Machine learning of cloud types shows higher climate sensitivity is associated with lower cloud biases

Abstract

Abstract. Uncertainty in cloud feedback in climate models is a major limitation in projections of future climate. Therefore, to ensure the accuracy of climate models, evaluation and improvement of cloud simulation is essential. We analyse cloud biases and cloud change with respect to global mean near-surface temperature (GMST) in climate models relative to satellite observations, and relate them to equilibrium climate sensitivity, transient climate response and cloud feedback. For this purpose, we develop a supervised deep convolutional artificial neural network for determination of cloud types from low-resolution (approx. 1°×1°) daily mean top of atmosphere shortwave and longwave radiation fields, corresponding to the World Meteorological Organization (WMO) cloud genera recorded by human observers in the Global Telecommunication System. We train this network on a satellite top of atmosphere radiation observed by the Clouds and the Earth’s Radiant Energy System (CERES), and apply it on the Climate Model Intercomparison Project phase 5 and 6 (CMIP5 and CMIP6) historical and abrupt-4xCO2 experiment model output and the ECMWF Reanalysis version 5 (ERA5) and the Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2) reanalyses. We compare these with satellite observations, link biases in cloud type occurrence derived from the neural network to change with respect to GMST to climate sensitivity, and compare our cloud types with an existing cloud regime classification based on the Moderate Resolution Imaging Spectroradiometer (MODIS) and International Satellite Cloud Climatology Project (ISCCP) satellite data. We show that there is a significant negative linear relationship between the root mean square error of cloud type occurrence derived from the neural network and model equilibrium climate sensitivity and transient climate response (Bayes factor 22 and 17, respectively). This indicates that models with a better representation of the cloud types globally have higher climate sensitivity. Factoring in results from other studies, there are two possible explanations: either high climate sensitivity models are plausible, contrary to combined assessments of climate sensitivity by previous review studies, or the accuracy of representation of present-day clouds in models is negatively correlated with the accuracy of representation of future projected clouds.

The response of hemispheric differences in Earth’s albedo to CO2 forcing in coupled models and its implications for shortwave radiative feedback strength

“…The link between mean state magnitude and forced response is ultimately not self-evident (McCoy et al, 2014;Zelinka et al, 2022;Kuma et al, 2022), and the time scales of the SH extratropical cloud changes involved may mean that albedo symmetryrestoring remote compensations do not act on the time span of the observations that we presently have (Frey et al, 2017;Gjermundsen et al, 2021).…”

Machine learning of cloud types in satellite observations and climate models

“…The datasets used in our analysis are publicly available: CERES (2022), GISTEMPv4 (GISTEMP Team, 2021), CMIP5 (2022), CMIP6 (2022), MERRA-2 (2022), ERA5 (2022) and IDD (Unidata, 2003). The code used in our analysis is open source and available on GitHub (Kuma, 2022) and Zenodo (https://doi.org/10.5281/zenodo.7400793, Kuma et al, 2022).…”

“…We try to answer the question of whether cloud type biases and change with respect to GMST are related to cloud feedback, ECS and TCR in the CMIP models. The ANN and the associated code are made available under an open-source license (Kuma et al, 2022).…”

Machine-Learned Cloud Classes From Satellite Data for Process-Oriented Climate Model Evaluation

“…However, in addition to being built on the rather low resolution of (280 km) 2 of the ISCCP-D1 [17] product, uncertainties and artifacts introduced by the satellite simulators can affect the results [20], [21]. In a recent study, a convolutional neural network was used on (4000 km) 2 grid cells to assign the amount of each of four cloud classes per cell [22]. In [22], the classes were derived from WMO classes detected from surface observations, and the method is applicable to climate model output.…”