遇见数据集

Code, analysis log and derived data for "Temporal leakage in annual satellite embeddings inflates landslide susceptibility assessments"

收藏
Zenodo2026-11-04 更新2026-11-05 收录
官方服务:

资源简介:

This record accompanies the manuscript "Temporal leakage in annual satellite embeddings inflates landslide susceptibility assessments", submitted to Journal of Remote Sensing (earlier version: EarthArXiv preprint https://doi.org/10.31223/X5RB93). Annual satellite embeddings such as AlphaEarth Foundations and TESSERA summarize a calendar year, so the layer of a landslide's event year can already contain post-event observations. The study audits 18 dated landslide events with paired year swaps under fixed spatial validation, adds controlled Sentinel-2 experiments with a fixed number of observation dates and a preceding-year seasonal reference, decomposes the effect by training and prediction year, compares AlphaEarth with TESSERA and annual Sentinel-2 composites, evaluates complete maps, and tests how much pre-event embeddings add beyond conventional factors and explicit landslide history. Version 2.0 adds the analyses of the journal revision (M3_zenodo_v2_20261101.zip): code, locally prespecified protocols and run manifests, per-sample predictions, optical date selections and scene metadata, radiometric checks, date-qualified historical records and control samples, within-fold AUC statistics, the patch-missingness scan and paired CNN refits, land-cover blocks, the figures of the submitted manuscript with the script that redraws them, and the public copy of the analysis log (in Chinese). An offline check (scripts/m3_v12/verify_archive.py) verifies every checksum, recomputes the revision statistics from the saved predictions and redraws the figures byte-identically. The version 1.0 archive of the original analyses (M3_zenodo_package_20261019.zip) is included unchanged. Licences: code under the MIT License; files derived from the Unified Global Landslide Catalogue (paths containing "COOLR") under CC BY-NC 4.0, following the licence of that catalogue; the bundled Liberation Sans fonts under the SIL Open Font License 1.1; everything else under CC BY 4.0. See README.md and LICENSE.md inside the archive for contents, reproduction steps and third-party data sources. Download note: Download all 83 parts (.001 through .083), split_manifest.json and reassemble.py. Run python reassemble.py to reconstruct and verify M3_zenodo_v2_20261101.zip. See SPLIT_ARCHIVE_README.txt. The reconstructed archive is byte-identical to the original.

提供机构:
Zenodo
创建时间:
2026-11-04
二维码
社区交流群
二维码
科研交流群
商业服务