Introduction
Since the Fukushima Daiichi Nuclear Power Plant (FDNPP) accident, several types of radiation monitoring methods have been introduced in Japan. Ambient dose rate (air dose rate) surveys were conducted using walking surveys and unmanned helicopters around Fukushima [1, 2]. Airborne radiation survey (ARS) using unmanned helicopters provides the advantage of covering large areas, including forested regions. However, the ARS measurement resolution is lower than that of walking surveys. In a previous study, Sasaki et al. [3, 4] developed an artificial neural network (ANN) to convert ARS data into air dose rates at 1 m above ground level using the data accumulated after the FDNPP accident. ANN-based conversions have been reported to exhibit higher accuracy than traditional methods. In their work, Sasaki et al. [3, 4] used unmanned helicopter measurements, terrain, and photographic data as explanatory variables for the ANN, with ground measurement data used as objective variables. The developed ANN was trained using data obtained from a lanthanum bromide (LaBr3) scintillation detector. However, the ANN developed here (F-ANN) was not applicable to measurement data obtained from detectors other than those used during training. Examples of extensive radiation contamination, such as Fukushima, are rare, and the accumulated data are highly valuable. Ensuring the applicability of the F-ANN obtained from Fukushima experience to measurements by other countries or organizations is an important contribution to the future advancement of ARS.
The purpose of this study is to make the F-ANN obtained from the Fukushima experience applicable to other institutions as well. In this study, we conducted verification to determine whether the F-ANN could be applied even when measured with detectors different from the F-ANN original detector. We present the results of tests and evaluations conducted to assess the applicability of the F-ANN to data obtained from monitoring of ambient radiation of KAERI-A1 (MARK-A1) Cadmium–Zinc–Telluride (CZT) detectors developed by the Korea Atomic Energy Research Institute (KAERI) as emergency radiation monitoring systems developed in Republic of Korea and cerium bromide (CeBr3) scintillation detectors developed by the Japan Atomic Energy Agency.
Materials and Methods
In this study, we applied the F-ANN to measurement data obtained from detectors different from the original detector used in the training of the F-ANN. The original detector used in the F-ANN is the LaBr3 scintillation detector (three detectors with 38 mm diameter and 38 mm height). Here, the detectors other than the original detector, which are the target for application with the F-ANN, are referred to as target detectors. In this study, we adopted the MARK-A1 (CZT detector; 15 mm×15 mm×7.5 mm, MARK-A1 [5]) developed by KAERI as an emergency radiation monitoring system developed in Republic of Korea, and the CeBr3 scintillation detector (two detectors with 50 mm diameter and 50 mm height) developed by the Japan Atomic Energy Agency. Fig. 1 shows the radiation spectra obtained when measuring with the original detector and the CZT detector used in this study, while Fig. 2 shows the radiation spectra obtained when measuring with the original detector and the CeBr3 detector, under the same environmental conditions. As shown in Figs. 1 and 2, the radiation spectrum information obtained from the original detector and the target detector differs even under the same measurement conditions due to the characteristics of the detectors. In this study, to apply the data from the target detectors to the F-ANN, the data from the target detectors were converted to the same level as the original detector. Fig. 3 illustrates the process of applying the data from the target detectors to the F-ANN in this study, and the conversion procedure is outlined below.
1. Conducting Calibration Flights
The original detector and the target detector were subjected to hovering flights under the same conditions, and measurement data were acquired. For CZT, hovering measurements were conducted for 5 minutes at altitudes of 80, 50, 40, 30, 20, and 10 m. For CeBr3, hovering measurements were conducted for 3 minutes at altitudes of 150, 120, 100, 80, 60, 40, 30, and 20 m.
2. Comparing the Information of Each Count Rate Obtained from the Calibration Flights
We obtained count rates within the energy bands of 50–450 keV, 450–900 keV, and 900–2,800 keV from the radiation spectra obtained during the calibration flights and examined the relationship between the count rates of the original detector and the target detector.
