Physics-Informed Landslide Dataset (PILD): Metadata, Protocols, and Release Assets for Cross-Domain Landslide Mapping

The Physics-Informed Landslide Dataset (PILD) release v1.0 is a public metadata and protocol package for cross-domain landslide mapping research. The release provides event-level indices, benchmark split tables, protocol summaries, source registries, and reproducibility-oriented documentation for a heterogeneous benchmark assembled from multiple landslide source families. The package is designed to support benchmark definition, protocol audit, and source-traceable reconstruction without redistributing all upstream raw imagery or environmental products, which remain governed by the licenses and distribution policies of their original providers. The release includes the canonical event master table, the strict_t2 event index, split files for in-domain, leave-one-dataset-out, and cross-trigger evaluation, a DLR-centered strict_t3 event summary, a data card, a data dictionary, a source registry, a source license matrix, upstream download pointers, a global benchmark coverage figure, and a release-safe rendering script for the coverage map. Companion code repository: https://github.com/Liu-Zhihang/geophysadapter Keywords - landslide - remote sensing - cross-domain learning - benchmark dataset - geophysical context - geospatial AI