EM-SNN: Efficiently Modulated Spiking Neural Network for Remote Sensing Image Dehazing

Jie Shao; Jiaqi Ma; Wenwen Min; Beihang Song; Ning Chen; Youfa Liu; Jun Wan

cs.CVOct 7, 2026

EM-SNN: Efficiently Modulated Spiking Neural Network for Remote Sensing Image Dehazing

Authors: Jie Shao, Jiaqi Ma, Wenwen Min, Beihang Song, Ning Chen, Youfa Liu, Jun Wan

Organizations: Zhongnan University of Economics and Law, Wuhan, China · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Yunnan University, Kunming, China · National Institute of Natural Hazards, Ministry of Emergency Management of China, Beijing, China · Wuhan University, Wuhan, China

Abstract

Although spiking neural networks (SNNs) provide an energy-efficient alternative to artificial neural networks (ANNs), their application to remote sensing image dehazing remains limited. A key challenge arises from the coupling between haze-induced high-frequency attenuation and discrete spike thresholding. This interaction suppresses weak responses and fundamentally limits the recovery of edges, textures, and fine details in spiking dehazing models. To address this challenge, we propose the Efficiently Modulated Spiking Neural Network (EM-SNN), a dedicated spiking framework tailored to remote sensing image dehazing. EM-SNN integrates a statistics-driven Threshold-Modulated Leaky Integrate-and-Fire (TM-LIF) neuron to adaptively compensate for haze-induced contrast compression, together with a Spike Sobel Modulation (SSM) module that enhances structural cues and reduces depth-wise attenuation during spiking feature propagation. By jointly modulating activation scales and structural representations, EM-SNN improves dehazing performance while preserving the inherent event-driven sparsity of SNNs. Experiments on HRSD, RICE, RRSHID, and SateHaze1K demonstrate that EM-SNN achieves competitive dehazing performance while consuming only one quarter of the energy of the strong ANN baseline SFRDP-Net.

Figures & tables

Figure 1: Visualization of haze-induced degradation and spiking responses. (a) Remote-sensing patch under clear (top) and hazy (bottom) conditions. (b) 3D intensity surface plots of the patches. (c) LIF membrane potentials over neuron index and simulation step. (d) PSNR versus energy consumption on RICE-2; bubble size indicates parameter.

Figure 2: Overall architecture of EM-SNN. EM-SNN adopts a multi-scale encoder–decoder built from SRB blocks with skip fusion via channel-wise concatenation (C), and uses PRB for reconstruction. TM-LIF is applied in all SCB units across both SRB blocks and the Down/Upsampling stages, while SSM is embedded across scales to provide structure-guided attention.

Figure 3: Frequency-domain visualizations under different neurons. The left shows clear and hazy remote sensing images with the region of interest highlighted, while the right compares ReLU, LIF, and TM-LIF in terms of spectra. We report $\Delta$ , the high-frequency energy ratio drop from clear to hazy, which indicates that haze suppresses high-frequency components. Compared with LIF, TM-LIF, via channel-wise adaptive thresholds and multi-level quantization, yields spectra closer to the continuous responses of ReLU and alleviates the haze-induced loss of high-frequency details.

Method	Venue	LHID		DHID		RICE1		RICE2		Complexity
Method	Venue	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	Param (M)	Energy (mJ)
DCP ( He et al. 2010 )	CVPR’09	15.86	0.7168	19.17	0.7561	16.03	0.6966	10.37	0.4491	–	–
DehazeNet ( Cai et al. 2016 )	TIP’16	15.91	0.7252	21.65	0.8389	20.94	0.8503	21.16	0.8412	0.008	2.48
FFA-Net ( Qin et al. 2020 )	AAAI’20	25.05	0.8800	27.79	0.8870	31.78	0.9543	26.81	0.7932	4.69	1322.64
4KDehazing ( Zheng et al. 2021b )	CVPR’21	25.73	0.8406	23.44	0.8076	27.15	0.9062	26.72	0.9140	34.55	477.94
DCI-Net † ( Zhang and Wang 2022 )	TGRS’22	26.00	0.8930	28.12	0.9080	35.03	0.9600	32.14	0.8600	26.51	123.56

Table 1: Quantitative comparison on LHID, DHID, RICE1, and RICE2. Bold and underlined indicate the best and second-best results, respectively. † Methods specifically designed for remote sensing dehazing.

