EarthNet2021 (EarthNet2021: Earth Surface Forecasting)

卫星图像是地球表面的快照。我们建议对它们进行预测。我们将地球表面预测定义为根据未来天气预测卫星图像的任务。 EarthNet2021 是一个适合在任务上训练深度神经网络的大型数据集。它包含分辨率为 $20$~m 的 Sentinel~2 卫星图像，匹配的地形和中尺度（$1.28$~km）气象变量打包成 $32000$ 的样本。此外，我们将 EarthNet2021 视为允许模型互比的挑战。结果预测将比数值模型中的空间分辨率大大提高 ($>\times50$)。这可以预测极端天气的局部影响，从而支持下游应用，例如作物产量预测、森林健康评估或生物多样性监测。在 www.earthnet.tech 上查找数据、代码以及参与方式。Satellite imagery consists of snapshots of the Earth's surface. We propose the task of forecasting such satellite imagery, which we define as predicting future satellite imagery based on forthcoming weather forecasts. EarthNet2021 is a large-scale dataset well-suited for training deep neural networks on this task. It contains 32,000 samples of Sentinel-2 satellite imagery with a 20-meter resolution, paired with matching topographic and mesoscale (1.28 km) meteorological variables. Additionally, EarthNet2021 is structured as a challenge benchmark that allows for direct comparison between different models. The spatial resolution of the resulting forecasts is substantially improved (>50-fold) compared to that of numerical models. This capability enables the prediction of localized impacts of extreme weather events, supporting downstream applications including crop yield forecasting, forest health assessment, and biodiversity monitoring. Find the dataset, code, and participation guidelines at www.earthnet.tech.

搜集汇总

背景与挑战

背景概述

EarthNet2021是一个用于地球表面预测的大型数据集，包含32000个高分辨率(20m)的Sentinel 2卫星图像样本及匹配的地形和气象数据，可用于预测极端天气的局部影响并支持农业、林业等下游应用。该数据集由多个德国研究机构联合发布，旨在促进深度学习模型在该任务上的比较研究。