Dual-Branch State-Displacement Network for Sea Surface Temperature Super-Resolution

Wankun Chen; Feng Gao; Yanhai Gan; Chuanzheng Gong; Xun Gong; Junyu Dong; Qian Du

eess.IVAug 15, 2026

Dual-Branch State-Displacement Network for Sea Surface Temperature Super-Resolution

Authors: Wankun Chen, Feng Gao, Yanhai Gan, Chuanzheng Gong, Xun Gong, Junyu Dong, Qian Du

Organizations: State Key Laboratory of Physical Oceanography, and also with the Frontiers Science Center for Deep Ocean Multispheres and Earth System, Ocean University of China, Qingdao 266100, China · State Key Laboratory of Physical Oceanography, Institute of Oceanographic Instrumentation, Shandong Academy of Sciences, Qingdao 266100, China · Department of Electrical and Computer Engineering, Mississippi State University, Starkville, MS 39762 USA

Abstract

Sea surface temperature (SST) is a critical indicator of global climate change, yet satellite-derived SST imagery often suffers from coarse spatial resolution, limiting the ability to capture fine-scale thermal structures such as ocean fronts. To address this, we propose a Dual-Branch State-Displacement Network (DBSD-Net) for SST super-resolution. DBSD-Net adopts a dual-branch architecture: a wavelet frequency branch that explicitly separates low and high-frequency components via discrete wavelet transform for targeted processing, and a VGGUNet branch that extracts multi-scale semantic features from a frozen pre-trained VGG backbone. Within the wavelet branch, we introduce a Structural State Space Module (SSSM) with a Gated Structure Refinement (GSR) unit to efficiently capture long-range dependencies and enhance structural integrity, and a Displacement Gate Module (DGM) that learns a displacement field for geometry-aware modulation of high-frequency details, thereby mitigating spatially varying degradation. Experiments on multiple public SST datasets demonstrate that DBSD-Net outperforms existing state-of-the-art methods.

Figures & tables

Fig. 1: The overview of the Dual-Branch State-Displacement (DBSD) framework proposed for the SR of SST data is presented. (a) Illustrates the overall architecture of DBSD. DBSD consists of multiple stacked Wavelet Gate-state Groups (WGSG) branch and a VGGUNet++ branch. (b) Shows the structure of the SSSM block, which is composed of the VSSM block and the GSR module. (c) Displays the details of the VSSM block. (d) Shows the details of 2D-SSM.

Fig. 2: Details of the Displacement Gate Module (DGM). The DGM learns a compact displacement field, maps it to bilinear interpolation coefficients, and generates channel‑wise gating signals for geometry‑aware modulation of high‑frequency sub‑bands.

Fig. 3: Details of the VGGUNet++ Module. The VGGUNet++ extracts features from four stages of a frozen VGG19, compresses them with grouped convolutions, and fuses them through dense nested skip connections with residual refinement for hierarchical multi‑scale semantic representation.

Fig. 4: The influence of loss functions under different coefficients while training.

Configuration	PSNR (dB) $\uparrow$	SSIM $\uparrow$	Params (M)
PT-Frozen	46.57	0.9909	30.59
Random-Frozen	46.53	0.9909	30.59
PT-Trainable	46.52	0.9908	43.80
Random-Trainable	46.51	0.9907	43.80

TABLE I: Ablation study of VGG19 initialization and trainability on the GHRSST dataset.

Number of Channels	PSNR $\uparrow$	SSIM $\uparrow$	FLOPs (G)	Params (M)
32	43.14	0.9713	2.31	2.39
48	44.39	0.9816	5.74	5.27
64	45.23	0.9853	13.32	9.29
96	46.84	0.9918	52.03	20.70
128	47.35	0.9930	147.48	36.68
144	47.39	0.9941	228.53	46.30

TABLE II: Impact of channel number on DBSD performance on GHRSST dataset.

Number of WGSG	PSNR $\uparrow$	SSIM $\uparrow$	FLOPs (G)	Params (M)
6	46.94	0.9922	74.29	18.42
8	46.76	0.9925	98.69	24.51
10	47.05	0.9925	123.08	30.59
12	46.87	0.9924	147.48	36.68
16	46.86	0.9922	196.27	48.85

TABLE III: Impact of the number of WGSG blocks in the proposed DBSD on GHRSST dataset.

Method	FLOPs	Params	Memory	Latency	Throughput
	(G)	(M)	(MiB)	(ms)	(images/s)
EDSR [ 71 ]	18.30	1.37	90.5	1.55	1586.66
CRAN [ 72 ]	73.31	8.00	167.0	8.32	480.77
ELAN [ 43 ]	29.57	1.43	138.5	22.13	180.75
DAT [ 73 ]	52.08	5.22	269.8	21.89	182.77
SAFM [ 74 ]	51.40	5.56	189.6	5.31	553.52

TABLE IV: Comparison of the super-resolution performance efficiency of different methods on simulated input, with input size of

[4,3,48,48]

, scaling factor of

\times 4

, and averaging over 100 iterations.

