Assessing the Capability of a Downscaled Urban Land Surface Temperature Time Series to Reproduce the Spatiotemporal Features of the Original Data

Panagiotis Sismanidis, Iphigenia Keramitsoglou, Chris Kiranoudis, Benjamin Bechtel
2016 Remote Sensing  
The downscaling of frequently-acquired geostationary Land Surface Temperature (LST) data can compensate the lack of high spatiotemporal LST data for urban climate studies. In order to be usable, the generated datasets must accurately reproduce the spatiotemporal features of the coarse-scale LST time series with greater spatial detail. This work concerns this issue and exploits the high temporal resolution of the data to address it. Specifically, it assesses the accuracy, correct pattern
more » ... n and the spatiotemporal inter-relationships of an urban three-month-long downscaled geostationary LST time series. The results suggest that the downscaling process operated in a consistent manner and preserved the radiometry of the original data. The exploitation of the data inter-relationships for evaluation purposes revealed that the downscaled time series reproduced the smooth diurnal cycle, but the autocorrelation of the downscaled data was higher than the original coarse-scale data. Overall, the evaluation process showed that the generation of high spatiotemporal LST data for urban areas is very challenging, and to deem it successful, it is mandatory to assess the temporal evolution of the urban thermal patterns. The results suggest that the proposed tests can facilitate the evaluation process. Surface UHI (SUHI) [13] [14] [15] [16] . In contrast to TA data, which are point measurements confined to local conditions [17] , thermal infrared (TIR) remote sensing is capable of providing a simultaneous and synoptic view of the urban thermal environment [18] . This enables the more detailed assessment of the urban hotspots and the relationship between the urban core and the surrounding natural lands [18] [19] [20] . Nevertheless, the use of satellite TIR data is not straightforward, and a number of limitations and problems, such as the atmospheric influence, the unknown emissivity and the effective anisotropy, have to be addressed prior to their exploitation [21] . To that end, one of the most important problems concerns the spatial and temporal resolution of the LST data. In particular, the available satellite sensors cannot provide datasets that capture the high spatial and temporal variability of SUHIs, and thus, their exploitation in urban climate studies is limited [22] [23] [24] [25] . For instance, the Sun-synchronous Landsat series satellites, which offer the appropriate high spatial resolution (~100 m), acquire LST data every 16 days; whereas the Spinning Enhanced Visible and Infrared Imager (SEVIRI) onboard Meteosat Second Generation (MSG) geostationary satellites provides~4 km TIR data with a more appropriate temporal resolution of 5-15 min. To overcome this problem, the statistical downscaling of geostationary LST data has been proposed. This process can lead to the generation of Downscaled LST (DLST) time series that combine high spatial and temporal resolution and preserve the radiometry of the original thermal data [22] [23] [24] [25] [26] . In detail, the LST statistical downscaling is a scaling process that aims to enhance the spatial resolution of coarse-scale LST imagery using fine-scale auxiliary datasets. These auxiliary datasets are usually referred to as LST predictors and are statistically correlated to the LST [26]. An LST downscaling scheme comprises two major parts. The first part is the set of LST predictors used for explaining the spatial variation of LST, while the second part is the regression tool used for associating the LST predictors with the LST data. These two parts are synergistically exploited in a three-stage procedure: firstly, the LST predictors are upscaled and co-registered to the coarse-scale LST data; then, a relationship between the coarse-scale LST data and the LST predictors is established using the regression tool; and finally, this relationship is applied to the fine-scale LST predictors, so as to generate the DLST data. Agam et al. [27] and Kustas et al. [28] include also a fourth process stage, which is an adjustment of the generated DLST data based on the differences of the observed and regressed coarse-scale LSTs. This post-downscaling processing aims to compensate the loss of LST variability due to the use of inflexible regression tools, such as least-square or linear fits [28] . In recent years, a large number of relevant works have been published [26] testing different LST predictors (separately or combined), such as: Vegetation Indices (VIs), emissivity data, land cover maps and topography data; and also regression tools, such as: linear regressors, least-square fits, Support Vector Regression Machines (SVMs) and Neural Networks (NNs) [22] [23] [24] [25] [26] [27] [28] [29] . From the available list of LST predictors, the most widely used is the Normalized Vegetation Difference Index (NDVI) [27] [28] [29] , which is strongly negatively correlated with summer daytime LST imagery [30] . Most of the aforementioned works focus on the downscaling of~4 km data down to 1 km. The downscaling to even higher spatial resolutions (<500 m) is still a very challenging task, mainly because the assumption that the LST data-predictor relationship is valid in both spatial scales weakens or ceases to apply [31] . The work of Bechtel et al. [24] is one of the few studies that discusses the downscaling of geostationary urban LST data down to 100 m with a Root-Mean-Square-Error (RMSE) of 2.2˝C (recent LST fusion studies [32, 33] also report similar downscaling factors). In this work, a large set of LST predictors was utilized that also included for the first time LST annual climatology data in the form of annual cycle parameters (ACPs) [34] [35] [36] . The use of the ACPs for downscaling LST data below the 1-km cap provided very promising results [24] . However, this type of LST predictor has not been used as extensively as the rest, and thus, the available relevant literature is still very limited. Besides the identification of more robust predictors, another issue that requires attention is the evaluation of the generated high spatiotemporal DLST time series, which is hampered by the lack of appropriate ground truth data [21] . Presently, most downscaling studies utilize independent LST image data (confined to certain time spots) with which they compare the generated data and
doi:10.3390/rs8040274 fatcat:owxp2oy34nghppg7yepmn4zpp4