3. Creating Conversion Coefficients
Conversion coefficients were obtained from the linear approximation of the count rates for both the original detector and the target detector as Equation (1):
where Yoriginal represents the counting rate of original detector (LaBr3), Ytarget represents the counting rate of target detector, and a denotes the conversion factor of the counting level. Using the obtained a here, the data from the target detector were converted to the same level as the original detector.
When there is a strong correlation between the counting rates of the original detector and the target detector, the counting rate levels can be directly converted and applied to the ANN. However, if no correlation is observed or if data cannot be obtained, those values cannot be used as explanatory variables for the ANN. Detailed results comparing the counting rates of the original detector and the target detector in this study will be discussed in the “Results” section. It was found that the CeBr3 detector exhibited a strong correlation with various counting rates, while the CZT detector showed a low detection rate, especially in the 900–2,800 keV range where the counting rate of the target detector was around 0–1 counts per second (cps). Therefore, in this study, when using CZT data, the counting rate in the 900–2,800 keV range was not included as an input variable for the ANN. Consequently, to accommodate cases where the input variables are insufficient with the data from the target detector, a modified version of the F-ANN with reduced input variables was reconstructed. It should be noted that the original F-ANN model was used for the CeBr3 detector.
The list of explanatory and objective variables used in the Fukushima ANN employed in this study is presented in Table 1. The reconstructed F-ANN was developed based on radiation monitoring data collected in the vicinity of the FDNPP from 2018 to 2021, utilizing the same dataset as the original Fukushima ANN [1, 2]. The construction process of the ANN followed the methodology proposed by Sasaki et al. [3] and Sasaki and Sanada [4].
To evaluate whether there are significant differences between the original detector and the target detector, ARS surveys were conducted in the areas shown in Figs. 4 and 5 to acquire evaluation data. Fig. 4 depicts ARS surveys conducted with the original detector and the CZT detector, while Fig. 5 shows surveys conducted with the original detector and the CeBr3 detector. Both survey areas were affected by the FDNPP accident.
The area covered in Fig. 4 was approximately 72,200 m2, with an average air dose rate at 1 m above ground level of 6.5 μSv/hr (measured through walking surveys). About 40% of the area consisted of forested regions. The area covered in Fig. 5 was approximately 48,500 m2, with an average air dose rate at 1 m above ground level of 7.3 μSv/hr. Originally, this area was a rice paddy, but it had become densely wooded.
The CZT measurements depicted in Fig. 4 were conducted using an unmanned helicopter at an altitude of 60–90 m, with a measurement width of 20 m and a speed of 2 m/s. Two flight methods, flat flight and step flight, were employed for the unmanned helicopter flights in Fig. 4. In flat flight, the unmanned helicopter maintained a constant altitude, while in step flight, it flew at various altitudes following the terrain. The CeBr3 measurements shown in Fig. 5 were conducted using an unmanned helicopter at an altitude of approximately 30 m, with a measurement width of 20 m and a speed of 2 m/s. Since the area depicted in Fig. 5 was flat, only flat flight was conducted. The data obtained from the measurements were meshed into 10 m grids to apply to the F-ANN as input data.
In addition to the conversion using ANN, the ARS data was also transformed using the commonly employed flat source model (FSM) for comparison [1, 6]. Equation (2) was utilized to convert the count rates obtained from ARS to air dose rates at 1 m above ground level, as:
where DFSM i (μSv·hr−1) denotes the value obtained by converting the measured value at measurement point i to the air dose rate at 1 m above ground level. Call (cps) denotes the gamma-ray count rate, CD (μSv·hr−1·cps−1) denotes the air dose rate conversion factor for converting the gamma-ray count rates to air dose rates, AF (m−1) is an altitude-correction factor calculated from the relationship between the ground altitude and gamma-ray count rate, Hstd (m) represents the converted altitude (=1 m), and Hm (m) indicates the measured altitude above ground level. In this study, the following values were utilized: CD=4,431, AF=−0.01164, and CBG=173.
