tacofoundation/DynamicEarthNet-video

DynamicEarthNet-video是一个存储效率高的DynamicEarthNet数据集的重新包装版本。该数据集包含了全球75个1024×1024像素区域（大约3米地面采样距离）的每日图像，时间跨度从2018年1月1日至2019年12月31日。每天提供的图像是四波段PlanetFusion表面反射率图像（红色B04、绿色B03、蓝色B02、窄带近红外B8A）。每月的像素级标签标注了七个土地覆盖类别：不透水、农业、森林、湿地、裸土、水和雪/冰。原始的GeoTIFF堆栈（大约525GB）通过xarrayvideo转换为12位H.265/HEVC格式，实现了显著的尺寸节省同时保持科学保真度。这种紧凑的视频格式消除了I/O瓶颈，使得可以直接从磁盘进行序列模型的端到端训练，快速进行4波段每日时间序列的实验，以及高效地共享变化检测和预测的基准。

DynamicEarthNet-video is a storage-efficient re-packaging of the original DynamicEarthNet collection. The dataset covers seventy-five 1024 × 1024 px regions (≈ 3 m GSD) across the globe, sampled daily from 1 January 2018 to 31 December 2019. Each day is delivered as four-band PlanetFusion surface-reflectance images (B04 Red, B03 Green, B02 Blue, B8A Narrow-NIR). Monthly pixel-wise labels annotate seven land-cover classes: impervious, agriculture, forest, wetlands, bare soil, water and snow/ice. The original GeoTIFF stacks (≈ 525 GB) are transcoded with xarrayvideo to 12-bit H.265/HEVC, achieving dramatic size savings while preserving scientific fidelity. The compact video format therefore removes I/O bottlenecks and enables end-to-end training of sequence models directly from disk, rapid experimentation on 4-band daily time-series, and efficient sharing of benchmarks for change detection and forecasting.