Comparative Study on the Application of Statistical Downscaling in Urban Climate Research

Often we obtain urban climatic observations from sensors within the city or satellites passing over the city. The data collected within the city is generally of high resolution but difficult to obtain, while the remote sensing data generally has coarser resolution but is more readily available. This creates problems when studying small domains, such as urban areas. Currently, remotely sensed data remains the one of the most accessible and reliable sources of climate observations, especially as the resolution continues to improve; however, it simply is not fine enough to adequately assess urban climate variables such as heat, precipitation, and evapotranspiration. These variables have large impacts on the assessment of hydroclimatic extremes (i.e., droughts and floods) as well as the generally livability of the city, therefore it is imperative that we obtain these observations at a high resolution. One such way to obtain this high resolution data is to use downscaling techniques on the readily available remote sensing observations. This study serves as a review of different statistical downscaling techniques for use in urban applications. Here, we show the results from three different statistical downscaling methods chosen for their applicability in urban settings. These three techniques, detailed below, were used to improve resolution of several urban climate variables, with a specific focus in aiding the ongoing research on urban floods and droughts. We also ran a dynamical downscaling model in addition to the statistical methods. This model takes physical principles into account to improve accuracy, but is very computationally expensive. The first method we investigated was cascading universal multifractals (UMs). This method, outlined by Gires et al. (2012), has a basis in the assumption that statistical moments of an arbitrary pixel in a data source is scalable. This method has two main steps: (1) estimate the UM parameters that will be used in the downscaling; and (2) iteratively run through a random discrete multiplicative cascade. In other words, we need to find the scaling factors then use those factors to determine the ‘random factor’. This random factor is used to determine the value of the child pixel following each cascade. Each cascade involves squaring the number of pixels in the dataset. Finally, you repeat the process until you reach the desired number of pixels. We also looked into two methods based on regression. In order to use regression to downscale climate data, you need to pick an appropriate high-resolution (HR) variable, or predictor, to compare with the low-resolution (LR) variable. The predictor needs to have a strong correlation with the LR variable. There are many processes to select the predictor variable, which is where the two regression methods chosen for this study differ. The first method uses principle component analysis (PCA), as outlined by Zakšek et al. (2012), and the second method uses minimum redundancy maximum relevance (MRMR), as outlined by Bechtel et al. (2012). The purpose of these different processes is to take several potential predictors and determine which is the best to use in the regression. In order to complete the regression, one needs to upscale the predictor data to the LR domain, obtain the linear function, then apply the function in the HR domain. Finally, we tested the ALARO model coupled with TEB, a dynamical downscaling method outlined by Hamdi et al. (2014). The ALARO model (Gerard et al., 2009) is an atmospheric model used to downscale climatic data, while the TEB model (Masson, 2000) is used to classify urban areas. This model allowed us to provide boundary conditions as well as integrate microphysics into the downscaling framework, something the other methods did not take into account. Following the downscaling work, we compared the results with high resolution climate data from several “smart city” projects, such as Chicago Array of Things, and remote sensing sources, such as AVIRIS. This data served as the ‘control run’ to which we compared the downscaling methods and assessed their accuracy. Finally, our study also includes a review of the current tools available in terms of urban downscaling and the areas in which improvement is needed, as well as an introduction to a joint portal between UNESCO, IISc, and Purdue that focuses on applications for urban hydroclimatic extremes. The overall aim of this study is to provide a broad review of urban downscaling techniques that will be shared via the urban downscaling portal mentioned above. We will demonstrate the accuracy (or inaccuracy) of several methods commonly used in downscaling research. These methods will be evaluated based their accuracy when compared to the high resolution control data collected from smarty city sensors and remote sensing sources. It is our goal to improve the resolution of urban climate observations so that solutions can be found to help prepare for and adapt to future climate variability.