MEOX：面向地球观测的轻量化多模态混合专家模型

Summary of Thought Process

Alright, let's break down this research summary and translate it into simplified Chinese. My primary focus is accuracy and maintaining the technical precision you'd expect from a paper in this field. I'm going to work sentence-by-sentence. First, I'll identify the key technical terms and concepts. Then, I'll draft a Chinese translation, ensuring I understand the nuances of the original English. Next, I'll refine it, paying close attention to phrasing, terminology, and overall flow to make it sound natural and professional in Chinese. I'll make sure to preserve paragraph breaks and any inline URLs, even though there aren't any in this source. I'm keeping my translation strictly to the content itself, with no added commentary. The goal is a clean, accurate, and readily understandable translation that maintains the paper's original intent. I will ensure that the Chinese terms chosen are those that are the most commonly used and that represent each term accurately.

近期对地观测表征学习的进展能够兼容异构传感器和缺失观测，但这往往依赖于更大规模的模型架构。我们提出了 MEOX（基于专家的多模态对地观测模型，Multimodal Earth Observation with eXperts），这是一种多模态掩码自编码器，其编码器包含 293.9 万个参数，总参数量仅为 311.5 万。传感器特定适配器、显式有效性信号以及共享稀疏专家模块在进行可学习的按块融合之前，保留了模态特异性处理。随后，四个元数据词元伴随单一空间序列共同通过后续的十四个编码器模块。带有专有低秩残差的共享专家投影限制了参数增长，而旋转注意力机制则支持不同于预训练尺度的下游空间网格。该模型采用模态平衡的掩码重构和结构化传感器丢弃，在 122.8 万个 MMEarth64 样本上进行了预训练。模型在 64 像素和 224 像素两种分辨率下的六项 GEO-Bench 任务上进行了冻结迁移评估。在 64 像素腰果分割任务中，该模型达到了 64.42% 的平均交并比（mIoU）；在 224 像素 EuroSAT 任务中达到了 90.56% 的平均准确率，均优于文献报道的对应 CSMoE 结果。BigEarthNet 微调达到了 72.95% 的微平均精度。路由诊断区分了专家参与度、空间依赖性、模态关联以及功能贡献。留出集上的 WorldCover 探测评估表明元数据带来了 0.64 个百分点的收益，而检索实验则解耦了同传感器语义与跨传感器对齐。这些结果表明，该模型在紧凑的参数预算下，实现了面向多种传感器的灵活表征学习和强大的任务迁移能力。