Unsupervised learning statistical downscaling based on residual network of attention mechanism
Buhong Ge, Tao Wu, Jing Hu, Jingjue Chen · 2022 IEEE 2nd International Conference on Data Science and Computer Application (ICDSCA) · 2022
Because the numerical model dynamic downscaling takes a long time and requires high computing resources, a statistical downscaling model of residual network based on unsupervised learning is proposed. First of all, this network is inspired by the “zero-shot” super-resolution, using ResNet50 as the network residual module, making the network deeper while simplifying the degree of training. It link the low-scoring information of low-scoring pictures to the depths of the network to improve the effect of the network on feature extraction. In addition, KernelGAN is used to train it with the low spatial resolution pictures output by the WRF mode instead of the down sampling blur kernel of the low resolution pictures corresponding to the input in ZSSR. Finally, the channel attention and the spatial attention mechanism are added to change the equality of feature extraction by the main body network. The simulated high temperature image (T2) of Tianjin Port and Bohai Bay two meters from the ground, which is better than Bicubic, Kriging, SRCNN, ZSSSR, KernelGAN+ZSSR and other methods.