A physical model-based method for retrieving urban land surface temperatures under cloudy conditions

Authors: FU, PENG - University Of Illinois; XIE, YANHUA - University Of Wisconsin; WENG, QIHAO - Indiana State University; MYINT, SOE - Arizona State University; MEACHAM-HENSOLD, KATHERINE - University Of Illinois; Bernacchi, Carl

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/10/2019
Publication Date: 9/1/2019
Citation: Fu, P., Xie, Y., Weng, Q., Myint, S., Meacham-Hensold, K., Bernacchi, C.J. 2019. A physical model-based method for retrieving urban land surface temperatures under cloudy conditions. Remote Sensing of Environment. 230:111191. https://doi.org/10.1016/j.rse.2019.05.010.
DOI: https://doi.org/10.1016/j.rse.2019.05.010

Interpretive Summary: The temperature at the land surface is important as it provides critical information needed to understand weather conditions. Plant and ecosystem functioning, such as plant growth rates and crop yields, are very dependent on these weather conditions. While thermometers are great for measuring land surface temperatures, it is impossible to place them everywhere. Therefore, science relies heavily on satellite data to understand what the temperatures are across the whole land surface of the planet. However, when clouds are present, measuring surface temperatures with satellites is traditionally very difficult. This experiment used satellite information, ground-based thermometers, and machine learning approaches to see whether land surface measurements from satellites can be determined under cloudy conditions. Machine learning approaches are complex tools to figure out relationships using multiple data sources to make important predictions. Using this approach helped to improve the ability to measure land surface temperatures under cloudy conditions and the results can help to improve weather forecasts and ecosystem predictions.

Technical Abstract: Satellite-derived land surface temperature (LST), due to its synoptic coverage, has been widely used for understanding surface energy and carbon fluxes at local, regional, and global scales. Despite great achievements to develop practical algorithms to estimate LSTs from satellite thermal infrared (TIR) data, the retrieval of LSTs for overcast moments has received much less attention. The existing techniques, such as passive microwave (PMW) measurements-based approaches, surface energy balance (SEB)-based models, and reconstruction algorithms relying on spatial/temporal information, for estimating LSTs under cloudy skies may not fulfill the needs for a spatially and temporally consistent LST dataset. Inspired by the recent advancements in the physically based urban canopy models (UCMs), this study synergistically used the coupled Weather Research and Forecasting Model (WRF)/UCM system and the random forest (RF) regression technique to estimate LSTs under cloudy conditions. Taking the Baltimore-Washington metropolitan region as a test site, the WRF/UCM simulations (LSTs) were performed from April 28 to May 20, 2011. The MODIS LST images of the same period were downloaded and used to test the developed approach. LSTs under cloudy conditions for a partially cloudy image were retrieved using the RF model calibrated by the clear-sky pixels from the same image. For a fully cloud-contaminated image, clear-sky pixels from neighboring images were used to calibrate the RF model for estimating LSTs under cloudy conditions. Results showed that the WRF/UCM system could well capture the diurnal temperature patterns but tended to underestimate air temperatures. The correlation coefficient value between the MODIS LSTs and simulated LSTs exhibited a wide range from 0.5 to 0.9 with the RMSE value varying from 1.0 K to 9.0 K across land covers. The utilization of the RF regression technique for estimating LSTs under cloudy conditions from a partially cloud-contaminated LST image greatly reduced the RMSE value to 1.7 K without compromising the correlation coefficient value. For a fully cloud-contaminated LST image, LSTs were estimated with the correlation coefficient and RMSE values of 0.74 and 2.1 K, respectively. It should be cautioned that the pixels selected for calibrating the RF model should be randomly distributed within the study area. Overall, the presented approach has potential to generate consistent LST dataset that would be useful for various environmental applications. Further research efforts can be made to understand the impacts of the temporal distance between neighboring images and the target image on the performance of the developed approach.