TY - JOUR
T1 - Comparative analysis of deep learning and traditional airborne electromagnetic data processing: A case study
AU - Asif, Muhammad Rizwan
AU - Kass, M Andy
AU - Herpe, Maiwenn
AU - Rawlinson, Zara
AU - Westerhoff, Rogier
AU - Larsen, Jakob Juul
AU - Christiansen, Anders Vest
PY - 2025
Y1 - 2025
N2 - State-of-the-art airborne electromagnetic (AEM) systems conduct large-scale surveys and produce thousands of line kilometers of data. These large volumes of AEM data are generally subjected to manual inspection for identifying and culling data corrupted by couplings from various sources, such as fences and power lines, as well as the assessment for late-time noise. Although effective, this traditional AEM data processing is subjective and time consuming. This study evaluates the efficacy of a recently proposed unsupervised deep-learning method for the automated processing of AEM data on a large-scale survey conducted in New Zealand. The automated system processes 6471 line kilometers of data in approximately 15 min — a task that could extend more than 5–16 days if done manually by a trained expert. Although manual processing is not the ultimate benchmark, both processing schemes have an agreement of approximately 88% for the low-moment data and approximately 96% for the high-moment data. The inversion results for manual and automated system processing also indicate comparable data fits and resistivity models. The automated system demonstrates strong generalization capabilities; however, it faces challenges when data affected by various sources, such as anthropogenic activities, mimic credible 1D responses. This underscores the need for manual inspection in contexts wherein it is necessary to distinguish between true anomalies and noise. This study not only validates the automated system’s applicability across diverse geologic settings but also emphasizes the importance of balancing automated and manual approaches for optimal AEM data processing.
AB - State-of-the-art airborne electromagnetic (AEM) systems conduct large-scale surveys and produce thousands of line kilometers of data. These large volumes of AEM data are generally subjected to manual inspection for identifying and culling data corrupted by couplings from various sources, such as fences and power lines, as well as the assessment for late-time noise. Although effective, this traditional AEM data processing is subjective and time consuming. This study evaluates the efficacy of a recently proposed unsupervised deep-learning method for the automated processing of AEM data on a large-scale survey conducted in New Zealand. The automated system processes 6471 line kilometers of data in approximately 15 min — a task that could extend more than 5–16 days if done manually by a trained expert. Although manual processing is not the ultimate benchmark, both processing schemes have an agreement of approximately 88% for the low-moment data and approximately 96% for the high-moment data. The inversion results for manual and automated system processing also indicate comparable data fits and resistivity models. The automated system demonstrates strong generalization capabilities; however, it faces challenges when data affected by various sources, such as anthropogenic activities, mimic credible 1D responses. This underscores the need for manual inspection in contexts wherein it is necessary to distinguish between true anomalies and noise. This study not only validates the automated system’s applicability across diverse geologic settings but also emphasizes the importance of balancing automated and manual approaches for optimal AEM data processing.
U2 - 10.1190/geo2024-0282.1
DO - 10.1190/geo2024-0282.1
M3 - Journal article
SN - 0016-8033
VL - 90
SP - WA103-WA112
JO - Geophysics
JF - Geophysics
IS - 3
ER -