Search

This paper exposes problems of the commonly used technique of splitting the available data in neural spatial interaction modelling into training, validation, and test sets that are held fixed and warns about drawing too strong conclusions from such static splits. Using a bootstrapping procedure, we compare the uncertainty in the solution stemming from the data splitting with model specific uncertainties such as parameter initialization. Utilizing the Austrian interregional telecommunication traffic data and the differential evolution method for solving the parameter estimation task for a fixed topology of the network model [i.e. J =8] this paper illustrates that the variation due to different resamplings is significantly larger than the variation due to different parameter initializations. This result implies that it is important to not over-interpret a model, estimated on one specific static split of the data. [ABSTRACT FROM AUTHOR]

Spatial predictions of wildfire spread are used operationally and in risk estimation. It is important that their outputs are validated to quantify predictive performance and uncertainty. There are numerous loss functions for this simulation validation. Our paper synthesises ten common and five novel loss functions for evaluating the performance of wildfire predictions against observed or simulated wildfires. We describe each loss function; test their sensitivity to scale, rotation and translation; and demonstrate their application with three case study wildfires in south-eastern Australia. Based on their purpose, there are three categories of loss functions: general overlap, partial overlap and distance. Within these categories, there is functional redundancy as many loss functions are appropriate and perform similarly. Using loss functions from across the three different categories is the most comprehensive for evaluating wildfire models, but the choice depends on the simulation purpose and interpretation. • Spatial predictions of wildfire spread are frequently used but not always evaluated. • We synthesised fifteen loss functions for evaluating predicted vs observed wildfire spread. • There is not a 'best' loss function, it depends on simulation use and interpretation. [ABSTRACT FROM AUTHOR]

Air pollution and associated human exposure are important research areas in Greater Sydney, Australia. Several field campaigns were conducted to characterize the pollution sources and their impacts on ambient air quality including the Sydney Particle Study Stages 1 and 2 (SPS1 and SPS2), and the Measurements of Urban, Marine, and Biogenic Air (MUMBA). In this work, the Weather Research and Forecasting model with chemistry (WRF/Chem) and the coupled WRF/Chem with the Regional Ocean Model System (ROMS) (WRF/Chem-ROMS) are applied during these field campaigns to assess the models' capability in reproducing atmospheric observations. The model simulations are performed over quadruple-nested domains at grid resolutions of 81-, 27-, 9-, and 3-km over Australia, an area in southeastern Australia, an area in New South Wales, and the Greater Sydney area, respectively. A comprehensive model evaluation is conducted using surface observations from these field campaigns, satellite retrievals, and other data. This paper evaluates the performance of WRF/Chem-ROMS and its sensitivity to spatial grid resolutions. The model generally performs well at 3-, 9-, and 27-km resolutions for sea-surface temperature and boundary layer meteorology in terms of performance statistics, seasonality, and daily variation. Moderate biases occur for temperature at 2-m and wind speed at 10-m in the mornings and evenings due to the inaccurate representation of the nocturnal boundary layer and surface heat fluxes. Larger underpredictions occur for total precipitation due to the limitations of the cloud microphysics scheme or cumulus parameterization. The model performs well at 3-, 9-, and 27-km resolutions for surface O3 in terms of statistics, spatial distributions, and diurnal and daily variations. The model underpredicts PM2.5 and PM10 during SPS1 and MUMBA but overpredicts PM2.5 and underpredicts PM10 during SPS2. These biases are attributed to inaccurate meteorology, precursor emissions, insufficient SO2 conversion to sulfate, inadequate dispersion at finer grid resolutions, and underprediction in secondary organic aerosol. The model gives moderate biases for net shortwave radiation and cloud condensation nuclei but large biases for other radiative and cloud variables. The performance of aerosol optical depth and latent/sensible heat flux varies for different simulation periods. Among all variables evaluated, wind speed at 10-m, precipitation, surface concentrations of CO, NO, NO2, SO2, O3, PM2.5, and PM10, aerosol optical depth, cloud optical thickness, cloud condensation nuclei, and column NO2 show moderate-to-strong sensitivity to spatial grid resolutions. The use of finer grid resolutions (3- or 9-km) can generally improve the performance for those variables. While the performance for most of these variables is consistent with that over the U.S. and East Asia, several differences along with future work are identified to pinpoint reasons for such differences. [ABSTRACT FROM AUTHOR]

