CrowdMag

Deviations in local magnetic field strengths from what is expected are known as magnetic anomalies, whose proper detection can reveal various geophysical features and aid in resource exploration. NOAA's current magnetic anomaly map is known as EMAG2_v3, which is a compilation of data from airborne, shipborne, and satellite measurements. However, the resolution of this map is limited by the spacing between these measurements. To improve its magnetic reference models and maps, NOAA introduced CrowdMag, a citizen science project that harnesses data contributed by the public through a mobile app that utilizes the magnetometers embedded in modern smartphones. But because the magnetometers in smartphones are lower quality and therefore influenced by surrounding magnetic interference, the collected data is extremely noisy. To overcome this challenge, a type of deep neural network known as an autoencoder (AE) is employed to denoise the data. An AE learns to reconstruct its input data by effectively compressing the information into a lower-dimensional representation and then reconstructing it to remove noise. This artificial intelligence (AI) technique is trained using intentionally noised and masked EMAG2_v3 data (Figure 1b) and its hyperparameters optimally tuned before being applied to the noisy CrowdMag data. This tuning includes but is not limited to, altering input size, epoch size, loss function, masking and randomization level, and compression ratio. The AE's performance is measured by visual comparison of the original EMAG2_v3 data to its reconstruction (Figure 1a/1c) and the optimal hyperparameters are established. Once the reconstruction of EMAG2_v3 is sufficient, the AE is applied to the noisy CrowdMag data, and an AI-ready calibrated and denoised dataset of magnetic anomalies is produced. To the best of our knowledge, this is the first time autoencoders have been applied to this sector of geoscience. The outcomes of this project can be extended to other fields of science where noisy datasets are prevalent. Additionally, this initiative promotes the utilization of citizen-science-collected measurements, fostering greater public involvement in scientific research.

In 2014, we started the "CrowdMag" Project to collect vector magnetic data from digital magnetometers in smartphones. In October 2014, we released our first-generation Android and iOS apps. Currently, the CrowdMag Project has more than 15,000 enthusiastic users contributing more than 12 million magnetic data points from around the world. NOAA's National Centers for Environmental Information (NCEI), in partnership with the University of Colorado's Cooperative Institute for Research in the Environmental Sciences (CIRES) develops magnetic field models to aid navigation, resource exploration and scientific research. We use observatories, satellites and ship/airborne surveys to measure the magnetic data. However, the measurements leave gaps in coverage, particularly for short-wavelength urban noise. Our ultimate goal is to use data from the CrowdMag Project to improve global magnetic data coverage. Here we present some early results from the analysis of the crowdsourced magnetic data. A global magnetic model derived solely based on CrowdMag data is generally consistent with satellite-derived models such as World Magnetic Model. A unique contribution of the CrowdMag Project is the collection of ground level magnetic data in densely populated regions with an unprecedented spatial resolution. For example, we show a magnetic map (by binning the data collected into 100x100m cells) of central Boulder using 170,000 data points collected by about 60 devices over the duration October 2014- January 2016. The median magnetic field value is consistent with the expected magnitude of the Earth's background magnetic field. The standard deviation of the CrowdMag total field (F) values is much higher than the expected natural (i.e., diurnal and geologic) magnetic field variation. However, the phone's magnetometer is sensitive enough to capture the larger magnitude magnetic signature from the urban magnetic sources. We discuss the reliability of crowdsourced magnetic maps and their applications to navigation, global models, and local geologic or environmental investigations.