26 June 2023 in Atmospheric Measurement Techniques
We demonstrate how machine learning can be easily applied to support the analysis of large quantities of excited hydroxyl (OH*) airglow imager data. We use a TCN (temporal convolutional network) classification algorithm to automatically pre-sort images into the three categories “dynamic” (images where small-scale motions like turbulence are likely to be found), “calm” (clear-sky images with weak airglow variations) and “cloudy” (cloudy images where no airglow analyses can be performed). The proposed approach is demonstrated using image data of FAIM 3 (Fast Airglow IMager), acquired at Oberpfaffenhofen, Germany, between 11 June 2019 and 25 February 2020, achieving a mean average precision of 0.82 in image classification. The attached video sequence demonstrates the classification abilities of the learned TCN. Within the dynamic category, we find a subset of 13 episodes of image series showing turbulence. As FAIM 3 exhibits a high spatial (23 m per pixel) and temporal (2.8 s per image) resolution, turbulence parameters can be derived to estimate the energy diffusion rate. Similarly to the results the authors found for another FAIM station (Sedlak et al., 2021), the values of the energy dissipation rate range from 0.03 to 3.18 W kg−1.
10 November 2022 in SN Computer Science
In this work, we propose a method that enforces explicit control over various attributes during the image generation process in a generative adversarial net. We propose a semi-supervised learning procedure that allows us to use a quantized approximation of object orientation for learning continuous object rotations. As a result, among many other attributes, our proposed method allows us to control object orientation in scenes that are rendered according to our specifications.