Lightning Signal Classification On A Low-Cost Community Detector Network: A Data Labelling And Machine Learning Pipeline Approach

Mahadhika, Muhammad Izzul Sinar (2026) Lightning Signal Classification On A Low-Cost Community Detector Network: A Data Labelling And Machine Learning Pipeline Approach. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221093-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (806kB) | Request a copy

Abstract

Lightning strikes pose significant risks to people and infrastructure, causing 5-10 deaths and up to 100 injuries annually in Australia alone. Large-scale lightning detection systems operated by national meteorological agencies rely on specialised, proprietary sensor networks that are too costly for community organisations to deploy. Low-cost, community-operated detector networks offer a more accessible alternative, but currently rely on simple peakdetection techniques that struggle to distinguish genuine lightning signals from background electrical noise, leaving much of collected waveform data underutilised. A key barrier to applying machine learning to this data is the absence of reliable ground-truth labels. This study first developed a labelling method that matched timestamps from storm logs verified by a subject matter expert (SME) against detector readings, with noise samples taken from quiet periods confirmed on a map-based visualisation. The resulting dataset of approximately 10,000 balanced samples was then used to develop and compare three classifiers: a feature-based model (XGBoost), a 1D convolutional neural network (CNN), and a hybrid CNN-XGBoost model. The probability gap, defined as the difference between the mean predicted probabilities of the lightning and noise classes, increased from 0.03-0.05 under GPS-second bucketing to 0.84-0.96 under the final labelling method. The hybrid model achieved the highest F1 score (0.991) and the lowest expected calibration error (ECE = 0.009) of the three models. These results were obtained on a balanced, randomly split dataset from a single network and should not be read as operational performance. The resulting classifier outputs confidence scores intended for use by downstream systems such as strike localisation and visualisation tools.
======================================================================================================================================
Sambaran petir menimbulkan risiko yang signifikan terhadap manusia dan infrastruktur, menyebabkan 5-10 kematian dan hingga 100 korban luka setiap tahunnya di Australia saja. Sistem deteksi petir skala besar yang dioperasikan badan meteorologi nasional mengandalkan jaringan sensor khusus dan proprietary yang terlalu mahal untuk diterapkan oleh organisasi komunitas. Jaringan detektor komunitas berbiaya rendah menawarkan alternatif yang lebih mudah diakses, tetapi masih mengandalkan deteksi puncak sederhana yang kesulitan membedakan sinyal petir asli dari derau elektrik latar belakang, sehingga data waveform selama 1 tahun kurang termanfaatkan. Hambatan utama penerapan machine learning pada data ini adalah tidak adanya label ground truth yang andal. Penelitian ini membangun metode pelabelan dengan mencocokkan timestamp log badai yang diverifikasi oleh subject matter expert (SME) terhadap pembacaan detektor, dengan sampel derau diambil dari periode tenang yang dikonfirmasi melalui visualisasi peta. Dataset hasilnya (sekitar 10.000 sampel seimbang) kemudian dipakai untuk mengembangkan dan membandingkan tiga classifier: XGBoost, CNN 1D, dan model hybrid CNN-XGBoost. Gap probabilitas, yaitu selisih rata-rata probabilitas prediksi kelas petir dan kelas derau, meningkat dari 0,03-0,05 pada pelabelan berbasis GPSsecond bucketing menjadi 0,84-0,96 setelah metode pelabelan final diterapkan. Model hybrid mencapai F1 score tertinggi (0,991) dan expected calibration error terendah (ECE = 0,009) di antara ketiga model. Hasil ini diperoleh pada data seimbang dari satu jaringan dengan pembagian acak sehingga belum mencerminkan kinerja operasional. Classifier yang dihasilkan memberikan skor kepercayaan yang dimaksudkan untuk dipakai oleh sistem downstream, seperti solver lokalisasi dan alat visualisasi.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Data Labelling, Lightning Classification, Machine Learning, Model Calibration, Signal Classification
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	MUHAMMAD IZZUL SINAR MAHADHIKA
Date Deposited:	05 Nov 2026 03:34
Last Modified:	05 Nov 2026 03:34
URI:	http://repository.its.ac.id/id/eprint/146031

Actions (login required)

View Item