Laura Crocetti


Last Name

Crocetti

First Name

Laura

ORCID

0000-0003-2538-4111

Organisational unit

09707 - Soja, Benedikt / Soja, Benedikt

Filters

2020 - 202662
Reset filters

Search Results

Publications1 - 10 of 62
  • Offset detection from GNSS time series data: comparing an (unsupervised) augmented Kalman filter approach and supervised machine learning methods
    Item type: Other Conference Item
    Hohensinn, Roland; Crocetti, Laura; Ren, Elizabeth; et al. (2023)
    AGU Fall Meeting Abstracts
  • Machine learning-based prediction of Alpine foehn events using GNSS troposphere products: first results for Altdorf, Switzerland
    Item type: Journal Article
    Aichinger-Rosenberger, Matthias; Brockmann, Elmar; Crocetti, Laura; et al. (2022)
    Atmospheric Measurement Techniques
    Remote sensing of water vapour using the Global Navigation Satellite System (GNSS) is a well-established technique and reliable data source for numerical weather prediction (NWP). However, one of the phenomena rarely studied using GNSS are foehn winds. Since foehn winds are associated with significant humidity gradients between two sides of a mountain range, tropospheric estimates from GNSS are also affected by their occurrence. Time series reveal characteristic features like distinctive minima and maxima as well as a significant decrease in the correlation between the stations. However, detecting such signals becomes increasingly difficult for large datasets. Therefore, we suggest the application of machine learning algorithms for the detection and prediction of foehn events by means of GNSS troposphere products. This initial study develops a new, machine learning-based method for detection and prediction of foehn events at the Swiss station Altdorf by utilising long-term time series of high-quality GNSS troposphere products. Data from the Automated GNSS Network Switzerland (AGNES) and various GNSS sites from neighbouring countries as well as records of an operational foehn index are used to investigate the performance of several different classification algorithms based on appropriate statistical metrics. The two best-performing algorithms are fine tuned and tested in four dedicated experiments using different feature setups. The results are promising, especially when reprocessed GNSS products are utilised and the most dense station setup is used. Detection- and alarm-based measures reach levels between 66 %-80 % for both tested algorithms and thus are comparable to those from studies using data from meteorological stations and NWP. For operational prediction, limitations due to the availability and quality of GNSS products in near-real time (NRT) exist. However, they might be mitigated to a significant extent by provision of additional NRT products and improved data processing in the future. Results also outline benefits for the results when including geographically relevant stations (e.g. high-altitude stations) in the utilised datasets.
  • Determination of high-precision tropospheric delays using crowdsourced smartphone GNSS data
    Item type: Journal Article
    Pan, Yuanxin; Kłopotek, Grzegorz; Crocetti, Laura; et al. (2024)
    Atmospheric Measurement Techniques
    The Global Navigation Satellite System (GNSS) is a key asset for tropospheric monitoring. Currently, GNSS meteorology relies primarily on geodetic-grade stations. However, such stations are too costly to be densely deployed, which limits the contribution of GNSS to tropospheric monitoring. In 2016, Google released the raw GNSS measurement application programming interface for smartphones running on Android version 7.0 and higher. Given that nowadays there are billions of Android smartphones worldwide, utilizing those devices for atmospheric monitoring represents a remarkable scientific opportunity. In this study, smartphone GNSS data collected in Germany as part of the Application of Machine Learning Technology for GNSS IoT Data Fusion (CAMALIOT) crowdsourcing campaign in 2022 were utilized to investigate this idea. Approximately 20 000 raw GNSS observation files were collected there during the campaign. First, a dedicated data processing pipeline was established that consists of two major parts: machine learning (ML)-based data selection and ionosphere-free precise point positioning (PPP)-based zenith total delay (ZTD) estimation. The proposed method was validated with a dedicated smartphone data collection experiment conducted on the rooftop of the ETH campus. The results confirmed that ZTD estimates of millimeter-level precision could be achieved with smartphone data collected in an open-sky environment. The impacts of observation time span and utilization of multi-GNSS observations on ZTD estimation were also investigated. Subsequently, the crowdsourced data from Germany were processed by PPP with the ionospheric delays interpolated using observations from surrounding satellite positioning service of the German National Survey (SAPOS) GNSS stations. The ZTDs derived from ERA5 and an ML-based ZTD product served as benchmarks. The results revealed that an accuracy of better than 10 mm can be achieved by utilizing selected high-quality crowdsourced smartphone data. This study demonstrates high-precision ZTD determination with crowdsourced smartphone GNSS data and reveals success factors and current limitations.
  • Atmospheric Delay Modeling and Uncertainty Quantification with Probabilistic Neural Networks
    Item type: Conference Poster
    Soja, Benedikt; Hadas, Tomasz; Orus Perez, Raul; et al. (2025)
    Accurate modeling of atmospheric delays is crucial for high-precision Global Navigation Satellite System (GNSS) applications. In the ESA-funded project “Probabilistic Neural Network Models for Accurate Atmospheric Modeling”, we utilize Probabilistic Neural Networks (PNNs) to estimate tropospheric and ionospheric delays together with realistic uncertainty measures. The uncertainties, often neglected in machine learning applications, allow us to improve the weighting of observations in geodetic parameter estimation and increase confidence in the results. The data-driven framework is particularly suitable to capture nonlinear atmospheric behavior that is difficult to describe with conventional models. Two strategies are explored: a general-purpose “blind” model using only time, location, and geometry, and an enhanced model that incorporates auxiliary data for demanding use cases such as the Galileo High Accuracy Service (HAS). Both are tested with synthetic and real-world GNSS data to evaluate their impact on Position, Velocity, and Time (PVT) solutions. In this presentation, we present initial results, showing that PNN-based estimates of zenith tropospheric delay and total electron content, along with their calibrated uncertainties, have the potential to improve positioning accuracy and convergence.