The projection of temperature extremes by climate models participating in the Intergovernmental Panel on Climate Change Fourth Assessment Report (AR4) are examined regionally over Australia. Minimum and maximum temperature extremes are defined as the 20 year return value calculated using extreme value theory. Three measures of model evaluation, a means-based, a distribution-based [via probability density functions (PDFs)] and an extreme-based (via the tails of PDFs) method, are used to compare daily model data to observed daily data over various climatic regions for a 20 year period. Model ensembles consisting of the 'better' and 'poorer' models determined by each measure of skill are created for each region. These are compared with an all-model ensemble to examine the difference in more skilled ensemble projections of temperature extremes in the A2 (high emissions) scenario for 2046-2065 and 2081-2100. If either of the distribution-based evaluation methods were used to distinguish models, the higher skilled models projected smaller increases in the 20 year return values than the all-model ensemble for both maximum temperature and minimum temperature. For some regions, the 90% confidence intervals of the better and poorer ensemble ranges did not overlap, indicating that projections are statistically significantly different. We show that the means-based evaluation produces less consistent results to the two distribution-based evaluation methods. We conclude that specific AR4 models, shown to be relatively poor over most regions of Australia by different skill metrics, bias the projected increase in the 20 year temperature extremes towards higher values. We also suggest that performance in simulating the mean climate is an unreliable measure of climate model capacity used to select models for projecting changes in extremes over Australia. Copyright © 2012 Royal Meteorological Society [ABSTRACT FROM AUTHOR]

The ability of meteorological models to accurately characterise regional meteorology plays a crucial role in the performance of photochemical simulations of air pollution. As part of the research funded by the Australian government's Department of the Environment Clean Air and Urban Landscape hub, this study set out to complete an intercomparison of air quality models over the Sydney region. This intercomparison would test existing modelling capabilities, identify any problems and provide the necessary validation of models in the region. The first component of the intercomparison study was to assess the ability of the models to reproduce meteorological observations, since it is a significant driver of air quality. To evaluate the meteorological component of these air quality modelling systems, seven different simulations based on varying configurations of inputs, integrations and physical parameterizations of two meteorological models (the Weather Research and Forecasting (WRF) and Conformal Cubic Atmospheric Model (CCAM)) were examined. The modelling was conducted for three periods coinciding with comprehensive air quality measurement campaigns (the Sydney Particle Studies (SPS) 1 and 2 and the Measurement of Urban, Marine and Biogenic Air (MUMBA)). The analysis focuses on meteorological variables (temperature, mixing ratio of water, wind (via wind speed and zonal wind components), precipitation and planetary boundary layer height), that are relevant to air quality. The surface meteorology simulations were evaluated against observations from seven Bureau of Meteorology (BoM) Automatic Weather Stations through composite diurnal plots, Taylor plots and paired mean bias plots. Simulated vertical profiles of temperature, mixing ratio of water and wind (via wind speed and zonal wind components) were assessed through comparison with radiosonde data from the Sydney Airport BoM site. The statistical comparisons with observations identified systematic overestimations of wind speeds that were more pronounced overnight. The temperature was well simulated, with biases generally between ±2 °C and the largest biases seen overnight (up to 4 °C). The models tend to have a drier lower atmosphere than observed, implying that better representations of soil moisture and surface moisture fluxes would improve the subsequent air quality simulations. On average the models captured local-scale meteorological features, like the sea breeze, which is a critical feature driving ozone formation in the Sydney Basin. The overall performance and model biases were generally within the recommended benchmark values (e.g., ±1 °C mean bias in temperature, ±1 g/kg mean bias of water vapour mixing ratio and ±1.5 m s−1 mean bias of wind speed) except at either end of the scale, where the bias tends to be larger. The model biases reported here are similar to those seen in other model intercomparisons. [ABSTRACT FROM AUTHOR]