Fermium Hazard Mapper: input data and outputs for flood and landslide susceptibility in Bangladesh (version 4)

Abstract

Version 4 (October 2026). Adds the robustness checks of the accompanying paper on the comparison of the flood model with the record of earlier flooding: the temporal test without ground flooded in every earlier event and outside ground the JRC record has seen under water (persistence_check.json), with 20 and 40 km blocks (block_size_check.json), with the model trained on two further targets (target_variants.json), and on flood masks re-mapped from Sentinel-1A alone (raw_s1a/ and output/sensitivity_s1a/), together with the terrain hazard surface scored only on held-out cells (hazard_heldout.json). Every comparison over block assignments now reports the corrected resampled t-test of Nadeau and Bengio (2003) as p_value, with the plain t-test as p_value_uncorrected. Scores, maps and every other output are unchanged from version 3, which first kept every asset and grid cell to Bangladesh. These outputs were produced by the code at commit 1faf040 of the repository, later than the software release Fermium Hazard Mapper v3.0. Input data and pipeline outputs for Fermium Hazard Mapper, which estimates flood risk to mapped infrastructure and people in three regions of Bangladesh (Rangpur and Rajshahi, Sylhet, and the south-west coast) and landslide susceptibility in the Chittagong Hill Tracts. Each region is one archive with the inputs the pipeline ingests (raw/) and the outputs it produces (output/), and a manifest giving the size, SHA-256 checksum and licence of every file. The flood models are trained on flood extents mapped from Sentinel-1 radar for thirteen events between 2017 and 2024 and validated on 10 km blocks held out from training. The outputs include the per-event flood masks, the asset scores and calibrated probabilities, the terrain and Kriged hazard surfaces, the composite risk grid, union, upazila and district summaries, and the metrics of every validation described in the accompanying paper: held-out blocks, observed floods, the model comparison over twenty block assignments, the label-source comparison, the temporal hold-out on the 2024 floods, the past-flooding feature test and the weight sensitivity analysis. Code: https://github.com/rakibhhridoy/rimes-mapathon (MIT). Every figure in the paper is generated from these outputs. Licences. CC BY 4.0 for the outputs. Inputs keep their own licences, listed per file in each manifest (OpenStreetMap under ODbL 1.0, WorldPop and geoBoundaries under CC BY 4.0, SRTM public domain, JRC Global Surface Water free with attribution, NASA COOLR open, Sentinel-1 derived flood masks under the Copernicus open licence). Changes from version 1 (March 2026). Version 1 contained GADM administrative boundaries, which GADM's licence does not allow to be redistributed. They are removed and replaced by geoBoundaries at the union, upazila and district levels. Version 1's outputs came from a pipeline with a coordinate-system error that left every training label at zero, and should not be used. All outputs here are from the corrected pipeline, run on 23 and 24 September 2026.