3:30 PM - 5:00 PM
[O08-P56] Development of the analysis- and detection-system for meteor appearance signals using deep learning
Keywords:Radio meteor observation, Python, Deep learning, Arduino
Introduction
The radio meteor observation is a worldwide method, using a receiver and antenna designed to receive radio waves of a specific frequency. This method is based on the characteristic that radio waves in a certain frequency band transmitted from the ground are partially reflected when a meteor traverses the atmosphere. However, it requires high costs for expensive devices and observation system construction. Therefore, we have developed a “meteor appearance notification system,” which automatically observes meteors and quickly broadcasts meteor appearance information so that anyone can easily observe meteor showers. In our previous study and development, we have completely automated the process from detecting meteor appearance signals to broadcasting meteor appearance information. Furthermore, it has also allowed us to collect quantitative data. However, our conventional detecting system still has issues that it does not work on weak signals, and it causes false detection of noise as meteors. This study aimed to detect meteor appearance signals with higher accuracy by using deep learning to analyze sound data obtained by radio meteor observation.
Method
Our system consists of three sections: 1) the radio meteor observation section using a 2-element antenna and radio wave receiver, 2) the analysis section using Arduino and Python, and 3) the bot-transmitting section using Python. In this study, we newly introduced the deep learning model (by PyTorch) into the analysis section.
Deep learning was performed on the sound data obtained by observing sporadic meteors in the following steps: 1) We divided the continuous sound data into data per second. 2) We extracted parts of sound data to reduce the number of processes. The extraction threshold of sound volume and frequency were > –30.29 dB (relative to the computer’s maximum sound volume) and 495 Hz, respectively. These thresholds resulted from our previous study and the automatic program that outputs sound values matching the conditions. The processed sound data consequently contained a large amount of meteor appearance signals and noise. 3) We converted the extracted sound data into spectrograms. Each was output as a single image file (n > 300). 4) We assigned a data label to each image, manually identifying the meteor appearance signal and the noise. Then, the dataset was used for deep learning. 5) We randomly selected some of the images and labels (see step 4) to be used for discrimination and excluded them from the data for deep learning. For the discrimination, we compared the data labels with the input data. When the pair was identical, it was marked as “correct”; otherwise, it was marked as “incorrect.” 6) We made the deep learning model learn about image recognition using the remaining images and labels that were not excluded in step 5.
We carried out three experiments (I, II, III) using the developed deep learning model; I) We identified whether the recorded sound data (obtained by sporadic meteors observation) were meteor appearance signals or noise using the deep learning model. We then calculate the percentage of correct answers (i.e., accuracy), comparing the identified results with the assigned labels. II) In order to make our system possible to detect weak meteor signals, we used all sound data, omitted step 2 (partial extraction of the sound data), and calculated accuracy. III) We used the sound data obtained by observing sporadic meteors at a different time from Experiments I and II, and then calculated accuracy.
Results and Discussion
In Experiment I, we validate a novel deep learning method to detect meteor appearance signals. The accuracy was 95.57%. This result suggests that our system would accurately distinguish meteor appearance signals and noise. In Experiment II, we developed a system to identify the meteor appearance signals from all sound data with seconds resolution. The accuracy was 95.57%, as in Experiment I. This result implies that our system has the ability to detect weak meteor signals. In Experiment III, we tested the applicability of our new system using the data different from which we constructed the deep learning program. The accuracy was 94.97%. Therefore, our detecting system would be practical and has sufficient accuracy to work with various meteor appearance signals and noise patterns.
Conclusion
We successfully introduced a deep learning model into radio meteor observation. Our novel analysis- and detection-system made it easier to detect weak meteor appearance signals. Moreover, it accurately processes noisy sound data, which led false detection. Our next goal is to apply our system to forecast and broadcast the appearance peaks of meteor showers.
The radio meteor observation is a worldwide method, using a receiver and antenna designed to receive radio waves of a specific frequency. This method is based on the characteristic that radio waves in a certain frequency band transmitted from the ground are partially reflected when a meteor traverses the atmosphere. However, it requires high costs for expensive devices and observation system construction. Therefore, we have developed a “meteor appearance notification system,” which automatically observes meteors and quickly broadcasts meteor appearance information so that anyone can easily observe meteor showers. In our previous study and development, we have completely automated the process from detecting meteor appearance signals to broadcasting meteor appearance information. Furthermore, it has also allowed us to collect quantitative data. However, our conventional detecting system still has issues that it does not work on weak signals, and it causes false detection of noise as meteors. This study aimed to detect meteor appearance signals with higher accuracy by using deep learning to analyze sound data obtained by radio meteor observation.
Method
Our system consists of three sections: 1) the radio meteor observation section using a 2-element antenna and radio wave receiver, 2) the analysis section using Arduino and Python, and 3) the bot-transmitting section using Python. In this study, we newly introduced the deep learning model (by PyTorch) into the analysis section.
Deep learning was performed on the sound data obtained by observing sporadic meteors in the following steps: 1) We divided the continuous sound data into data per second. 2) We extracted parts of sound data to reduce the number of processes. The extraction threshold of sound volume and frequency were > –30.29 dB (relative to the computer’s maximum sound volume) and 495 Hz, respectively. These thresholds resulted from our previous study and the automatic program that outputs sound values matching the conditions. The processed sound data consequently contained a large amount of meteor appearance signals and noise. 3) We converted the extracted sound data into spectrograms. Each was output as a single image file (n > 300). 4) We assigned a data label to each image, manually identifying the meteor appearance signal and the noise. Then, the dataset was used for deep learning. 5) We randomly selected some of the images and labels (see step 4) to be used for discrimination and excluded them from the data for deep learning. For the discrimination, we compared the data labels with the input data. When the pair was identical, it was marked as “correct”; otherwise, it was marked as “incorrect.” 6) We made the deep learning model learn about image recognition using the remaining images and labels that were not excluded in step 5.
We carried out three experiments (I, II, III) using the developed deep learning model; I) We identified whether the recorded sound data (obtained by sporadic meteors observation) were meteor appearance signals or noise using the deep learning model. We then calculate the percentage of correct answers (i.e., accuracy), comparing the identified results with the assigned labels. II) In order to make our system possible to detect weak meteor signals, we used all sound data, omitted step 2 (partial extraction of the sound data), and calculated accuracy. III) We used the sound data obtained by observing sporadic meteors at a different time from Experiments I and II, and then calculated accuracy.
Results and Discussion
In Experiment I, we validate a novel deep learning method to detect meteor appearance signals. The accuracy was 95.57%. This result suggests that our system would accurately distinguish meteor appearance signals and noise. In Experiment II, we developed a system to identify the meteor appearance signals from all sound data with seconds resolution. The accuracy was 95.57%, as in Experiment I. This result implies that our system has the ability to detect weak meteor signals. In Experiment III, we tested the applicability of our new system using the data different from which we constructed the deep learning program. The accuracy was 94.97%. Therefore, our detecting system would be practical and has sufficient accuracy to work with various meteor appearance signals and noise patterns.
Conclusion
We successfully introduced a deep learning model into radio meteor observation. Our novel analysis- and detection-system made it easier to detect weak meteor appearance signals. Moreover, it accurately processes noisy sound data, which led false detection. Our next goal is to apply our system to forecast and broadcast the appearance peaks of meteor showers.