RingMOSS: A Comprehensive Multi-Modal Pre- Training Datase

To establish a versatile RSFM adaptable to diverse tasks, RingMoE requires a comprehensive and diverse pre-training dataset that accounts for significant variations in imaging modalities, spatial resolutions, temporal dynamics, geographic regions, and scene complexities. To meet this challenge, we curate RingMOSS, a large-scale multi-modal RS dataset comprising 400 million images from nine satellite platforms, covering a broad spectrum of Earth observation scenarios. 

为构建适配多样任务的通用遥感基础模型（Remote Sensing Foundation Model, RSFM），RingMoE模型需要一套全面且多样化的预训练数据集，以覆盖成像模态、空间分辨率、时间动态特性、地理区域以及场景复杂度等多维度的显著差异。为应对这一挑战，我们精心构建了RingMOSS——一款大规模多模态遥感数据集，其包含来自9个卫星平台的4亿幅图像，覆盖了广泛的地球观测场景。

搜集汇总

背景与挑战

背景概述

RingMOSS是一个大规模多模态遥感预训练数据集，包含来自九个卫星平台的4亿张图像，覆盖了广泛的地球观测场景。该数据集适用于遥感图像处理和多模态学习任务。