Presentation Information
[O12-P76]Creating an Automated Meteor Observation Programme Combining Machine Learning and Matrix Factorization
*Shinmura Atsuki1, *Sasaki Ryusei1, *Naito Rin1 (1. Tokyo Gakugei University Senior High School)
Keywords:
Meteor,Radio Wave Observation,Machine Learning,Non-negative Matrix Factorization
Quantifying the inflow flux of extraterrestrial organic matter through meteor observation holds significant academic importance for identifying the sources of prebiotic materials on early Earth and elucidating the mechanisms of the origin of life in astrobiology. Consequently, long-term meteor observations have been conducted globally. However, conventional radio meteor observation via spectrogram visualization suffers from inefficiency due to the necessity of manual data interpretation. To address this, this research aimed to develop software capable of automated meteor detection based on existing radio observation frameworks.
The methodology involved receiving radio waves transmitted from Otsu City, Shiga Prefecture, which were reflected by meteors. The conversion of these signals into audio enabled visualization through MROFFT. Subsequent analysis of recordings with and without meteor echoes using Discrete Fourier Transform (DFT) produced power spectral density (PSD) graphs. These graphs revealed that meteor occurrences caused the PSD to exceed -40 dB. However, direct implementation of this threshold resulted in false positives triggered by lightning and external noise.
To distinguish meteor-derived signals from non-meteor interference, supervised machine learning was employed for feature extraction. The analysis indicated that the Signal-to-Noise Ratio (SNR) is highly effective for classification, likely because meteor echoes concentrate energy within narrow frequency bands, whereas lightning and external noise disperse energy across broader spectra. Based on this finding, a detection algorithm using SNR thresholds was developed. Operations over a specific period confirmed that this SNR-based approach remains robust against noise variations as it does not depend on the receiver's gain settings.
Despite its high accuracy, the SNR-based method exhibited limited environmental adaptability due to its reliance on noise-predefined calculations. This necessitated the introduction of Non-negative Matrix Factorization (NMF), a mathematically grounded acoustic analysis method adaptable to diverse environments. NMF reconstructs original audio by decomposing it into a basis matrix (timbre) and an activation matrix (volume). The developed program integrates this method with machine learning through the following process:
A 30-second initial calibration to learn background noise and estimate the basis matrix. NMF processing of 1-second input segments to calculate activation matrices. Reconstruction of the input signal using the estimated basis and calculated activation matrices. Calculation and normalization of the reconstruction error via I-divergence (assuming a Poisson distribution) to define "Likelihood." Detection of meteors based on a 5σ Likelihood threshold. Continuous per-second updates of the background noise model to accommodate environmental changes.
A one-week automated trial yielded 55 detections, consisting of 25 true meteors and 30 false positives, with four missed events. These results were attributed to sudden fluctuations in background noise and external interference being identified as anomalies. To mitigate these issues, a hybrid approach combining NMF and SNR was proposed. Machine learning conducted on 120 datasets (60 meteor and 60 non-meteor) demonstrated that the inclusion of NMF features improved accuracy by approximately 4% compared to SNR-only models.
Current operations utilize this integrated NMF-SNR algorithm. Future research aims to quantitatively evaluate the program's precision by analyzing detection counts of true positives, false positives, and false negatives. Furthermore, replacing SNR with relative indices, such as Z-scores, is expected to enhance environmental adaptability and versatility. Subsequent developments will involve cross-referencing data with optical observations from security cameras and estimating meteor positions through radio wave direction-of-arrival estimation.
The methodology involved receiving radio waves transmitted from Otsu City, Shiga Prefecture, which were reflected by meteors. The conversion of these signals into audio enabled visualization through MROFFT. Subsequent analysis of recordings with and without meteor echoes using Discrete Fourier Transform (DFT) produced power spectral density (PSD) graphs. These graphs revealed that meteor occurrences caused the PSD to exceed -40 dB. However, direct implementation of this threshold resulted in false positives triggered by lightning and external noise.
To distinguish meteor-derived signals from non-meteor interference, supervised machine learning was employed for feature extraction. The analysis indicated that the Signal-to-Noise Ratio (SNR) is highly effective for classification, likely because meteor echoes concentrate energy within narrow frequency bands, whereas lightning and external noise disperse energy across broader spectra. Based on this finding, a detection algorithm using SNR thresholds was developed. Operations over a specific period confirmed that this SNR-based approach remains robust against noise variations as it does not depend on the receiver's gain settings.
Despite its high accuracy, the SNR-based method exhibited limited environmental adaptability due to its reliance on noise-predefined calculations. This necessitated the introduction of Non-negative Matrix Factorization (NMF), a mathematically grounded acoustic analysis method adaptable to diverse environments. NMF reconstructs original audio by decomposing it into a basis matrix (timbre) and an activation matrix (volume). The developed program integrates this method with machine learning through the following process:
A 30-second initial calibration to learn background noise and estimate the basis matrix. NMF processing of 1-second input segments to calculate activation matrices. Reconstruction of the input signal using the estimated basis and calculated activation matrices. Calculation and normalization of the reconstruction error via I-divergence (assuming a Poisson distribution) to define "Likelihood." Detection of meteors based on a 5σ Likelihood threshold. Continuous per-second updates of the background noise model to accommodate environmental changes.
A one-week automated trial yielded 55 detections, consisting of 25 true meteors and 30 false positives, with four missed events. These results were attributed to sudden fluctuations in background noise and external interference being identified as anomalies. To mitigate these issues, a hybrid approach combining NMF and SNR was proposed. Machine learning conducted on 120 datasets (60 meteor and 60 non-meteor) demonstrated that the inclusion of NMF features improved accuracy by approximately 4% compared to SNR-only models.
Current operations utilize this integrated NMF-SNR algorithm. Future research aims to quantitatively evaluate the program's precision by analyzing detection counts of true positives, false positives, and false negatives. Furthermore, replacing SNR with relative indices, such as Z-scores, is expected to enhance environmental adaptability and versatility. Subsequent developments will involve cross-referencing data with optical observations from security cameras and estimating meteor positions through radio wave direction-of-arrival estimation.