Prediction of accurate wind speed is necessary for a variety of applications such as energy production, agriculture, climate modeling, and weather forecasting. Various satellites orbiting the earth measure the wind speed, which is particularly useful as they provide measurements of wind speed over large areas and in remote locations that might be difficult to measure using other methods. However, satellite-based wind speed measurements have relatively low spatial resolution compared to other methods, such as ground-based radar. In this research, we develop WindSR and a lightweight tiny-WindSR to improve the resolution of satellite wind speed data by four times from the NASA’s GEOS-5 Nature Run dataset. WindSR has SRResNet-based architecture consisting of several Residual-in-Residual Dense Blocks to compute features from low spatial resolution (28 km) wind speed for upscaling. We train WindSR with more than 20,000 pairs of low-resolution (28 km) and corresponding high-resolution (7 km) wind speed data and evaluate its performance on the validation set consisting of 2,102 wind speed images. Experimental results show that WindSR outperforms classical upsampling algorithms, such as Bicubic interpolation and Lanczos interpolation by 17.89% and general-purpose super-resolution GANs such as BSRGAN and SwinIR by up to 11.35% on the RMSE metric. The dataset developed in this research is publicly available at: https://github.com/sekilab/WindSR_DatasetTest.
Full Text Finder