Finally, the data from the original detector and the converted target detector were applied to the Fukushima ANN, and the resulting outcomes were compared. The root mean squared error (RMSE) was employed for the comparative evaluation. The calculation formula for RMSE is presented in Equation (3):
where N is the total number of data points. In addition, histograms of the differences between the two values were created for comparative evaluation, and the dispersion of values was evaluated. Additionally, comparisons were made between the results obtained from the application of ANN and those from FSM to evaluate the differences in various conversion outcomes.
Results and Discussion
Fig. 6 shows the comparison results of count rates between the original detector (LaBr3) and the target detector (CZT) during calibration flights. The correlation coefficients between the count rates of the original and CZT detectors were 0.999 for the 50–450 keV range, 0.999 for the 450–900 keV range, and 0.865 for the 900–2,800 keV range. Strong correlations were observed in the data for the 50–450 keV and 450–900 keV ranges. After performing linear approximations for each count rate, the conversion coefficients a to transform the count levels of the CZT detector to the level of the original detector were determined to be 0.0398 and 0.0129 for the 50–450 keV and 450–900 keV ranges, respectively.
Fig. 7 illustrates the comparison results of count rates between the original detector (LaBr3) and the target detector (CeBr3) during calibration flights. The correlation coefficients between the count rates of the original and CeBr3 detectors were 1.000 for the 50–450 keV range, 1.000 for the 450–900 keV range, and 0.980 for the 900–2,800 keV range, indicating strong correlations in each case. After performing linear approximations for each count rate, the conversion coefficients a to transform the count levels of the target detector to the level of the original detector were determined to be 1.109, 1.516, and 2.460 for the 50–450 keV, 450–900 keV, and 900–2,800 keV ranges, respectively.
Fig. 8 depicts the ratio of count rates between the original detector and the target detector at various altitudes during calibration flights. For both CZT and CeBr3 detectors, the ratio between the two detectors remained constant across measurement altitudes, observed to be consistent within the 50–450 keV and 450–900 keV ranges, regardless of altitude. However, the ratio of 900–2,800 keV for the CZT detector in Fig. 8A was not constant. This result is attributed to the small size of the CZT crystal, resulting in very low count rates (0–1 cps) for the CZT detector in the 900–2,800 keV range. Therefore, in the ANN conversion for CZT, only the count rates in the 50–450 keV and 450–900 keV ranges were utilized as input variables, while the count rates in the 900–2,800 keV range were excluded. In Fig. 8B, a decreasing trend in the ratio of 900–2,800 keV was observed as the altitude increased for the CeBr3 detector. The CeBr3 detector used in this experiment has larger crystal size compared to the original detector, and it does not exhibit self-contamination due to its scintillator properties. Therefore, it was able to capture influences from various natural radiation sources better than LaBr3, resulting in higher count rates for CeBr3 and a decrease in the ratio. The spectrum in Fig. 2 represents the average spectrum of the measurement flights in the area shown in Fig. 5, where peaks from various natural radiation sources (K-40, Tl-208) are observed. Although there were slight fluctuations in the ratio, a strong correlation was confirmed in Fig. 7C. Thus, for the application of CeBr3, the count rates in the 900–2,800 keV range were also included as input variables.
Fig. 9 shows the results of converting the measurements obtained from the original detector and the CZT detector to air dose rates at 1 meter above ground level using the reconstructed F-ANN. Fig. 9C and 9F compare the conversion values obtained from each detector, while Fig. 9G and 9H compare the conversion values for each flight. Fig. 9I presents a histogram of the differences between the data from the CZT detector and the original detector. Fig. 10 shows the results of converting the measurements obtained from the original detector and the CZT detector using the FSM.
Fig. 11 shows the results of converting the measurements obtained from the original detector and the CeBr3 detector to air dose rates at 1 m above ground level using the Fukushima ANN. Fig. 11C compares the conversion values obtained from each detector, while Fig. 11D presents a histogram of the differences between the data from the CeBr3 detector and the original detector. Fig. 12 shows the results of converting the measurements obtained from the original detector and the CeBr3 detector using the FSM.
