Advanced analysis techniques and deep learning for atmospheric measurements
Date issued
Authors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
License
Abstract
This work explores a wide range of data analysis and signal processing methods for
different possible applications in atmospheric measurements. While these methods
and applications span a wide area of disciplines, the evaluation of applicability and
limitations and the results of this evaluation show many similarities.
In the first study, a new framework for the temporal characterization of airborne
atmospheric measurement instruments is provided. Allan-Werle-plots are applied to
quantify dominant noise structures present in the time series. Their effects on the drift
correction capabilities and measurement uncertainty estimation can be evaluated via
simulation. This framework is applied to test flights of an airborne field campaign
and reveals an appropriate interval between calibration measurements of 30 minutes.
During ground operation, the drift correction is able to reduce the measurement uncertainty
from 1.1% to 0.2 %. Additional short-term disturbances during airborne
operation increase the measurement uncertainty to 1.5 %.
In the second study, the applicability and limitations of several noise reduction methods
are tested for different background characteristics. The increase in signal-noiseratio
and the added bias strongly depend on the background structure. Individual
regions of applicability show almost no overlap for the different noise reduction methods.
In the third study, a fast and versatile Bayesian method called sequential Monte Carlo
filter is explored for several applications in atmospheric field experiments. This algorithm
combines information provided via the measurements with prior information
from the dominant chemical reactions. Under most conditions the method shows
potential for precision enhancement, data coverage increase and extrapolation. Limitations
are observed that can be analyzed via the entropy measure and improvements
are achieved via the extension by an additional activity parameter.
In the final study, state-of-the-art neural network architectures and appropriate data
representations are used to reduce the effect of interference fringes in absorption spectroscopy.
Using the neural network models as an alternative to linear fitting yields
a large bias which renders the model approach not applicable. On the task of background
interpolation the neural network approach shows robust de-noising behavior
and is shown to be transferable to a different absorption spectrometer setup. Application
of the interpolation to the test set lowers the detection limit by 52%.
This work highlights the importance of in-depth analysis of the effects and limitations
of advanced data analysis techniques to prevent biases and data artifacts and to determine
the expected data quality improvements. An elaboration of the limitations
is particularly important for deep learning applications. All presented studies show
great potential for further applications in atmospheric measurements.