Format

Send to

Choose Destination
Sci Total Environ. 2019 May 15;665:300-313. doi: 10.1016/j.scitotenv.2019.02.077. Epub 2019 Feb 10.

Reconstruction of high spatial resolution surface air temperature data across China: A new geo-intelligent multisource data-based machine learning technique.

Author information

1
Key Laboratory of Environmental Change and Natural Disaster, Ministry of Education, Beijing Normal University, Beijing 100875, China; Faculty of Geographical Science, Academy of Disaster Reduction and Emergency Management, Beijing Normal University, Beijing 100875, China; State Key Laboratory of Earth Surface Processes and Resources Ecology, Beijing Normal University, Beijing 100875, China.
2
Key Laboratory of Environmental Change and Natural Disaster, Ministry of Education, Beijing Normal University, Beijing 100875, China; Faculty of Geographical Science, Academy of Disaster Reduction and Emergency Management, Beijing Normal University, Beijing 100875, China; State Key Laboratory of Earth Surface Processes and Resources Ecology, Beijing Normal University, Beijing 100875, China. Electronic address: zhangq68@bnu.edu.cn.
3
Department of Geosciences and Hydrology, University of Oslo, N-0316 Oslo, Norway.
4
College of Territorial Resource and Tourism, Anhui Normal University, Anhui 241002, China.

Abstract

Good knowledge of the surface air temperature (SAT) is critical for scientific understanding of ecological environment changes and land-atmosphere thermodynamic interactions. However, sparse and uneven spatial distribution of the temperature gauging stations introduces remarkable uncertainties into analysis of the SAT pattern. From a geo-intelligent perspective, here we proposed a new SAT reconstruction method based on the multisource data and machine learning technique which was developed by considering autocorrelation of the in situ observed SAT in both space and time, or simply STAML, i.e. Geoi-SVM (Geo-Intelligent Support Vector Machine), Geoi-BPNN (Geo-Intelligent Back Propagation Neural Network) and Geoi-RF (Geo-Intelligent Random Forest). The multisource data used in this study include the in situ observed SAT and multisource remotely sensed data such as MODIS land surface temperature, NDVI (Normalized Difference Vegetation Index) data. Intermodel comparisons amidst reconstructed SAT data were done to evaluate reconstructing performance of abovementioned models. Besides, the SAT reconstructed by CART (Classification and Regression Tree) was also included to evaluate the reconstructing performance of the models considered in this study when compared to SAT data by CART algorithm. We found that the estimation error of the reconstructed SAT by the STAML is smaller than 0.5K (Kelvin). In addition, it is interesting to note that the Geoi-RF performs better with Mean Absolute Error (MAE) of lower than 0.25K, and Root Mean Squared Error (RMSE) and Standard Deviation (SD) of lower than 0.5K respectively. Correlation coefficients between the reconstructed SAT by Geoi-RF and the observed SAT are close to 1. Besides, the estimation accuracy of the SAT by the Geoi-RF technique is 18.51-63.17% higher than that by the other techniques considered in this study. This study provides a new idea and technique for reconstruction of SAT over large spatial extent at regional and even global scale.

KEYWORDS:

Machine learning algorithm; Multisource data; Spatial resolution; Spatiotemporal autocorrelation; Surface air temperature

Supplemental Content

Full text links

Icon for Elsevier Science
Loading ...
Support Center