This Global Dated Landslide Database (GDLDB) is part of the project WeMonitor (Weakly Supervised Deep Learning Models for Detecting and Monitoring Spatio-Temporal Anomalies in Optical and Radar Satellite Time Series), funded by the Helmholtz Imaging Platform. The aim is to develop a deep learning model that uses satellite image time series from Sentinel1/2 to automatically monitor changes caused, for example, by landslides, deforestation, large fires, dam failures, or the emergence of waste dumps. To train such a model, a reference dataset is required that shows the area and date of the changes as precise as possible. To allow for a generic and transferable model, the reference data also needs to cover the diversity of the process to be detected. Thus, the aim of the GDLDB is to comprise landslides of different sizes, shapes, and types, occurring at different seasons and in different regions with varying natural conditions and different triggering mechanisms such as rainfall and earthquake-induced landslides.

To build the GDLDB, available local and regional landslide inventories from around the world are combined into one coherent database by verifying their location and date of occurrence with high-resolution remote sensing data. The selection criteria for the source inventories are the definition of the landslide location as polygons, at least a rough indication of the landslide origin date, and that the landslides occurred during the Sentinel-2 data availability from 2016 onwards. A total of 16 individual inventories are included (Table 1), one each from the USA, Dominica, Italy, Zimbabwe, southern India, Nepal, China, Papua New Guinea, and New Zealand, and two each from Kyrgyzstan, Japan, and the Philippines. In addition, a global inventory was added, including a small number of landslides from the USA, Peru, Chile, Europe, Pakistan, Nepal, India, and Taiwan, and a larger number of landslides from Indonesia.

From each inventory, approximately 100 landslides were randomly selected to ensure an unbiased selection of landslides in terms of shape, size, and location. The original source inventories are produced using a variety of methods, including manual mapping in airborne data with ground verification and automatic identification in satellite remote sensing data. As a result, the mapping quality of the inventories varies greatly. In cases where landslides could not be verified by us using available optical remote sensing data (e.g. Sentinel-2, Planet Scope, and data available in Google Earth) new polygons are selected until the number of approximately 100 landslides is reached. In some inventories, the number of 100 landslides could not be guaranteed, due to a lack of suitable landslides (e.g., small size, incorrect classification) or the total number of landslides in the selected inventory was less than 100. For inventories with a lot of small landslides, that were difficult or impossible to observe, a size threshold of 1000m2 was introduced.