Scholarlink

This study presents FedSACNN, a federated convolutional neural network with spatial attention mechanisms, specifically designed for Martian terrain hazard segmentation from decentralized rover images. FedSACNN is the first interplanetary federated learning (FL) system combined with Squeeze-and-Excitation (SE) with a lightweight U-Net backbone enhanced by Atrous Spatial Pyramid Pooling (ASPP) and a Spatial Attention Module, enabling multi-scale feature extraction while emphasizing hazard-predictive regions. The framework allows for several Mars rovers to collaboratively train a shared model without transmitting raw images, thereby ensuring data privacy and significantly reducing communication overhead, both of which are necessary restrictions in spaceborne contexts. Our model enhances the network to pick out hazard-critical features such as slopes and rocks while eliminating background noise. Mimicking real-world decentralized behavior, the AI4Mars dataset was split into three disjoint subsets corresponding to independent rovers. This work verifies the feasibility of the use of federated learning for the distributed surface analysis of extraterrestrial terrain. By bypassing the requirement for central data collection and low-bandwidth model synchronization, FedSACNN offers a privacy-conscious and scalable path for embedded hazard prediction in future Mars missions. Ultimately, FedSACNN is a groundbreaking step toward reliable, distributed, and autonomous terrain analysis for planetary missions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Federated Spatial Attention CNN for Predicting Surface Hazards on Mars via Distributed Rover Networks (FedSACNN)

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mahinur Rahman Tani,
  • Nusrath Jahan Shawon,
  • Maimuna Rashid,
  • Romana Rashid,
  • Sara Karim

摘要

This study presents FedSACNN, a federated convolutional neural network with spatial attention mechanisms, specifically designed for Martian terrain hazard segmentation from decentralized rover images. FedSACNN is the first interplanetary federated learning (FL) system combined with Squeeze-and-Excitation (SE) with a lightweight U-Net backbone enhanced by Atrous Spatial Pyramid Pooling (ASPP) and a Spatial Attention Module, enabling multi-scale feature extraction while emphasizing hazard-predictive regions. The framework allows for several Mars rovers to collaboratively train a shared model without transmitting raw images, thereby ensuring data privacy and significantly reducing communication overhead, both of which are necessary restrictions in spaceborne contexts. Our model enhances the network to pick out hazard-critical features such as slopes and rocks while eliminating background noise. Mimicking real-world decentralized behavior, the AI4Mars dataset was split into three disjoint subsets corresponding to independent rovers. This work verifies the feasibility of the use of federated learning for the distributed surface analysis of extraterrestrial terrain. By bypassing the requirement for central data collection and low-bandwidth model synchronization, FedSACNN offers a privacy-conscious and scalable path for embedded hazard prediction in future Mars missions. Ultimately, FedSACNN is a groundbreaking step toward reliable, distributed, and autonomous terrain analysis for planetary missions.