A Bayesian Kriging Regression Method to Estimate Air Temperature Using Remote Sensing Data

• Published in: Remote sensing (Basel, Switzerland) Vol. 11; no. 7; p. 767
• Main Authors: Zhang, Zhenwei, Du, Qingyun
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.04.2019
• Subjects: Air temperature; Algorithms; Artificial intelligence; Bayesian analysis; Bayesian framework; Climate; conterminous United States; Ground stations; Heat; Kriging interpolation; Kriging regression; Laboratories; Land surface temperature; Machine learning; Mathematical models; Methods; Parameter estimation; Performance prediction; Regression analysis; Remote sensing; Satellite imagery; Satellites; Spacetime; Spatial data; Statistical analysis; Statistical inference; Statistical methods; Statistical models; Studies; Temperature effects; Uncertainty analysis; Unevenness; Urban areas; Weather stations; New York; United States > US; China
• DOI: 10.3390/rs11070767
• ISSN: 2072-4292

Surface air temperature (Ta) is an important physical quantity, usually measured at ground weather station networks. Measured Ta data is inadequate to characterize the complex spatial patterns of Ta field due to low density and unevenness of the networks. Remote sensing can provide satellite imagery with large scale spatial coverage and fine resolution. Estimating spatially continuous Ta by integrating ground measurements and satellite data is an active research area. A variety of methods have been proposed and applied in this area. However, the existing studies primarily focused on daily Ta and failed to quantify uncertainties in model parameter and estimated results. In this paper, a Bayesian Kriging regression (BKR) method is proposed to model and estimate monthly Ta using satellite-derived land surface temperature (LST) as the only input. The BKR is a spatial statistical model with the capacity to quantify uncertainties via Bayesian inference. The BKR method was applied to estimate monthly maximum air temperature (Tmax) and minimum air temperature (Tmin) over the conterminous United States in 2015. An exploratory analysis shows a strong relationship between LST and Ta at the monthly scale, indicating LST has the great potential to estimate monthly Ta. 10-fold cross-validation approach was adopted to compare the predictive performance of the BKR method with the linear regression method over the whole region and the urban areas of the contiguous United States. For the whole region, the results show that the BKR method achieves a competitively better performance with averaged RMSE values 1.23 K for Tmax and 1.20 K for Tmin, which are also lower than previous studies on estimation of monthly Ta. In the urban areas, the cross-validation demonstrates similar results with averaged RMSE values 1.21 K for Tmax and 1.27 K for Tmin. Posterior samples for model parameters and estimated Ta were obtained and used to analyze uncertainties in the model parameters and estimated Ta. The BKR method provides a promising way to estimate Ta with competitively predictive performance and to quantify model uncertainties at the same time.