LAUSR

Timely griddedgross domestic product (GDP) data is a fundamental indicator in many applications. It is critical to characterize the complex relationship between GDP and its auxiliary information for accurately estimating gridded GDP. However, few knowledge is available about the performance of deep learning approaches for learning this complex relationship. This article develops a novel convolutional neural network based GDP downscaling approach (GDPnet) to transform the statistical GDP data into GDP grids by integrating various geospatial big data. An existing autoencoder-based downscaling approach (Resautonet) is employed to compare with GDPnet. The latest county-level GDP data of China and the multiple geospatial big data are adopted to generate the 1-km gridded GDP data in 2019. Due to the different related auxiliary data of each GDP sector, the two downscaling approaches are first separately built for each GDP sector and then the results are merged to the gridded total GDP data. Experimental results show that the two deep learning approaches had good predictive power with R2 over 0.8, 0.9, and 0.92 for the three sectors tested by county-level GDP data. Meanwhile, the proposed GDPnet outperformed the existing Resautonet. The average R2 of GDPnet was 0.034 higher than that of Resautonet in terms of county-level GDP test data. Furthermore, GDPnet had higher accuracy (R2 = 0.739) than Resautonet (R2 = 0.704) assessed by town-level GDP data. In addition, the proposed GDPnet is faster (about 78% running time) than the Resautonet. Hence, the proposed approach provides a valuable option for generating gridded GDP data.