Fig. 5: Qualitative analysis of different models on the HYCOM dataset with SR scale of 4, four representative regions selected from the validation dataset, each region intercepted a fixed size of data, and scaled to the same size, in order to best observe the details.

Method	SR scale	NP		AGM		EWP		NIO
Method	SR scale	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
EDSR [ 71 ]	×2	36.95	0.9837	45.63	0.9893	43.93	0.9928	37.64	0.9813
	×3	31.34	0.9651	39.72	0.9812	37.25	0.9736	33.89	0.9685
	×4	29.77	0.9341	38.57	0.9771	32.16	0.9332	32.31	0.9465
CRAN [ 72 ]	×2	37.63	0.9871	45.96	0.9903	44.43	0.9950	37.92	0.9853
	×3	31.34	0.9651	39.72	0.9812	37.25	0.9736	33.89	0.9685

TABLE V: Super-resolution performance of different methods on the HYCOM Dataset. PSNR and SSIM are calculated on the test set. Red : best, Blue : second-best.

Fig. 6: Qualitative analysis of different models on the OISST dataset with SR scale of 4, randomly selected two regions in the test dataset, each region intercepted a fixed size of data, and scaled to the same size in order to best observe the details.

Scale		EDSR [ 71 ]	CRAN [ 72 ]	ELAN [ 43 ]	DAT [ 73 ]	SAFM [ 74 ]	MambaIR [ 20 ]	DBSD-Net
$\times$ 2	PSNR $\uparrow$	41.83	42.14	46.88	42.74	45.18	46.31	48.86
$\times$ 2	SSIM $\uparrow$	0.9887	0.9905	0.9956	0.9913	0.9933	0.9943	0.9959
$\times$ 3	PSNR $\uparrow$	36.36	37.42	38.80	37.72	39.59	39.02	39.66
$\times$ 3	SSIM $\uparrow$	0.9757	0.9793	0.9869	0.9806	0.9865	0.9862	0.9896
$\times$ 4	PSNR $\uparrow$	34.84	35.11	35.28	33.15	35.66	35.87	36.08
$\times$ 4	SSIM $\uparrow$	0.9566	0.9578	0.9679	0.9568	0.9669	0.9705	0.9731

TABLE VI: Super-resolution performance of different methods on the OISST Dataset. PSNR and SSIM are calculated on the test set. Red : best, Blue : second-best.

Fig. 7: Qualitative analysis of different models on the GHRSST dataset with SR scale of 4, randomly selected two regions in the test dataset, each region intercepted a fixed size of data, and scaled to the same size in order to best observe the details.

Scale		EDSR [ 71 ]	CRAN [ 72 ]	ELAN [ 43 ]	DAT [ 73 ]	SAFM [ 74 ]	MambaIR [ 20 ]	DBSD-Net
$\times$ 2	PSNR $\uparrow$	49.36	48.37	51.57	51.13	50.20	52.42	54.05
$\times$ 2	SSIM $\uparrow$	0.9911	0.9960	0.9971	0.9973	0.9942	0.9973	0.9982
$\times$ 3	PSNR $\uparrow$	45.46	46.22	49.87	48.97	48.42	48.56	50.66
$\times$ 3	SSIM $\uparrow$	0.9941	0.9926	0.9958	0.9958	0.9934	0.9955	0.9967
$\times$ 4	PSNR $\uparrow$	42.99	44.21	46.79	46.06	45.11	47.46	47.92
$\times$ 4	SSIM $\uparrow$	0.9721	0.9901	0.9921	0.9909	0.9875	0.9931	0.9937

TABLE VII: Super-resolution performance of different methods on the GHRSST Dataset. PSNR and SSIM are calculated on the test set. Red : best, Blue : second-best.

Region	Metric	EDSR [ 71 ]	CRAN [ 72 ]	ELAN [ 43 ]	DAT [ 73 ]	SAFM [ 74 ]	MambaIR [ 20 ]	DBSD-Net
EWP	Grad-MAE $\downarrow$	7.8649	7.6997	7.6383	7.6207	7.6137	7.6139	7.6095
	Grad-Corr $\uparrow$	0.5785	0.6033	0.6133	0.6190	0.6197	0.6205	0.6212
	Front-F1 $\uparrow$	0.5588	0.5793	0.5818	0.5784	0.5798	0.5785	0.5794
	FCD $\downarrow$	1.5558	1.5387	1.5526	1.5536	1.5408	1.5571	1.5555
AGM	Grad-MAE $\downarrow$	49.1438	48.4802	48.1875	47.9735	47.9740	47.9335	47.9028
	Grad-Corr $\uparrow$	0.6186	0.6410	0.6460	0.6511	0.6506	0.6524	0.6532

TABLE VIII: The detail-aware reconstruction performance of different super-resolution models under the HYCOM dataset (

\times

4 upscaling). Red : best, Blue : second-best.

