IceCloudNet: Cirrus and mixed-phase cloud prediction from SEVIRI input learned from sparse supervision (2310.03499v1)

Abstract: Clouds containing ice particles play a crucial role in the climate system. Yet they remain a source of great uncertainty in climate models and future climate projections. In this work, we create a new observational constraint of regime-dependent ice microphysical properties at the spatio-temporal coverage of geostationary satellite instruments and the quality of active satellite retrievals. We achieve this by training a convolutional neural network on three years of SEVIRI and DARDAR data sets. This work will enable novel research to improve ice cloud process understanding and hence, reduce uncertainties in a changing climate and help assess geoengineering methods for cirrus clouds.

• The paper introduces IceCloudNet, a CNN using sparse DARDAR data to predict ice cloud properties (IWP) from readily available SEVIRI satellite input.
• IceCloudNet demonstrates superior performance compared to linear regression and XGBoost baselines, achieving lower mean absolute errors (0.49 for cirrus and 0.47 for mixed-phase predictions).
• This novel approach provides high spatio-temporal resolution predictions of ice clouds, offering valuable observational constraints to improve global climate models and enhance understanding of cloud processes.

Analysis of "IceCloudNet: Cirrus and mixed-phase cloud prediction from SEVIRI input learned from sparse supervision"

The paper "IceCloudNet: Cirrus and mixed-phase cloud prediction from SEVIRI input learned from sparse supervision" presents a novel approach to predict regime-dependent ice cloud properties using convolutional neural networks (CNNs). The authors address a significant challenge in climate science: the accurate prediction and understanding of ice cloud microphysics, which has critical implications for climate models and future climate projections.

The paper leverages the interplay of geostationary satellite data from SEVIRI and the vertically-resolved data from DARDAR. The key innovation is the use of the IceCloudNet, a CNN based on a U-Net architecture enhanced by ConvNeXt blocks, to predict the ice water path (IWP) for both cirrus and mixed-phase clouds. This approach offers a high spatio-temporal prediction capability that is not afforded by traditional polar-orbiting active satellite instruments due to their limited temporal coverage.

Dataset and Methodology

The dataset integrates three years of SEVIRI's multi-spectral images with the DARDAR dataset, which in turn combines CALIPSO and CloudSat retrievals. Notably, the integration results in a sparse data problem because the DARDAR overpass swath is narrow in comparison to SEVIRI's imaging capability. To counter this sparsity, the authors resampled DARDAR data to match SEVIRI’s horizontal resolution and created a dataset of image patches that contain co-located DARDAR data, allowing for the sparse supervision of the CNN.

The U-Net architecture used in IceCloudNet is adept at handling the spatial relations inherent in the input data, a factor key to predicting microphysical properties across both cirrus and mixed-phase regions. The training of IceCloudNet involved the transformation of outputs using logarithmic scaling and an augmented dataset with rotated image patches to improve model robustness.

Results and Implications

IceCloudNet demonstrated superior performance over linear regression and XGBoost baselines, as evidenced by lower mean absolute error and higher Pearson correlation and accuracy metrics. Specifically, IceCloudNet achieves a mean absolute error of 0.49 and 0.47 for cirrus and mixed-phase predictions, respectively, significantly outperforming baseline methods.

The research outcome has substantial implications. Practically, the ability to predict IWP with high spatio-temporal resolution extends beyond the limited data provided by traditional methods, enabling researchers to monitor and paper cloud systems dynamically. Theoretically, this work provides a critical observational constraint that will aid in understanding cloud formation processes, improving global climate model predictions, and potentially influencing the assessment of geoengineering interventions targeting cirrus clouds.

Future Developments

Looking ahead, extending IceCloudNet could involve addressing its limitations in handling smaller cloud formations and enhancing generalization further through larger and more diverse datasets. Additionally, incorporating geographical and meteorological information into IceCloudNet could strengthen predictive performance in various atmospheric conditions.

The methodology proposed in this work paves the way for future research in climate science and atmospheric modeling, where machine learning models can provide significant insights into complex systems by bridging gaps in traditional observational methods. As machine learning techniques evolve, their integration with earth observation data will likely deepen our understanding of climate-related phenomena and improve our capacity to make informed predictions.