🎈 News

[2025.9.15] Training and testing code released.

[2026.5.29] Our paper was accepted by Neurocomputing.(Paper link: https://www.sciencedirect.com/science/article/pii/S0925231226015328)

⭐ Abstract

Salient Object Detection in Optical Remote Sensing Images (ORSI-SOD) is vital for applications such as urban planning and disaster monitoring. Yet, existing deep models remain energy-intensive and unsuitable for edge deployment. We tackle this challenge by proposing a Low-Energy Spiking Network Neural for Remote Sensing Saliency(LENS-Net), the first fully spiking neural network for ORSI-SOD. LENS-Net employs a Spike-driven Transformer v3 encoder to extract multi-scale features with low energy cost, and a Spike Multi-scale Attention Decoder that fuses contextual cues via spike-driven channel attention and up-convolution, ensuring effective saliency representation under sparse computation. Moreover, a {sigmoid-based soft surrogate gradient} replaces hard truncation, stabilizing training and enhancing boundary recognition in complex scenes. Across three benchmark datasets (ORSSD, EORSSD, and ORI-4199), LENS-Net demonstrates outstanding performance while maintaining high energy efficiency, outperforming all lightweight ANN counterparts. For instance, on the ORSSD dataset under a timestep T=4 during inference, it achieves a $S_{\alpha}$ of 92.79%, an $MAE$ of 0.0109, and consumes only 11.48 mJ of energy. These results establish an efficient, low-energy solution for practical ORSI-SOD deployment.

🚀 Contribution

• First SNN-based ORSI-SOD Method
  We present the first SNN-based method for ORSI-SOD, filling an important gap and establishing a strong baseline for future neuromorphic research in remote sensing.
• Novel Multi-scale Spiking Decoder (SpikeMAD)
  We design a novel multi-scale spiking decoder termed as SpikeMAD, which incorporates membrane potential residual connections and a multi-scale fusion strategy to achieve efficient and effective feature integration in remote sensing imagery.
• Soft-Clip Spike Firing Approximation
  We propose a soft-clip spike firing approximation backpropagation function that ensures smooth gradient transitions, eliminates discontinuities from hard truncation, and enhances boundary recognition in salient object detection.
• Extensive Experimental Validation
  Extensive experiments on three benchmark datasets demonstrate that LENS-Net achieves superior accuracy compared to all lightweight ANN counterparts, while consuming significantly less energy (e.g., only 11.61 mJ on ORSSD with T=4).

📆 TODO

• Release code

🎮 Getting Started

1. Install Environment

conda create -n LENS_Net python=3.8
conda activate LENS_Net
conda install pytorch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 -c pytorch

2. Prepare Datasets

• Download datasets: EORSSD, ORSSD, ORI-4199 from this link.

3. Train the LENS-Net

python train.py

3. Test the LENS-Net

python inference.py
python metric.py

🖼️ Comparative experiment of LENS-Net on three datasets.

🖼️ Ablation experiment of LENS-Net on ORSSD dataset.

✨ Visualization of saliency prediction maps by different methods on the EORSSD dataset

✨ Visualization ablation studies on the ORSSD dataset

💡 Layer-wise average spiking firing rates of LENS-Net on the EORSSD test dataset(600 images)

😄Checkpoint

A checkpoint can be downloaded from Google Drive.

🔧Reference.

Zhai L, Pietroń M, Corizzo R, et al. LENS-Net: Low-energy spiking neural network for remote sensing saliency[J]. Neurocomputing, 2026, 697: 134134.