Case	SSSM	DGM	VGGUNet++	NIO		EWP		NP		AGM
Case	SSSM	DGM	VGGUNet++	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Case 1	$\checkmark$	$\times$	$\times$	43.21	0.9801	41.74	0.9846	36.27	0.9532	32.99	0.9470
Case 2	$\times$	$\checkmark$	$\times$	42.97	0.9808	41.40	0.9833	36.16	0.9523	32.91	0.9453
Case 3	$\times$	$\times$	$\checkmark$	42.78	0.9766	41.45	0.9813	35.92	0.9469	32.71	0.9372
Case 4	$\times$	$\checkmark$	$\checkmark$	43.53	0.9826	41.80	0.9853	36.39	0.9566	33.05	0.9490
Case 5	$\checkmark$	$\times$	$\checkmark$	43.53	0.9828	41.79	0.9854	36.41	0.9571	33.09	0.9497

TABLE IX: Ablation study of the SSSM, DGM, and VGGUNet modules. The PSNR and SSIM are reported on GHRSST datasets.

Remote Sensing Image Super-Resolution

Remote sensing image super-resolution reconstructs fine spatial detail in satellite and aerial imagery from lower-resolution observations.

See all 35 papers

Deep Learning-Based Statistical Downscaling of Sea Surface Temperature Using a Residual Corrective Neural Network
Aug 9, 2026Onkar Jadhav, Tim French, Ivica Janekovic +2Climate Science Residual Learning
School of Earth and Oceans, UWA Oceans Institute, University of Western Australia, Crawley, WA, Australia. · School of Physics, Maths and Computing, Computer Science and Software Engineering, University of Western Australia, Crawley, WA, Australia. · School of Engineering, UWA Oceans Institute, University of Western Australia, Crawley, WA, Australia.
When Earth Foundation Models Meet Diffusion: An Application to Land Surface Temperature Super-Resolution
Apr 18, 2026Yiheng Chen, Zihui Ma, Peishi Jiang +6Diffusion Model Guidance Remote Sensing Image Super-Resolution
University of Alabama · Emory University · University of Michigan +2
Physics Encoded Spatial and Temporal Generative Adversarial Network for Tropical Cyclone Image Super-resolution
Feb 19, 2026Ruoyi Zhang, Jiawei Yuan, Lujia Ye +2Remote Sensing Image Super-Resolution Generative Adversarial Networks
Nanjing University of Information Science and Technology. · University of Reading. · Shanghai Typhoon Institute, China Meteorological Administration.

Aug 9, 2026J/K move · Enter open · S save

Deep Learning-Based Statistical Downscaling of Sea Surface Temperature Using a Residual Corrective Neural Network

Onkar Jadhav, Tim French, Ivica Janekovic, Nicole L. Jones+1

School of Earth and Oceans, UWA Oceans Institute, University of Western Australia, Crawley, WA, Australia. · School of Physics, Maths and Computing, Computer Science and Software Engineering, University of Western Australia, Crawley, WA, Australia. · School of Engineering, UWA Oceans Institute, University of Western Australia, Crawley, WA, Australia.

Climate Science Residual Learning Convolutional Neural Networks Climate Downscaling

Open paper AskPDF arXiv

The large-scale oceanic and atmospheric forecasts provided by global climate models typically lack sufficient resolution to accurately capture the response of the coastal ocean to atmospheric forcing and coastal circulation that drive fine-scale SST variability. Dynamical downscaling is computationally prohibitive, when applied to extensive coastlines, predictive ensembles, or long time periods. Therefore, this work presents a statistical downscaling of sea surface temperature (SST) from the seasonal coupled ocean-atmosphere forecast system (ACCESS-S2) using machine learning techniques. This study proposes a novel deep learning framework that uses a U-Net to generate an initial high-resolution SST estimate, which is subsequently refined using a residual corrective approach. The target SST fields are derived from the Regional Ocean Modeling System (ROMS). This two step approach called Residual Corrective Neural Network (RCNN) progressively refines initial U-Net predictions by incorporating dynamically scaled residuals at each step, enabling accurate capture of broad patterns and fine-grained features such as eddies and fronts. We also introduce a custom loss-assisted RCNN variant to improve performance during extreme events, which may be absent from training data due to climate-driven shifts in SST extremes. The framework efficiently downscales SST along the west coast of Australia. A 2011 marine heatwave case study shows that the RCNN improves ACCESS-S2 SST predictions by increasing horizontal resolution from 25 km to 2 km, enabling identification of fine-scale anomalies unresolved in the ACCESS-S2 dataset. This balance between computational efficiency and accuracy supports applications in coastal impact assessment and marine ecosystem studies.

Abstract

Remote Sensing Image Super-Resolution

Explore similar work

Deep Learning-Based Statistical Downscaling of Sea Surface Temperature Using a Residual Corrective Neural Network

When Earth Foundation Models Meet Diffusion: An Application to Land Surface Temperature Super-Resolution

Physics Encoded Spatial and Temporal Generative Adversarial Network for Tropical Cyclone Image Super-resolution