Figure 1. General architecture of DL4DS. A low-resolution gridded dataset can be downscaled, with the help of auxiliary predictor and static variables, and a high-resolution reference dataset. The mapping between the low- and high-resolution data is learned with either a supervised or a conditional generative
adversarial DL model.Figure 2. Panel (a) shows the main blocks and layers implemented in DL4DS. Panel (b) shows the structure of the main spatial convolutional block, a succession of two convolutional layers with interleaved regularization operations, such as dropout or normalization. Blocks and operations shown with dashed lines are optional.Figure 3. DL4DS supervised DL models, as well as generators, are composed of a backbone section (examples in panels [a–d]) and an output module (panel [e]). Panel (a) shows the backbone of models for downscaling pre-upsampled spatial samples using either residual or dense blocks. Panel (b) presents the backbone of a model for downscaling spatial samples using ConvNext-like blocks and one of the post-upsampling blocks described in Section 3.4.1. Panel (c) shows the backbone of a model for downscaling pre-upsampled spatial samples using an encoder-decoder structure. Panel (d) shows the backbone of a model for downscaling spatiotemporal samples using recurrent-convolutional blocks and a post-upsampling block. These backbones are followed by the output module (see Section 3.4.2) shown in panel (e). The color legend for the blocks used here is shown in Figure 2a.Figure 4. Example of a conditional generative adversarial model for spatiotemporal samples in postupsampling mode (see Section 3.4.1). Two networks, the generator shown in panel (a), and discriminator shown in panel (b), are trained together optimizing an adversarial loss (see Section 3.5). The color legend for the blocks used here is shown in Figure 2a.Figure 5. A reference NO2 surface concentration field from the low-resolution CAMS global reanalysis is shown in panel (a). In panel (b), we present a resampled version, via bicubic interpolation, of the lowresolution reference field. This interpolated field looks overly smoothed and showcases the inefficiency of simple resampling methods at restoring fine-scale information. Panel(c): the corresponding highresolution field from the CAMS regional reanalysis. Both low- and high-resolution grids were taken from the holdout set for the same time step. The maximum value shown corresponds to the maximum value in the high-resolution grid.Figure 6. Examples of downscaled products obtained with DL4DS, corresponding to the reference grid
shown in Figure 5a. The models corresponding to each panel are detailed in Table 1.Figure 7. Pixel-wise Pearson correlation for each model, computed for the whole year of 2018.Figure 8. Pixel-wise RMSE for each model, computed for the whole year of 2018. The dynamic range is shared for all the panels, with a fixed maximum value to facilitate the visual comparison.
更多请参考 DL4DS—Deep learning for empirical downscaling[1]
