As mentioned in our introduction page, DL4DS allows combining different backbones, and modules to create different configurations. The following example from the original paper demonstrates how those combinations can affect inference accuracy. 

For example, table 1 shows eight different combination of different numerically downscaled input data, backbone type, and loss functions used for experiments. The second column column of the table shows the method used for downscaling the training data. The third column shows the machine learning paradigm used, and the fourth column shows the backbone. The last column also shows the loss functions used with different settings. 

Table 1. DL4DS model showcase. 

Next, Figure 1 shows the some sample input data used for downscaling. Figure 1(a) shows the low-resolution surface NO2 concentration from the CAMS global reanalysis model. Figure 1(b) shows a numerically obtained downscaled sample; in this case, bi-cube interpolation method is used. It can be seen that the image is too smooth and do not capture all the features. Figure 1(c) shows the high-resolution data from the CAMS regional reanalysis. It is clear that downscaling is not possible with simple numeric method, it requires complex reanalysis, which is time consuming.

Figure 1. Reference data, NO2 surface concentration. (a) A low-resolution sample. (b) Bi-cube numerical interpolation. (c) High-resolution CAMS data for regional reanalysis. 

Difference of accuracies

Figure 2 shows the results of applying different learning configurations from Table 1.  Eight different combinations are used in Table 1 and they are labeled from A to H, corresponding results are shown in Figure 2.  It can be seen that different learning configuration produces different down-scaled images. Therefore, accuracy metrics are calculated and shown in Table 2.    

Figure 3. Examples of downscaled images, obtained from different learning configurations mentioned in Table 1. 

Next, Table 2 shows the accuracy of different learning configurations.  The metrics used are Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Pearson correlation coefficient (RearCorr),  Structural Similarity Index Measure (SSIM), and Peak Signal-to-noise Ratio (PSNR). The results show that, the learning configuration E has the best values for MAE, RMSE, and PearCorr, whereas configuration D has the best values for SSIM and PSNR values. Thus, different learning configuration provide different downscaled image quality. 

Table 2. Accuracy matrix for different downscaled configuration.