Satellite images are snapshots of the Earth surface. We propose to forecastthem. We frame Earth surface forecasting as the task of predicting satelliteimagery conditioned on future weather. EarthNet2021 is a large dataset suitablefor training deep neural networks on the task. It contains Sentinel 2 satelliteimagery at 20m resolution, matching topography and mesoscale (1.28km)meteorological variables packaged into 32000 samples. Additionally we frameEarthNet2021 as a challenge allowing for model intercomparison. Resultingforecasts will greatly improve (>x50) over the spatial resolution found innumerical models. This allows localized impacts from extreme weather to bepredicted, thus supporting downstream applications such as crop yieldprediction, forest health assessments or biodiversity monitoring. Find data,code, and how to participate at www.earthnet.tech