refs

Author: simon-donike

paper/paper.md

@@ -182,7 +182,7 @@ This work has been supported by the European Space Agency (ESA) $\Phi$-Lab, with
 # Appendix
 ## Appendix A – Architecture and Training Components
 
-**Table A1. Implemented generator types and their characteristics.** {#tbl:arch}
+Table A1. **Implemented generator types and their characteristics.** {#tbl:arch}
 
 | **Generator Type** | **Description** |
 |:-------------------|:----------------|
@@ -194,7 +194,7 @@ This work has been supported by the European Space Agency (ESA) $\Phi$-Lab, with
 | `cgan` [@cgan]| Stochastic Conditional Generator with `NoiseResBlock`. |
 
 
-**Table A2. Implemented discriminator types and their purposes.** {#tbl:disc}
+Table A2. **Implemented discriminator types and their purposes.** {#tbl:disc}
 
 | **Discriminator Type** | **Description** |
 |:-----------------------|:----------------|
@@ -203,7 +203,7 @@ This work has been supported by the European Space Agency (ESA) $\Phi$-Lab, with
 | `esrgan` [@rrdb] | ESRGAN discriminator with configurable base channels and linear head size to complement RRDB generators. |
 
 
-**Table A3. Implemented training features for stable adversarial optimization.** {#tbl:train}
+Table A3. **Implemented training features for stable adversarial optimization.** {#tbl:train}
 
 | **Feature** | **Description** |
 |:-------------|:----------------|
@@ -220,7 +220,7 @@ This work has been supported by the European Space Agency (ESA) $\Phi$-Lab, with
 | `Training.gpus` | Enables distributed data-parallel training when multiple GPU indices are listed, scaling training efficiently via PyTorch Lightning. |
 
 
-**Table A4. Supported loss components and configuration parameters.** {#tbl:loss}
+Table A4. **Supported loss components and configuration parameters.** {#tbl:loss}
 
 | **Loss Type** | **Description** |
 |:---------------|:----------------|
@@ -235,7 +235,7 @@ This work has been supported by the European Space Agency (ESA) $\Phi$-Lab, with
 
 During training, scalar metrics are continuously logged in **Weights & Biases**. These indicators quantify loss dynamics, adversarial balance, and stability. Table B1 summarises the most relevant internal metrics tracked by *OpenSR-SRGAN*.
 
-**Table B1. Key internal metrics tracked during training and validation.** {#tbl:metrics}
+Table B1. **Key internal metrics tracked during training and validation.** {#tbl:metrics}
 
 | **Metric** | **Description and Expected Behaviour** |
 |:-----------|:--------------------------------------|
@@ -269,7 +269,7 @@ Qualitative results show sharper fields, buildings, and roads compared to bicubi
 
 ![False-color visual comparison for 4× RGB SR on SEN2NAIP. Left to right: LR input, model output, HR reference.](figures/rgb_example.png){#fig:exp1}
 
-**Table C1. Configuration summary for the SEN2NAIP RGB experiment.** {#tbl:exp1config}
+Table C1. **Configuration summary for the SEN2NAIP RGB experiment.** {#tbl:exp1config}
 
 | **Parameter** | **Setting** |
 |:---------------|:------------|
@@ -280,7 +280,7 @@ Qualitative results show sharper fields, buildings, and roads compared to bicubi
 | Training schedule | Pretrain 150k steps; Ramp 50k steps; EMA β = 0.999 |
 | Hardware | Dual A100 (DDP), 16-bit precision |
 
-**Table C2. Validation performance of the SEN2NAIP RGB experiment (4×).** {#tbl:exp1results}
+Table C2. **Validation performance of the SEN2NAIP RGB experiment (4×).** {#tbl:exp1results}
 
 | **Model** | **PSNR↑** | **SSIM↑** | **LPIPS↑** | **SAM↓** |
 |:-----------|:----------:|:----------:|:-----------:|:----------:|
@@ -299,7 +299,7 @@ Figure C2 shows sharper edges and preserved spectral structure relative to bicub
 
 ![Visual comparison for 8× multispectral SR (6-band Sentinel-2). Left to right: LR input, model output, HR reference.](figures/swir_example.png){#fig:exp2}
 
-**Table C3. Configuration summary for the 6-band Sentinel-2 experiment.**
+Table C3. **Configuration summary for the 6-band Sentinel-2 experiment.**
 
 | **Parameter** | **Setting** |
 |:---------------|:------------|
@@ -310,7 +310,7 @@ Figure C2 shows sharper edges and preserved spectral structure relative to bicub
 | Training schedule | Pretrain 100k steps; Ramp 40k steps; EMA disabled |
 | Hardware | Dual A100 GPU, 32-bit precision |
 
-**Table C4. Validation performance of the 6-band Sentinel-2 experiment (8×).**
+Table C4. **Validation performance of the 6-band Sentinel-2 experiment (8×).**
 
 | **Model** | **PSNR↑** | **SSIM↑** | **LPIPS↑** | **SAM↓** |
 |:-----------|:----------:|:----------:|:-----------:|:----------:|