AuraTracer智迹闻
中文

EVENT DOSSIER

MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation

2026-09-07 12:00 Models across 2 days 🔥 47.2 heat score
2sources
2days unfolding
47.2heat score
1mentions
SummaryAI generated

In September 2026, researchers proposed the MEOX model, a compact multi-modal hybrid expert (MoE) architecture designed for Earth observation. The model contains approximately 3.115 million parameters and uses sensor-specific adapters, explicit validity signals, and shared sparse expert blocks to process multi-modal dependent information. The model was pre-trained on 1.228 million MMEarth64 samples and evaluated for frozen迁移 in six GEO-Bench tasks. Experimental results showed that MEOX achieved an average F1 score of 64.42% in the 64-pixel-scale Cash Fruit segmentation task and an average accuracy of 90.56% in the 224-pixel-scale EuroSAT task. Additionally, the model improved performance by 0.64 percentage points for the WorldCover detection task using retained metadata tokens, and achieved a micro-average accuracy of 72.95% during BigEarthNet fine-tuning.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
MEOX

Integrated timelineUNIFIED TIMELINE

  1. 2026-09-04

    MEOX: Compact Multimodal Mixture-of-Exp…

    MEOX 提出一种紧凑的多模态混合专家模型,其编码器参数为 293.9 万,总参数量为 311.5 万。该模型采用传感器特定适配器、显式有效性信号及共享稀疏专家块进行预处理,随后通过十四层编码器块与四个元数据令牌完成融合。模型在 122.…

  2. 2026-09-07

    MEOX: Compact Multimodal Mixture-of-Exp…

    MEOX 模型在六个 GEO-Bench 任务上进行了冻结迁移评估,其中在 64 像素尺度下的现金果分割任务中平均交并比达到 64.42%,在 224 像素尺度的 EuroSAT 任务中平均准确率高达 90.56%。该模型采用多模态掩码自…

SignalsSIGNALS

Keyword heat
  • MEOX2

All reports (2)SOURCES

A arXiv cs.CV en 2026-09-05 00:59

MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation

MEOX 提出一种紧凑的多模态混合专家模型,其编码器参数为 293.9 万,总参数量为 311.5 万。该模型采用传感器特定适配器、显式有效性信号及共享稀疏专家块进行预处理,随后通过十四层编码器块与四个元数据令牌完成融合。模型在 122.8 万个 MMEarth64 样本上进行预训练,并在六个 GEO-Bench 任务中评估冻结迁移效果。在 64 像素尺度下,其现金果分割平均交并比达 64.42%,超过 CSMoE 结果;在 224 像素尺度下,EuroSAT 平均准确率为 90.56%。此外,BigEarthNet 微调微平均精度为 72.95%,元数据带来 0.64 个百分点的提升。

A arXiv cs.CV en 2026-09-07 12:00

MEOX: Compact Multimodal Mixture-of-Experts for Earth Observation

MEOX 模型在六个 GEO-Bench 任务上进行了冻结迁移评估,其中在 64 像素尺度下的现金果分割任务中平均交并比达到 64.42%,在 224 像素尺度的 EuroSAT 任务中平均准确率高达 90.56%。该模型采用多模态掩码自编码器架构,包含 293.9 万参数编码器和总计 311.5 万参数,通过传感器特定适配器、显式有效性信号及共享稀疏专家块处理模态依赖信息。模型在 122.8 万个 MMEarth64 样本上进行预训练,利用模态平衡掩码重建和结构化传感器 Dropout 策略,并在下游空间网格上支持旋转注意力机制。实验表明,保留的元数据 token 为 WorldCover 探测任务带来 0.64 个百分点的提升,且路由诊断有效区分了专家参与、空间依赖及模态关联等功能贡献。