tacofoundation/DeepExtremeCubes-video

DeepExtremeCubes-video数据集是一个高效存储、分析就绪的Sentinel-2小型立方体（minicubes）的视频格式重新包装。该数据集包含了2016年至2022年间42,000个Sentinel-2小型立方体，经过xarrayvideo工具转码为H.265/HEVC视频，实现了大约12倍的感知无损压缩，同时保持了科学上的保真度。这种紧凑的表示形式消除了在空间时间地球观测数据上训练深度学习模型的主要瓶颈，适用于诸如复合热浪和干旱事件的影响制图、极端情况下植被压力的预测以及长反射率序列的自监督预训练等任务。

DeepExtremeCubes-video dataset is a storage-efficient, analysis-ready re-packaging of Sentinel-2 minicubes in video format. It includes 42,000 Sentinel-2 minicubes from 2016 to 2022, transcoded with xarrayvideo into H.265/HEVC videos, achieving approximately 12 times lossless-perceptual compression while preserving scientific fidelity. This compact representation removes the primary bottleneck for training deep-learning models on spatio-temporal Earth-observation data and is suitable for tasks such as impact mapping of compound heat-wave and drought events, forecasting vegetation stress during extremes, and self-supervised pre-training on long reflectance sequences.