Identifying landslides from remote imagery is critical for rapid responses after landslide hazards and for assessing their environmental impacts. Existing datasets for landslide detection models are mostly based on multi-sourced, high-resolution (e.g., 1-5 m) satellite imagery from commercial companies (e.g., Planet Labs) and ultra-high-resolution (e.g., \textless 1m) photos from unmanned aerial vehicle (UAV) surveys. However, obtaining those data is often economically expensive and labor-intensive, limiting their applicability. Here we present ‘LandsatQuake,’ a benchmark dataset composed of 31 landslide inventories from 21 earthquake-prone regions across the world covering a total area of (5.56 \times 10^7 ) km(^2) and spanning the last 40 years. This dataset emphasizes practicality by using satellite images acquired by Landsat, the only satellite system that has recorded Earth’s land surface for >40 years. The public availability, high coverage of the world, and longevity make the Landsat data ideal for developing historical and recent landslide inventories caused by known triggers (e.g., earthquakes or rainstorms). Additionally, we demonstrate the challenges of applying existing computer vision algorithms to practical landslide detection problems by evaluating several baselines.