Journal of Geo-information Science ›› 2019, Vol. 21 ›› Issue (6): 799-813.doi: 10.12082/dqxxkx.2019.190014

    Next Articles

Spatiotemporal Estimation of High-Accuracy and High-Resolution Meteorological Parameters based on Machine Learning

Ying FANG1,2(), Lianfa LI1,2,*()   

  1. 1. State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Science, Beijing 100101, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2019-01-19 Revised:2019-03-04 Online:2019-06-15 Published:2019-06-15
  • Contact: Lianfa LI E-mail:fangying@lreis.ac.cn;lilf@lreis.ac.cn
  • Supported by:
    National Natural Science Foundation of China, No.41471376, 41871351;Priority Research Program of the Chinese Academy of Science, No.XDA19040501

Abstract:

The meteorological stations are sparsely distributed across Mainland China. In terms of generating high-resolution surfaces of meteorological parameters, the estimation accuracy of existing models is limited for air temperature, and is poor for relative humidity and wind speed (few studies reported). With the measurement data of 824 monitoring stations covering the mainland of China in 2015, this study compared the typical Generalized Additive Model (GAM) and autoencoder-based residual neural network (here after, residual network for short) in terms of predicting three meteorological parameters, i.e. air temperature, relative humidity, and wind speed. The performances of the two models were evaluated through 10-fold cross-validation. Basic variables including latitude, longitude, elevation, and the day of the year are used in the air temperature models. In addition to the basic variables, the relative humidity models use air temperature and ozone concentration as covariates, and the wind speed models use wind speed coarse-resolution reanalysis data as covariates. In our spatiotemporal models, spatial coordinates capture the spatial variation and time index of the day captures the time variation. Compared with GAM, residual network significantly improved the prediction accuracy: on average, CV (Cross Validation) R2of the three meteorological factors increased by 0.21, CV RMSE decreased by 37%, and the relative humidity model improved the most (CV R2: 0.85 vs. 0.52, CV RMSE: 7.53% vs. 13.59%). With incorporation of the monthly index in the relative humidity models, the accuracy was greatly improved, indicating that the different levels of time factors are important for the relative humidity models. Furthermore, we also discussed the effectiveness and limitations of coarse resolution reanalysis data and nearest neighbor values as covariates. This study shows that the residual network model can greatly improve the accuracy of national high spatial (1 km) and temporal (daily) resolution meteorological data as opposed to traditional GAMs. Our findings provide implications for high-accuracy and high-resolution mapping of meteorological parameters in China.

Key words: meteorological factors, machine learning, residual autoencoder, Mainland China, GAM, deep learning, high resolution