Atmospheric classification as a cloud and precipitation evaluation tool in models and observations

Abstract

Atmospheric classifications are created for two regions using an automated clustering technique first described in Marchand et al. 2009. This method applies an iterative clustering algorithm to regional snapshots of dynamic and thermodynamic variables from the ERA-Interim reanalysis to define atmospheric states. An atmospheric state in this context is a frequently occurring weather pattern for the region. In creating the states, a time series of atmospheric state for the region is created. These states then serve as a basis for compositing other weather observations, creating distributions of associated weather variables such as cloud occurrence, precipitation, and radiative fluxes. The states are also suitable for sorting output from general circulation models, allowing the comparison of observed and modeled cloud variables on a state-by-state basis. Analysis of which atmospheric states lead to the most error in modeled values provides insight into the particular atmospheric conditions and processes that are problematic for models. A classification for a region surrounding Darwin, Australia is used to define periods of monsoon activity and investigate the interannual and intraseasonal variability of the Australian monsoon, as well as long-term trends in precipitation at Darwin. The classification creates a time series of atmospheric states, two of which are identified as corresponding to the active monsoon and the monsoon break. Occurrence of these states is used to define onset, retreat, and intensity of the monsoon season and the timing of individual active periods. Previous studies investigated the role of the MJO during the monsoon season, but have differed on whether the MJO creates a characteristic period or duration of active monsoon periods. We use our classification-based metrics of monsoon activity to examine the timing of individual active periods each season relative to the phase of the MJO, showing that the passage of the convective anomaly over Darwin helps both trigger and end individual active periods. Lastly, we look at trends in the occurrence of the atmospheric states and find that the number of active monsoon days has increased over the previous 33 years. We show that this, rather than changes in the daily rainfall rate during active monsoon periods, is responsible for an increasing trend in annual precipitation at Darwin during that time. A second classification for the region surrounding the Southern Great Plains (SGP) site in Oklahoma produced 21 atmospheric states. Analysis of these states showed that they represented different stages of passing synoptic systems, with some states representing warm fronts, others cold fronts, still others high pressure systems, and so on. Snapshots from a 2° run of the AM3 model from GFDL were sorted according to these states, producing a time series of state within the model. The time series of model state was used to composite ISCCP simulator output from the model according to each state. These simulated ISCCP composites are compared to composites of observed ISCCP data. In all states the model does not produce nearly enough high thin cloud. The representation of other cloud types depends on the state. We show that for states which have large-scale ascent (warm fronts and cold fronts) the model does not produce enough deep thick cloud, while for states that have parameterized convection (e.g. high pressure systems) the model produces too much cloud. The former we interpret as the model struggling to resolve the dynamics of fronts, while the latter indicates that the deep convective parameterization triggers too often. We demonstrate that the overall cloud amount bias is driven primarily by within-state errors in cloud amount rather than by errors in the relative frequency of occurrence of states. We compare the results of the 2° run of the AM3 to a second run of the model at 0.5° resolution. We find that thick cloud increases in states with large-scale ascent, and interpret this result as a result of fronts being closer to resolved in the higher resolution run of the model. We also find that high thick cloud decreases in states with parameterized convection. We find that this result is due to the distribution of CAPE values becoming more skewed at higher resolution, and thus triggering the deep convective parameterization less often. As a result of these improvements in within-state cloud amount error, the overall bias in cloud amount is the higher resolution run is as much due to errors in the frequency of occurrence of states as it is due to within-state errors. Interestingly, while individual state cloud amount biases decrease, the overall bias increases, due to changes in the distribution of states. Whether the higher resolution run constitutes a better simulation of clouds is thus an interesting question which we address.