  • Mapping GNSS radio occultation climatologies using machine learning
    Item type: Other Conference Item
    Shehaj, Endrit; Leroy, Stephen; Cahoy, Kerri; et al. (2023)
    Abstract Volume 21st Swiss Geoscience Meeting
  • A Global Machine Learning-based Zenith Wet Delay Forecasting Model
    Item type: Conference Poster
    Crocetti, Laura; Schartner, Matthias; Schindler, Konrad; et al. (2023)
  • A Cloud-native Approach for Processing of Crowdsourced GNSS Observations and Machine Learning at Scale: A Case Study from the CAMALIOT Project
    Item type: Journal Article
    Kłopotek, Grzegorz; Pan, Yuanxin; Sturn, Tobias; et al. (2024)
    Advances in Space Research
    The era of modern smartphones, running on Android version 7.0 and higher, facilitates nowadays acquisition of raw dual-frequency multi-constellation GNSS observations. This paves the way for GNSS community data to be potentially exploited for precise positioning, GNSS reflectometry or geoscience applications at large. The continuously expanding global GNSS infrastructure along with the enormous volume of prospective GNSS community data bring, however, major challenges related to data acquisition, its storage, and subsequent processing for deriving various parameters of interest. In addition, such large datasets cannot be managed manually anymore, leading thus to the need for fully automated and sophisticated data processing pipelines. Application of Machine Learning Technology for GNSS IoT data fusion (CAMALIOT) was an ESA NAVISP Element 1 project (NAVISP-EL1-038.2) with activities aiming to address the aforementioned points related to GNSS community data and their exploitation for scientific applications with the use of Machine Learning (ML). This contribution provides an overview of the CAMALIOT project with information on the designed and implemented cloud-native software for GNSS processing and ML at scale, developed Android application for retrieving GNSS observations from the modern generation of smartphones through dedicated crowdsourcing campaigns, related data ingestion and processing, and GNSS analysis concerning both conventional and smartphone GNSS observations. With the use of the developed GNSS engine employing an Extended Kalman Filter, example processing results related to the Zenith Total Delay (ZTD) and Slant Total Electron Content (STEC) are provided based on the analysis of observations collected with geodetic-grade GNSS receivers and from local measurement sessions involving Xiaomi Mi 8 that collected GNSS observations using the developed Android application. For smartphone observations, ZTD is derived in a differential manner based on a single-frequency double-difference approach employing GPS and Galileo observations, whereas satellite-specific STEC time series are obtained through carrier-to-code leveling based on the geometry-free linear combination of GPS and Galileo observations. Although the ZTD and STEC time series from smartphones were derived on a demonstration basis, a rather good level of consistency of such estimates with respect to the reference time series was found. For the considered periods, the RMS of differences between the derived smartphone-based time series of differential zenith wet delay and reference values were below 3.1 mm. In terms of satellite-specific STEC time series expressed with respect to the reference STEC time series, RMS of the offset-reduced differences below 1.2 TECU was found. Smartphone-based observations require special attention including additional processing steps and a dedicated parameterization in order to be able to acquire reliable atmospheric estimates. Although with lower measurement quality compared to traditional sources of GNSS data, an augmentation of ground-based networks of fixed high-end GNSS receivers with GNSS-capable smartphones would however, form an interesting source of complementary information for various studies relying on GNSS observations.
  • Land motion in Europe imaged by GNSS
    Item type: Other Conference Item
    Hübner, Laura; Brockmann, Elmar; Crocetti, Laura; et al. (2023)
    EGUsphere
    The densification of high-quality, permanent GNSS stations in Europe enables a large-scale investigation of deformation processes on the Earth’s surface. This work aims to interpolate the horizontal and vertical GNSS station velocities and thus produce velocity fields showing the land motion for Switzerland, the Alps and Europe. The GNSS station velocities are provided by the EUREF Working Group on European Dense Velocities. The data set contains horizontal (east, north) and vertical velocities for around 8000 stations in Europe. Five interpolation methods are implemented and compared, namely, Inverse Distance Weighting (IDW), Ordinary Kriging, K - Nearest Neighbors (KNN), Random Forest and Multilayer Perceptron (MLP). Latitude and longitude of the station locations are used as input features for the interpolation. Additional input features will be engineered for Random Forest and MLP. Generally, the performance of all five interpolation methods with latitude and longitude as features evaluated on the test data by Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Bias Error (MBE) is comparable for all velocity components in Switzerland, the Alps and Europe. RMSE and MAE vary among the methods by hundredths of a mm/year. The only exceptions are the horizontal velocity components for the extent of Europe, where MLP and Ordinary Kriging perform slightly worse than the other methods. The key findings of the qualitative analysis are that MLP and Ordinary Kriging produce the smoothest velocity fields, while IDW, KNN and Random Forest produce artifacts due to their mode of operation. All methods interpolate similar velocity fields where the station data is dense and greater differences when it is sparse. Especially for extrapolation areas where no data is available their performance is not verified. The interpolation of the GNSS station velocities in Switzerland, the Alps and Europe for this work shows that it is possible to produce velocity fields with accuracy level below 1 mm/year and the different phenomena of land motion can be clearly identified.
  • Detection of Heavy Precipitation using GNSS Zenith Wet Delay
    Item type: Other Conference Item
    Saini, Dev; Crocetti, Laura; Soja, Benedikt (2025)
Publications1 - 10 of 62