IceCloudNet: 3D reconstruction of cloud ice from Meteosat SEVIRI

TL;DR

The produced dataset increases the availability of vertical cloud profiles, for the period when DARDAR is available, by more than six orders of magnitude and moreover, IceCloudNet is able to produce vertical cloud profiles beyond the lifetime of the recently ended satellite missions underlying DARDAR.

Abstract

IceCloudNet is a novel method based on machine learning able to predict high-quality vertically resolved cloud ice water contents (IWC) and ice crystal number concentrations (N$_\textrm{ice}$). The predictions come at the spatio-temporal coverage and resolution of geostationary satellite observations (SEVIRI) and the vertical resolution of active satellite retrievals (DARDAR). IceCloudNet consists of a ConvNeXt-based U-Net and a 3D PatchGAN discriminator model and is trained by predicting DARDAR profiles from co-located SEVIRI images. Despite the sparse availability of DARDAR data due to its narrow overpass, IceCloudNet is able to predict cloud occurrence, spatial structure, and microphysical properties with high precision. The model has been applied to ten years of SEVIRI data, producing a dataset of vertically resolved IWC and N$_\textrm{ice}$ of clouds containing ice with a 3 kmx3 kmx240 mx15 minute resolution in a spatial domain of 30°W to 30°E and 30°S to 30°N. The produced dataset increases the availability of vertical cloud profiles, for the period when DARDAR is available, by more than six orders of magnitude and moreover, IceCloudNet is able to produce vertical cloud profiles beyond the lifetime of the recently ended satellite missions underlying DARDAR.

Figures (11)

• Figure 1: Conceptual visualization of the idea behind IceCloudNet. Panel (i) shows a sample of Meteosat SEVIRI and panel (ii) the vertical iwc profile along the co-located DARDAR satellite overpass. The strengths and weaknesses of both datasets are indicated by the icons next to the panels. Through the use of ml a novel synthetic dataset is created (iii), which combines the spatio-temporal resolution and coverage of SEVIRI with the vertical resolution of DARDAR. The shown data is from 2006-08-22 13:12:40 in a spatial domain ranging from 3°W to 7°E, 9°N to 19°N.
• Figure 2: 10 µ m channel for single SEVIRI observation from 2007-10-24 00:12:43 with co-located DARDAR IWP. The DARDAR swath is magnified for better visibility.
• Figure 3: A visualization of the IceCloudNet model architecture, including sample inputs and predictions, as well as the flow of losses during optimization. Abbreviations and acronyms used in the figure: Vis. - Visible; MLP - Multi Layer Perceptron; Conv2d - 2d convolution; ReLU - Rectified Linear Unit; $\mathcal{L}_{D}$ - Discriminator Loss; $\mathcal{L}_{R}$ - Regression loss; $\mathcal{L}_{G}$ - Loss of the main model generating cloud profiles.
• Figure 4: Regression and classification metrics calculated per pixel of all (A), daytime (B), nighttime (C) DARDAR overpasses for different height levels for 2010. Each height level in (A),(B), and (C) consists of three height levels in the original resolution (240 m). The classification metrics are calculated on a post-processed cloud mask and $R^2$ is calculated on logarithmically transformed iwc and nice. All day (D)/ daytime (E) / nighttime (F) cloud occurrence for different height levels for DARDAR overpasses (red), and IceCloudNet predictions along DARDAR overpasses (blue)
• Figure 5: Cloud cover along DARDAR overpasses: (A) Frequency of cloud cover along DARDAR overpass profiles for DARDAR reference data and IceCloudNet predictions. A single data point here represents the cloud cover along the DARDAR overpass in a single patch ( 256$\times$256 pixels ) as used in the training of IceCloudNet. The red line shows the linear regression fitted between the predictions and ground truth data. The marginal distributions of cloud cover frequencies are shown on the right and top, respectively. Cloud cover is calculated such that an atmospheric column is cloudy as soon as there is a cloudy pixel at any height, for example, if all columns within an overpass profile contain at least one cloudy pixel, the cloud cover is considered 100 %. (B): The coefficient of determination ($R^2$) calculated for different height levels between DARDAR and IceCloudNet cloud cover for all overpass profiles of the test dataset. A height level bin in this figure corresponds to 3 levels of the actual data, that is, 720 m.