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

• Haze attenuates high-frequency edges and textures, pushing weak responses below fixed spike thresholds and progressively losing structural information with depth.Repeated spike quantization and membrane leakage further suppress weak responses.Source: Introduction · p. 1
• EM-SNN achieves competitive dehazing performance while significantly reducing energy consumption compared with existing approaches.TM-LIF and SSM jointly modulate activation scales and structural representations.Source: Introduction · p. 2
• A statistics-driven, channel-wise adaptive TM-LIF threshold calibrates spike quantization to mitigate information loss from haze-induced contrast compression.It is designed to stabilize multispike representations.Source: Introduction · p. 2
• SSM uses Sobel-derived structural priors to compensate for depth-wise attenuation of weak details during deep spiking feature propagation.The module is described as deployment-friendly and energy-aware.Source: Introduction · p. 2

Single Image Dehazing

• Hand-crafted dehazing priors can falter when their assumptions fail; deep models improve representation, while remote-sensing approaches use spatial-frequency fusion or diffusion restoration.Lightweight alternatives may sacrifice representation capacity, whereas strong models often require substantial computation.Source: Single Image Dehazing · p. 2

Spiking Neural Networks

• SNNs have been applied to vision tasks such as detection, deraining, and super-resolution, but dense restoration and atmospheric-scattering degradation remain underexplored.Prior SNN dehazing methods largely use spiking-inspired approximations rather than fully spike-driven restoration pipelines.Source: Spiking Neural Networks · p. 2

Overall Architecture

• EM-SNN uses a multiscale SRB encoder–decoder with channel-wise concatenation for skip fusion and a PRB to reconstruct decoded spike features.A 3 × 3 convolution produces shallow input features, and a final 3 × 3 convolution produces the output.Source: Overall Architecture · p. 2, Figure 2 · p. 3

Threshold-Modulated LIF (TM-LIF)

• TM-LIF calibrates channel-wise firing thresholds from running membrane-potential dispersion statistics, then freezes them for deterministic inference.The statistic is computed across batch and spatial dimensions; the channel-specific thresholds are conditioned on the training distribution and remain fixed across test samples.Source: Threshold-Modulated LIF (TM-LIF) · pp. 3–4

Spike Sobel Modulation (SSM)

• SSM uses fixed Sobel gradients as a structural prior for channel and spatial modulation, selectively enhancing edge and texture responses with lightweight overhead while preserving event-driven sparsity.The structural prior highlights edge- and texture-dominant regions to guide modulation.Source: Spike Sobel Modulation (SSM) · pp. 4–5

Experimental Setup

• Evaluation covers synthetic HRSD and SateHaze1K and real-world RICE and RRSHID; models use 256 × 256 training crops and are trained for 1000 epochs.All experiments run on an NVIDIA RTX 4090 GPU; FLOPs and SOPs are reported at 256 × 256 input resolution.Source: Experimental Setup · p. 5

Experimental Results

• On SateHaze1K, EM-SNN restores clearer structures and sharper boundaries with fewer residual haze artifacts and color distortions across varying haze densities.Figure 4 shows enlarged patches and absolute error maps against ground truth; darker blue indicates lower error.Source: Figure 4 · p. 6
• EM-SNN reaches SOTA on LHID, RRSHID, and RICE2 versus ANN methods, with higher PSNR on DHID (28.83 dB) and RICE1 (35.99 dB) than (Cong et al. 2024).The RRSHID comparison spans haze densities.Source: Experimental Results · pp. 5–6, Table 3 · p. 6
• 25% of the FLOPs-to-energy-converted upper bound for SFRDP-Net is consumed by EM-SNN, whose average energy consumption is 43.62 mJ.Energy is evaluated across datasets; the reported reference is the upper bound for SFRDP-Net.Source: Experimental Results · pp. 6–7, Table 4 · p. 6

Ablation Study

• With TM-LIF and SSM both active, PSNR gains +1.03 dB and SSIM gains +0.0074 while energy falls by 1.49 mJ versus the baseline.The factorial ablation averages results over SateHaze1K's thin, moderate, and thick haze subsets.Source: Ablation Study · p. 7, Table 5 · p. 7
• With SSM fixed, TM-LIF improves PSNR/SSIM from 24.80/0.9033 to 25.63/0.9093 and reduces energy from 48.36 to 46.27 mJ; removing SSM lowers restoration quality.The lower-group ablation also reports reduced energy but degraded quality when the structural prior is removed or Sobel is replaced by Scharr.Source: Ablation Study · p. 7, Table 6 · p. 7
• Thresholds are lower and more concentrated in shallow layers, while heavier-haze training expands their range and produces a more pronounced high-threshold tail.The learned frozen channel-wise thresholds reflect distribution-dependent shifts in membrane-potential scales, rather than sample-wise adaptation during inference.Source: Ablation Study · p. 7, Figure 5 · p. 7
• Increasing time steps T or spike depth D generally improves reconstruction but raises energy; the default (T=1, D=4) gains +0.90 dB/+0.0095 over (T=1, D=1) at +15.86 mJ.The lightweight (T=1, D=1) setting yields 24.73 dB/0.8998 at 30.41 mJ.Source: Ablation Study · p. 7, Table 7 · p. 7

Conclusion

• EM-SNN combines a statistics-driven TM-LIF neuron and SSM to counter haze-induced contrast compression and spike-generation discretization, supporting structural preservation during deep spiking propagation.The paper reports experiments on public dehazing datasets.Source: Conclusion · p. 7
• Experiments show competitive restoration quality while maintaining sparse spiking activity and significantly reducing energy consumption.The conclusion does not specify a numerical energy reduction here.Source: Conclusion · p. 7
• Future work will address robustness under more diverse real-world sensing conditions and extend deployment to neuromorphic hardware platforms.Source: Conclusion · p. 7