Search

Objectives: Traditional remote sensing-based air temperature estimation method usually used the global models, which ignored the effects of spatiotemporal heterogeneity, especially for the researches in large regional areas. Methods: Taking the Yangtze River Economic Zone as a typical research area, this paper introduces the geographically and temporally neural networks for high-precision temperature estimation. The influence of spatiotemporal heterogeneity was considered by establishing the local models in the generalized regression neural network. Remote sensing data, assimilation data and station data were fused to obtain the spatially continuous near-surface temperature. The model performance was evaluated by the site-based ten-fold cross-validation method. Results: The results show that the geographically and temporally weighted neural network had effectively improved the estimation accuracy with the root mean square error(RMSE) =1.899 ℃, mean absolute error (MAE) =1.310 ℃ and correlation coefficient (R) =0.976. Compared with the multiple linear regression method and the traditional global neural network, the MAE value decreased by 1.112 ℃ and 0.378 ℃, respectively. Conclusions: The temperature mapping results indicated that the model used in this paper can well reflect the spatial distribution differences, which means that this study is possible to provide a new way for temperature estimation with high precision. [ABSTRACT FROM AUTHOR]