Table 1.

Quantitative results computed on the hold-out test dataset. Results are reported for the proposed DSen2-CR network in different configurations: trained on the proposed ${\mathcal{L}}_{\mathrm{CARL}}$ loss, trained on the plain L1 target loss ${\mathcal{L}}_{\mathrm{T}}$, and trained on ${\mathcal{L}}_{\mathrm{CARL}}$ and ${\mathcal{L}}_{\mathrm{T}}$ but without the SAR input. In the tables, Target refers to the error computed between the predicted image and the target cloud-free image. This is the loss as optimized using ${\mathcal{L}}_{\mathrm{T}}$. Reprod denotes the reproduction error, namely the error between the predicted image and the clear parts of the input image. This is part of the ${\mathcal{L}}_{\mathrm{CARL}}$ loss that is explicitly optimized. Recon is the reconstruction error, namely the error between the predicted image and the target image inside the reconstructed clouds and shadow regions.

(a) Test results on pixel-wise metrics
	MAE (${\rho }_{\mathrm{TOA}}$)			RMSE (${\rho }_{\mathrm{TOA}}$)	PSNR (dB)
Method	Target	Reprod	Recon	Target	Target
DSen2-CR on ${\mathcal{L}}_{\mathrm{CARL}}$	0.0290	0.0204	0.0266	0.0366	28.7
DSen2-CR on ${\mathcal{L}}_{\mathrm{T}}$	0.0270	0.0398	0.0266	0.0343	29.3
DSen2-CR on ${\mathcal{L}}_{\mathrm{CARL}}$ w/o SAR	0.0306	0.0188	0.0282	0.0387	27.6
DSen2-CR on ${\mathcal{L}}_{\mathrm{T}}$ w/o SAR	0.0284	0.0389	0.0281	0.0361	28.8
pix2pix	0.0292	0.0210	0.0274	0.0424	28.2

(b) Test results on spectral and structural fidelity metrics.
	SAM (°)			SSIM
Method	Target	Reprod	Recon	Target
DSen2-CR on ${\mathcal{L}}_{\mathrm{CARL}}$	8.15	3.94	8.04	0.875
DSen2-CR on ${\mathcal{L}}_{\mathrm{T}}$	8.07	6.33	8.13	0.878
DSen2-CR on ${\mathcal{L}}_{\mathrm{CARL}}$ w/o SAR	8.98	3.86	8.97	0.870
DSen2-CR on ${\mathcal{L}}_{\mathrm{T}}$ w/o SAR	8.97	6.17	9.05	0.873
pix2pix	13.68	13.93	12.67	0.844