Figure 4: Qualitative comparisons on SateHaze1K under varying haze densities. The orange box marks the zoom-in region. The bottom rows show enlarged patches and absolute error maps w.r.t. the ground truth (darker blue indicates lower error). Our method recovers clearer structures and sharper boundaries with fewer residual haze artifacts and color distortions.

Method	SateHaze1K-thin		SateHaze1K-moderate		SateHaze1K-thick		SateHaze1K-average
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
DCP	20.47	0.8705	20.83	0.8971	16.71	0.7435	19.34	0.8370
DehazeNet	16.82	0.8210	15.75	0.7720	15.12	0.7430	15.90	0.7787
FFA-Net	25.16	0.9168	24.79	0.9296	22.36	0.8453	24.10	0.8972
4KDehazing	23.35	0.8939	23.73	0.9209	20.61	0.8227	22.56	0.8792
DCI-Net †	24.39	0.9022	25.03	0.9315	21.82	0.8390	23.75	0.8909

Table 2: Quantitative comparison on the SateHaze1K.

Method	RRSHID-thin		RRSHID-moderate		RRSHID-thick		RRSHID-average
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
DCP	18.46	0.4564	17.80	0.4856	18.39	0.4843	18.22	0.4754
FFA-Net	17.08	0.4452	17.40	0.5385	16.71	0.4792	17.06	0.4876
GridDehazeNet	22.77	0.6145	22.62	0.6468	23.96	0.7112	23.12	0.6575
4KDehazing	22.83	0.6177	22.47	0.6505	22.55	0.6912	22.62	0.6531
SCANet	18.37	0.4718	18.11	0.5380	19.07	0.5966	18.52	0.3741

Table 3: Performance comparison on RRSHID across haze densities. Results of baseline methods follow ( Zhu et al. 2025 ) .

Dataset	FR	SOPs (G)	Energy (mJ)	Dataset	FR	SOPs (G)	Energy (mJ)
SateHaze1K-thin	0.1729	34.938	46.166	DHID	0.2235	45.149	55.356
SateHaze1K-moderate	0.1688	34.110	45.421	LHID	0.1216	24.563	36.829
SateHaze1K-thick	0.1788	36.127	47.236	RRSHID-thin	0.1450	29.298	41.090
RICE1	0.1626	32.845	44.283	RRSHID-moderate	0.1322	26.709	38.761
RICE2	0.1547	31.257	42.854	RRSHID-thick	0.1291	26.078	38.193

Table 4: Energy evaluation of EM-SNN across different datasets. FR denotes the average firing rate.

ID	TM-LIF	SSM	PSNR $\uparrow$	SSIM $\uparrow$	Energy (mJ) $\downarrow$
(1)			24.60	0.9019	47.76
(2)	✓		24.93	0.9012	49.50
(3)		✓	24.80	0.9033	48.36
(4)	✓	✓	25.63	0.9093	46.27

Table 5: Factorial ablation study of two modules on SateHaze1K.

ID	Variant	PSNR $\uparrow$	SSIM $\uparrow$	Energy (mJ) $\downarrow$
(1)	LIF + SSM	24.80	0.9033	48.36
(2)	I-LIF + SSM	25.00	0.9058	46.30
(3)	NI-LIF + SSM	25.00	0.9081	37.26
(4)	TM-LIF + SSM	25.63	0.9093	46.27
(5)	TM-LIF + w/o SSM	24.93	0.9012	49.50
(6)	TM-LIF + SSM (full)	25.63	0.9093	46.27

Table 6: Additional ablations on SateHaze1K. The upper group compares neuron designs with SSM fixed, and the lower group compares SSM variants with TM-LIF fixed. Results are averaged over the thin/moderate/thick subsets.

$T$	$D$	PSNR $\uparrow$	SSIM $\uparrow$	Energy (mJ) $\downarrow$
1	1	$24.73$	$0.8998$	$30.41$
1	2	$25.10$	$0.9076$	$40.04$
2	2	$25.37$	$0.9072$	$65.19$
1	4	$25.63$	$0.9093$	$46.27$
2	4	$25.67$	$0.9116$	$80.19$
4	4	$25.81$	$0.9135$	$158.37$

Table 7: Joint ablation of time steps

T

and spike depth

D

on SateHaze1K dataset.

