Abstract: High-resolution ocean remote sensing images are critical for oceanographic research, yet available data often suffers from low resolution due to temporal averaging or cloud interference. This paper presents a modified Very Deep Super-Resolution (VDSR) model specifically optimized for satellite-derived Sea Surface Temperature (SST) images. By integrating features from multiple network depths, the proposed architecture enhances the reconstruction of complex ocean structures like coastal fronts. Evaluated on several AVHRR-based databases, the model demonstrates superior performance in both Peak Signal-to-Noise Ratio (PSNR) and perceptual loss compared to state-of-the-art models and traditional interpolation methods.
Introduction
In oceanography, high-resolution imagery is essential for studying air-sea interactions, weather prediction, and marine ecosystems. However, researchers often face a "resolution gap." Satellite data is frequently provided as weekly or monthly averages, which smooths out fine-grained details. Furthermore, "instant" satellite images are often plagued by cloud contamination, necessitating the use of lower-resolution composite products.
This research focuses specifically on ocean fronts—the boundaries between water masses with different physical properties like temperature and salinity. These fronts are vital for biodiversity, yet they are difficult to study when the resolution is low. While general-purpose super-resolution models exist for natural photos, they often struggle with the unique, turbulent textures of ocean temperature data, particularly in complex coastal regions.
Methodology
The authors build upon the Very Deep Super-Resolution (VDSR) model, a 20-layer convolutional neural network (CNN) that uses residual learning to predict the difference between a blurry low-resolution image and its high-resolution counterpart.
The "Modified" part of the approach involves integrating feature maps from multiple layers of the network. While deep layers capture abstract features, low-level layers contain the most reliable raw information from the input. By bridging these different levels of abstraction, the model can better reconstruct the intricate, non-linear patterns found in coastal waters.
To evaluate the results, two primary metrics were used:
- PSNR (Peak Signal-to-Noise Ratio): Measures pixel-to-pixel accuracy.
- Perceptual Loss: Uses a pre-trained VGG19 network to ensure the spatial structure and content of the reconstructed image are perceptually accurate, rather than just matching raw pixel values.
Datasets
The study utilized several specialized datasets derived from Advanced Very High Resolution Radiometers (AVHRR):
- China Ocean SST Database: 600 daily averaged images covering China’s coastal waters.
- Ocean SST Database: 1,500 images covering a broader range of the Pacific.
- SST Ocean-Front Database: 1,500 images specifically processed to highlight ocean fronts using the Microcanonical Multiscale Formalism (MMF) method.
- Set291: A standard collection of natural images used as a baseline for non-satellite data performance.
Key Results
Quantitative improvements were most dramatic in challenging coastal water tests. The modified VDSR achieved a PSNR of 21.53 dB and a remarkably low Perceptual Loss of 5.89, significantly outperforming the standard VDSR (17.67 dB PSNR, 53.77 Perceptual Loss) and traditional bicubic interpolation (17.04 dB PSNR, 21.07 Perceptual Loss).
Qualitatively, visual comparisons demonstrate that the modified model successfully eliminates the "blur" around coastal edges and complex fronts. This provides a much more accurate map for oceanographers to locate and study dynamic marine structures.
Discussion and Future Work
A key insight from this study is that domain specificity matters. Interestingly, the authors found that the EDSR (Enhanced Deep Super-Resolution) model—which typically outperforms VDSR on natural images—did not perform as well on SST data. This suggests that "bigger is not always better" when dealing with specialized geophysical data, and architecture must be tuned to the specific textures of the domain.
Future work will focus on training with real-world pairs of low-resolution and high-resolution satellite images rather than simulated downsampled data. Additionally, the team plans to apply this architecture to other remote sensing data types, such as chlorophyll concentration and sea surface height.
Summary updated