Evaluating machine learning maturity and operational efficacy in database indexing and satellite-based flood mapping: a three-part evidence assessment

Abstract

This study evaluates machine learning maturity and operational efficacy across database indexing and satellite-based flood mapping, emphasizing methodological rigor and real-world reliability rather than reported performance alone. A PRISMA 2020 systematic review of 65 studies on one-dimensional learned database indexes finds reported 2–4× median point-lookup speedups over B-trees but low certainty of evidence due to evaluation bias and incomplete variance reporting. A PRISMA-ScR review of 73 Sentinel-1/2-based flood-mapping studies identifies a shift toward transformers, diffusion, and self-supervised architectures, with multi-sensor fusion showing directional improvements in most studies, while cloud dependence, unreleased datasets, and limited code availability constrain deployment. An applied benchmark of six SAR flood-detection methods against the Copernicus Global Flood Monitoring service across four monsoon events in Roumari Upazila, Bangladesh, demonstrates higher performance by the best local method, while acknowledging that the comparison does not isolate specific methodological factors. A crop-calendar overlay further contextualizes flood exposure during T. Aman transplanting and Aus harvesting, without inferring agricultural-loss rankings. Overall, the study highlights the gap between reported machine learning performance and reproducible, reliable operational deployment.