Figure 5: Layer-wise distributions of frozen channel-wise thresholds $v_{\mathrm{th}}$ learned on (a) thin, (b) moderate, and (c) thick haze subsets (dark-to-light: shallow-to-deep layers).

Explore similar work

HazeSpikeMamba: Coupling Spiking-Inspired and State-Space Features for Self-Supervised Real-World Dehazing
Aug 7, 2026Haoran Liu, Huibin Li, Mingzhe Liu +2Single Image Dehazing U-Net
School of Artificial Intelligence and Electronic Engineering, Sichuan Technology and Business University, Chengdu 611745, China · Chengdu University of Technology, College of Nuclear Technology and Automation Engineering, Chengdu 610059, China · Sigray, Inc., Concord, CA 94520, USA
IncepDeHazeGAN: Novel Satellite Image Dehazing
Apr 17, 2026Tejeswar Pokuri, Shivarth RaiSingle Image Dehazing Computer Vision
No Institute Given
SpikeRestormer: Towards Energy-Efficient All-in-One Image Restoration via Unified Event Reasoning
Aug 3, 2026Shengkai Hu, Jie Shao, Jiaqi Ma +5Query-Specific Degradation-Aware Representation Spiking Neural Networks
School of Information Engineering, Zhongnan University of Economics and Law, Wuhan, China · Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates · School of Computer Science, Wuhan University, Wuhan, China +1

Aug 7, 2026, cs.CVJ/K move · Enter open · S save

HazeSpikeMamba: Coupling Spiking-Inspired and State-Space Features for Self-Supervised Real-World Dehazing

Haoran Liu, Huibin Li, Mingzhe Liu, Peng Li+1

School of Artificial Intelligence and Electronic Engineering, Sichuan Technology and Business University, Chengdu 611745, China · Chengdu University of Technology, College of Nuclear Technology and Automation Engineering, Chengdu 610059, China · Sigray, Inc., Concord, CA 94520, USA

Single Image Dehazing U-Net Adaptive Feature Fusion State Space Models Spiking Neural Networks Real-World

Open paper AskPDF arXiv

Dehazing networks are commonly trained on synthetic hazy-clear pairs, but their performance often drops on real photographs. Synthetic haze generated using the atmospheric scattering model does not fully capture the variability of real haze, and paired real hazy-clear images are scarce. In this work, we propose HazeSpikeMamba, a compact dehazing framework that combines a spiking-inspired local path and an attentive state-space global path in a multi-scale U-Net. The local path uses TPCNNSpike, a new spike-emission scheme inspired by the neighborhood coupling of Pulse-Coupled Neural Network (PCNN). Unlike grouped directional scanning, TPCNNSpike updates all neurons in parallel using the previous firing states of their Gaussian-weighted neighborhoods. The global path adapts the Attentive State-Space Module of MambaIRv2, retaining semantic prompting and sequence reordering while removing the window self-attention branch. Its state-space processing models long-range dependencies with complexity linear in sequence length. For target-domain adaptation, a frozen degradation network, pretrained on paired NH-HAZE data, re-synthesizes haze from the dehazed prediction. The reconstruction error updates only the final restoration layers of HazeSpikeMamba without haze-free labels during adaptation. A shared checkpoint is adapted once on each complete unlabeled target set, making the evaluation dataset-level and transductive rather than zero-shot or per-image optimization. The forward network contains 2.02M active parameters and requires 13.27G nominal MACs (measured with thop at 256x256 input). This adaptation consistently improves BRISQUE and NIMA on RTTS, URHI, and HSTS. On RTTS, BRISQUE decreases from 30.13 to 27.72 and NIMA increases from 4.13 to 4.87. Under this transductive protocol, the adapted model also achieves the best BRISQUE and NIMA on URHI and HSTS among the compared methods.

Abstract

Explore similar work

HazeSpikeMamba: Coupling Spiking-Inspired and State-Space Features for Self-Supervised Real-World Dehazing

IncepDeHazeGAN: Novel Satellite Image Dehazing

SpikeRestormer: Towards Energy-Efficient All-in-One Image Restoration via Unified Event Reasoning