Table 2 shows the RMSE, median difference, and standard deviation of the difference calculated from the data between the CZT detector and the original detector. Table 3 shows the RMSE, median difference, and standard deviation of the difference calculated from the data between the CZT detector and the original detector for both flat flight and step flight. Table 4 shows the RMSE, median difference, and standard deviation of the difference calculated from the data between the CeBr3 detector and the original detector.
When examining the comparison between the CZT detector and the original detector, as depicted in Fig. 9 and the results in Table 2, it is evident that using conversion coefficients to apply to the ANN, even with data measured by the CZT detector, yields results that generally match those obtained when measured with the original detector. Additionally, as observed from Fig. 9 and the results in aTble 3, the conversion outcomes of the ANN demonstrate consistent results, indicating no significant variation in distribution among different flight methods.
However, upon examining the scatter plots in Fig. 9C and 9F, as well as the median values in Table 2, it is noticeable that the results from the CZT detector were slightly lower than those from the original detector. This discrepancy could potentially be attributed to the difference in crystal sizes between the detectors. Since CZT crystals were smaller than LaBr3 crystals, the CZT detector, as shown in Fig. 6, barely detected counts above 900 keV. This difference in detection efficiency might have influenced the slightly lower values obtained when using the CZT detector.
Comparing Fig. 9 with Fig. 10, it is evident that there were no significant differences in the distribution of results between the flat flight and step flight in the ANN conversion outcomes. However, notable differences were observed in the FSM conversion results. Referring to Table 3, the RMSE of the differences between flat flight and step flight in the ANN conversion results ranged from approximately 0.4 to 0.5, whereas in the FSM results, it was around 0.9, indicating a larger disparity. Additionally, the median values showed a noticeable positive bias in the FSM results, suggesting a higher tendency in the results of the flat flight. This could be attributed to the higher measurement altitude during flat flight, resulting in higher counts potentially due to the influence of scattered radiation.
Looking at the results comparing the CeBr3 detector with the original detector, as depicted in Fig. 11 and Table 4, it is evident that even data obtained from the CeBr3 detector can yield results that generally align with those obtained from the original detector when applied with conversion coefficients through ANN. However, upon examining the scatter plot in Fig. 11C, it is noticeable that the dispersion is larger compared to the CZT case shown in Fig. 9C and 9F. In Fig. 11C, the points highlighted in orange represent differences between the CeBr3 detector and the original detector of −1 or less.
Fig. 13 compares the counts rates of the CeBr3 detector and the original detector in the 900–2,800 keV range, based on the ARS measurements shown in Fig. 5. The count rates of the CeBr3 detector in Fig. 13 have already been converted to the level of the original detector. The points highlighted in orange in the scatter plot of Fig. 13 represent differences between the CeBr3 detector and the original detector of −1 or less, similar to Fig. 11C. Upon examination of Fig. 13, it is apparent that points highlighted in orange deviate significantly from the y=x line. This indicates that the dispersion observed in the comparison between the CeBr3 detector, and the original detector can be attributed to the count rates in the 900–2,800 keV range.
These results suggest that while slight differences may arise in the conversion outcomes of ANN due to variations in detector size and scintillation efficiency, the applicability of ANN has been validated through the use of conversion coefficients. Furthermore, inconsistencies were observed in the distribution patterns of FSM across different flight methods, whereas ANN consistently produced results regardless of the flight method employed. In this study, CZT and CeBr3 were adopted as the target detectors. Future evaluations using detectors from other institutions and different types of detectors will likely further enhance the versatility of the Fukushima ANN.
Conclusion
In this study, we demonstrated the potential utility of an ANN (F-ANN) trained on data from Fukushima’s experiences in converting ambient radiation monitoring system data to ground level 1 m ambient dose rates, even when obtained from target detectors different from those used to train the original detectors in F-ANN. By employing conversion coefficients to adjust the count levels of target detectors to those of original detectors, Fukushima ANN can be reconfigured for target detectors, enabling the conversion of measurement data from various detectors with the same precision as the original detectors. These results offer valuable insights into the potential applicability of the Fukushima ANN developed based on abundant monitoring data from Fukushima to various radiation measurements.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Print 
