Abstract:In recent decades, satellites have played a pivotal role in observing ocean dynamics, providing diverse datasets with varying spatial resolutions. Notably, within these datasets, sea surface height (SSH) data typically exhibit low resolution, while sea surface temperature (SST) data have significantly higher resolution. This study introduces a Transfer Learning-enhanced Generative Adversarial Network (TLGAN) for reconstructing high-resolution SSH fields through the fusion of heterogeneous SST data. In contrast to alternative deep learning approaches that involve directly stacking SSH and SST data as input channels in neural networks, our methodology utilizes bifurcated blocks comprising Residual Dense Module and Residual Feature Distillation Module to extract features from SSH and SST data, respectively. A pixelshuffle module-based upscaling block is then concatenated to map these features into a common latent space. Employing a hybrid strategy involving adversarial training and transfer learning, we overcome the limitation that SST and SSH data should share the same time dimension and achieve significant resolution enhancement in SSH reconstruction. Experimental results demonstrate that, when compared to interpolation method, TLGAN effectively reduces reconstruction errors and fusing SST data could significantly enhance in generating more realistic and physically plausible results.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the spatial resolution of sea surface height (SSH) data by fusing heterogeneous data, especially sea surface temperature (SST) data. Specifically, the paper presents a method based on transfer - learning - enhanced generative adversarial network (TLGAN) for reconstructing high - resolution SSH fields through heterogeneous data fusion. Traditional interpolation methods often lead to loss of details and increased errors when dealing with low - resolution SSH data, while the TLGAN model aims to overcome these limitations and significantly improve the reconstruction quality of SSH data by using high - resolution SST data. ### Main Problems and Solutions 1. **Problems**: - In ocean dynamics observations, SSH data usually has a low spatial resolution, while SST data has a high spatial resolution. - Directly using interpolation methods for up - sampling SSH data will lead to loss of details and increased errors. 2. **Solutions**: - The TLGAN model is introduced. This model combines transfer learning and generative adversarial network (GAN) to extract features from low - resolution SSH data and high - resolution SST data and perform feature fusion in the latent space. - The model contains three key modules: Residual Dense Module (RDM), Residual Feature Distillation Module (RFDM), and Upscaling Block. - Through a multi - stage training strategy, the model can effectively process data in different time dimensions and generate more realistic and physically reasonable high - resolution SSH images. ### Technical Details - **Network Architecture**: - **Generator**: It contains three main modules, which are used for extracting SSH features, extracting SST features, and up - sampling respectively. - **Discriminator**: The VGG model is used as a basis to evaluate the authenticity and quality of the generated high - resolution SSH images. - **Loss Functions**: - **Content Loss**: The mean square error (MSE) is used to measure the difference between the generated image and the high - resolution image. - **Adversarial Loss**: It evaluates the ability of the generator to deceive the discriminator. - **Training Strategies**: - **Stage 1**: Use adversarial training to conduct preliminary training on the entire TLGAN model. - **Stage 2**: Extract the feature space of SST data without updating the parameters of the first stage. - **Stage 3**: Use a small number of SSH data samples to further optimize the model without updating the parameters of the previous two stages. ### Experimental Results - **Visual Effects**: The high - resolution SSH images generated by TLGAN are closer to the original high - resolution images than traditional interpolation methods, especially in strong - gradient regions. - **Quantitative Evaluation**: Evaluation is carried out using indicators such as MSE, MAE, PSNR, and SSIM. The results show that TLGAN outperforms traditional interpolation methods in all indicators. - **Application Verification**: By calculating geostrophic velocity and kinetic energy (KE), the superiority of the high - resolution SSH data generated by TLGAN in ocean dynamics applications is verified. ### Conclusion This research successfully demonstrates the effectiveness of the TLGAN model in improving the spatial resolution of SSH data by fusing SST data. Compared with traditional interpolation methods, TLGAN is not only superior in visual effects but also shows higher accuracy and physical rationality in quantitative evaluation. This method provides new tools and ideas for ocean dynamics research.

A Transfer Learning-Enhanced Generative Adversarial Network for Downscaling Sea Surface Height through Heterogeneous Data Fusion

Super-resolution Reconstruction Algorithms Based on Fusion of Deep Learning Mechanism and Wavelet

A Fast Generative Adversarial Network Combined With Transformer for Downscaling GRACE Terrestrial Water Storage Data in Southwestern China

Generative Diffusion Model-based Downscaling of Observed Sea Surface Height over Kuroshio Extension since 2000

Unpaired Satellite Precipitation Downscaling via Multiple Generative Constrains

Single Remote Sensing Image Super-Resolution Via a Generative Adversarial Network with Stratified Dense Sampling and Chain Training

Enhancing the Resolution of Satellite Ocean Data Using Discretized Satellite Gridding Neural Networks

Spatial Downscaling of Sea Surface Temperature Using Diffusion Model

A High-Resolution Land Surface Temperature Downscaling Method Based on Geographically Weighted Neural Network Regression

GSDNet: A deep learning model for downscaling the significant wave height based on NAFNet

Advancing Satellite-Derived Precipitation Downscaling in Data-Sparse Area Through Deep Transfer Learning

Spatiotemporal Data Augmentation of MODIS‐Landsat Water Bodies Using Adversarial Networks

Deep learning-based spatial downscaling and its application for tropical cyclone detection in the western North Pacific

Efficient downscaling of satellite oceanographic data with convolutional neural networks

Deep learning for ocean remote sensing: an application of convolutional neural networks for super-resolution on satellite-derived SST data

A New Satellite-Based Precipitation Downscaling Scheme for Data-sparse Areas Using Deep Learning and Transfer Learning

Generative Adversarial Models for Extreme Geospatial Downscaling

Underwater Image Enhancement Using a Multiscale Dense Generative Adversarial Network

Deep Learning for Feature-Level Data Fusion: Higher Resolution Reconstruction of Historical Landsat Archive

Remote Sensing Image Spatiotemporal Fusion Using a Generative Adversarial Network

Hierarchical attention aggregation with multi-resolution feature learning for GAN-based underwater image enhancement