions

In this paper, we have investigated the possible atmospheric and ionospheric anomalies associated with the large magnitude earthquake of Mw 7.1 in Japan from various ground-and space-measurements. The main findings are: For the daytime and nighttime analysis, the unusual variations in the atmospheric constituents beyond the confidence bounds occurred −6 and −5 days before the earthquake main shock. Moreover, the wavelet transformation and NARX and LSTM train data also endorse the anomalous behavior of atmospheric anomalies on the same days as the statistical method. The abnormal variations in OLR, SST, AT, and RH validate the coupling of the earthquake and atmosphere by the LAIC hypothesis. The sudden drop in RH confirmed the cooling of hot gases from seismogenic regions in LAIC phenomenon, followed by a sudden increase in OLR and SST due to abnormal ions and water particle drift from the lower atmosphere toward the upper atmosphere. Moreover, we also found synchronized and co-located ionosphere anomalies −6 and −5 days before the main shock with the two available stations (USUD and MTKA) within the Dobrovolsky region during quiet geomagnetic storms. All these anomalies in the atmosphere and ionosphere are due to electric field generation around the epicenter, where the charge particle propagated to the atmosphere and ionosphere via the p-hole and radon emission to form the LAIC mechanism. The statistical and ML data showed clear earthquake variations; however, more techniques are needed to develop forecasting of possible earthquake precursors from the available satellite clusters.