DeepMAD: Deep Learning for Magnetic Anomaly Detection and Denoising

Abstract

In this paper we introduce an end-to-end deep learning (DL) framework for magnetic anomaly detection (MAD) and denoising. This framework consists of two neural networks: a binary classification network for magnetic anomaly detection and a regression network for geomagnetic noise suppression. The two networks work in a cascade mode: the magnetic field measurement is first sent to the detection network to check the existence of the anomaly signal, and then to the denoising network for extracting the signal from the geomagnetic noise if the detection result is positive. The core idea of our proposed method is that the characteristics of both the magnetic anomaly signal and the geomagnetic noise can be learned from massive training data. The experimental results show that: (1) under the same false alarm rate constraint, the probability of detection of our proposed method is above 80% when the signal-to-noise ratio (SNR) equals -6 dB, while the orthogonal basis function (OBF) method fails when the SNR is below 0 dB; (2) for geomagnetic noise suppression, an improvement of 10 to 15 dB is achieved for data with input SNRs between -5 and 15 dB. Our results paved the way for data-driven magnetic anomaly detection and denoising.