PADR-Net reproducibility archive: Physics-Informed Reservoir Learning for Shallow-Water Flood Modelling

Le résumé fourni par la source

This archive provides the reproducibility package for “Physics-Informed Reservoir Learning for Shallow-Water Flood Modelling.” It supports the full computational workflow of PADR-Net — the Physics-Aware Deep Reservoir Network — a physics-informed reservoir-learning framework for flood-depth reconstruction, hydrodynamic consistency assessment, and flood-severity prediction under sparse and partially observed flood data. This version corresponds to the MATG rebuild release of the PADR-Net archive. Compared with the earlier v1.0.0-matg-resubmission package, this release expands the mathematical and computational formulation by including an observation-aware data model, explicit source-term closures, a shared spatial reservoir formulation, a formal PADR-Net training and transfer-validation algorithm, residual-stability diagnostics, baseline comparisons, and additional validation tables. The study evaluates PADR-Net on controlled synthetic flood regimes and on a harmonized archive of 243 African flood events spanning 2000–2024 across three macro-regions: West Africa (Niger–Benue), East Africa (Nile headwaters), and Southern Africa (Limpopo–Zambezi). The archive is designed to make the manuscript’s numerical experiments transparent, repeatable, and auditable by providing the scripts, processed metadata, split definitions, result tables, environment files, and publication figures needed to reproduce the reported analyses. Contents of this release This reproducibility package contains: Python scripts for data preparation, feature construction, model training, baseline evaluation, ablation analysis, transfer validation, uncertainty analysis, residual-stability diagnostics, and figure generation. Pre-computed result tables for nested predictor comparisons, M0–M8 ablation experiments, physics-weight sensitivity, source-term closure sensitivity, residual-stability diagnostics, stencil-sensitivity experiments, leave-one-region-out (LORO) validation, leave-one-year-out (LOYO) validation, and bootstrap confidence intervals. Publication figures in multiple formats, including PNG, SVG, and EPS (see previous version). Environment files, including environment.yml and requirements.txt, to reproduce the computational setup. Metadata for the 243-event African flood inventory, including regional assignment, event timing, severity class, split membership, and validation role. Source metadata and download instructions for public external datasets, including ERA5 reanalysis from the Copernicus Climate Data Store and satellite flood observations from the Global Flood Database. Scripts to regenerate the main manuscript tables and publication figures from the pre-computed result files without requiring users to re-download all raw external data. Mathematical and methodological additions in this version This MATG rebuild release supports the expanded manuscript formulation, including: An observation-aware data model that distinguishes satellite flood extent, benchmark/reference depth, hydrodynamic residual collocation points, and event-level impact labels. A source-term closure operator for precipitation, infiltration, drainage loss, lateral inflow, and Manning roughness fields. A shared spatial Echo State reservoir formulation in which reservoir weights are fixed and shared across grid cells, while local forcing, terrain, and exposure descriptors vary by event, time, and grid cell. A physics-informed objective combining observation losses, shallow-water-equation residual loss, and output-weight regularization. A residual-distance bound showing that, under a local residual-stability condition, the distance to the physically admissible shallow-water residual manifold decreases at the rate (O(\lambda_{\mathrm{phys}}^{-1/2})). A prediction-head decoupling result showing that the physics weight can be tuned using hydrodynamic diagnostics without directly changing the fixed-design severity-ranking head. A formal PADR-Net training and transfer-validation algorithm. Residual-stability and stencil-sensitivity diagnostics designed to support interpretation of the physics regularization. Reproducing the main outputs The main publication outputs can be reproduced from the pre-computed result tables using: conda env create -f environment/environment.yml conda activate padrnet python code/scripts/06_make_figures.py A full rerun of the computational workflow can be performed when the external datasets are available locally using the scripts in code/scripts/. The archive is organized so that users can either reproduce the publication tables and figures directly from released result files or rerun the complete PADR-Net workflow from processed inputs. Key reported results The full PADR-Net configuration uses rainfall, hydroclimatic memory, exposure, and hydrodynamic descriptors (R+M+E+H). In the archived experiment set, the selected physics-informed configuration uses (\lambda^{*}=0.10). On the African flood archive, the full model achieves: Spearman rank correlation: (\rho_s = 0.671) PR-AUC: 0.703 Depth reconstruction skill: NSE(_{\mathrm{depth}}) = 0.643 MAE: 3.569 The nested predictor analysis shows that antecedent memory substantially improves severity ranking, exposure descriptors improve high-impact discrimination, and hydrodynamic descriptors improve PR-AUC, MAE, and depth reconstruction. The sensitivity experiments further show that the physics weight controls the hydrodynamic reconstruction and residual-consistency trade-off, while the fixed-design severity head remains invariant during the physics-weight sweep. Links Code repository: https://github.com/earthai-tech/padrnetTagged release: https://github.com/earthai-tech/padrnet/releases/tag/v1.1.0-matg-rebuildDocumentation: https://base-attentive.readthedocs.io/en/latest/ERA5 data access: https://cds.climate.copernicus.euGlobal Flood Database: https://global-flood-database.cloudtostreet.ai This archive is intended to support reproducibility, code transparency, mathematical traceability, and validation auditability for the PADR-Net flood modeling experiments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.