Knowledge about the impact craters on rocky planets is crucial for understanding the evolutionary history of the universe. Compared to traditional visual interpretation, deep learning approaches have improved the efficiency of crater detection. However, single-source data and divergent data quality limit the accuracy of crater detection. In this study, we focus on valuable features in multi-modal remote sensing data from Chang’e lunar exploration mission and propose an Attention-based Dual-branch Segmentation Network (ADSNet). First, we use ADSNet to extract the multi-modal features via a dual-branch encoder. Second, we introduce a novel attention for data fusion where the features from the auxiliary modality are weighted by a scoring function and then being fused with those from the primary modality. After fusion, the features are transferred to the decoder through skip connection. Lastly, high-accuracy crater detection is achieved based on the learned multi-modal data features through semantic segmentation. Our results demonstrate that ADSNet outperforms other baseline models in many metrics such as IoU and F1 score. ADSNet is an effective approach to leverage multi-modal remote sensing data in geomorphological feature detection on rocky planets in general.

中文翻译：

---

基于嫦娥像的双分支多模态收敛网络


了解岩石行星上的撞击坑对于理解宇宙的演化历史至关重要。与传统的视觉解释相比，深度学习方法提高了陨石坑检测的效率。然而，单一来源数据和不同的数据质量限制了陨石坑检测的准确性。在本研究中，我们重点关注嫦娥探月任务多模态遥感数据中的宝贵特征，并提出了一种基于注意力的双分支分割网络 （ADSNet）。首先，我们使用 ADSNet 通过双分支编码器提取多模态特征。其次，我们引入了一种新的数据融合关注，其中来自辅助模态的特征由评分函数加权，然后与来自主要模态的特征融合。融合后，特征通过 skip 连接传输到解码器。最后，基于学习到的多模态数据特征，通过语义分割实现高精度的火山口检测。我们的结果表明，ADSNet 在 IoU 和 F1 分数等许多指标上都优于其他基线模型。ADSNet 是利用多模态遥感数据进行岩石行星地貌特征检测的有效方法。