Linear Versus Non‐linear Models in Predicting Low‐Cloud Regime Occurrences From Cloud‐Controlling Factors

Abstract

Accurately predicting low‐cloud occurrences from environmental conditions is essential for reducing uncertainties in cloud‐climate feedbacks. This study evaluates linear and non‐linear frameworks—specifically, anomaly‐based ridge regression and neural network (NNet) that predict raw target values—to predict cloud regime occurrence frequencies from low‐cloud‐amount‐indices (LCAIs) and other cloud‐controlling factors at monthly, 4° resolution, using data from selected subtropical study regions. While the combined use of multiple LCAIs notably enhances linear model performance within the training domains, these models exhibit poor geographic transferability due to the diversity of cloud‐environment relationships. In contrast, an NNet model trained using raw physical variables consistently outperforms the linear models and captures established physical mechanisms, such as the capping inversion effect. Nevertheless, global application of the baseline NNet model reveals a “generalization gap,” particularly in extra‐tropical and tropical warm‐pool environments that are not represented in the training data set. To address this limitation, an expanded model is trained using additional samples from poorly predicted regions. This approach substantially reduces regional biases and improves mean absolute error by 30%–50%. However, it also introduces greater nonlinearity and regional trade‐offs, including slightly inferior performance within the original training regions. These findings suggest that, although a generalized cloud‐regime model may be achievable, it requires high‐capacity non‐linear architectures and diverse sampling strategies to address the variable nature of cloud–environment relationships. Once successfully developed, however, such a framework could serve as a computationally efficient alternative to satellite simulators for CR‐based evaluation of Earth System Model outputs.