zinnia82/flood-binary-hnm-benchmark

Flood Binary HNM Benchmark

Street-level flood/non-flood binary classification imagery, used in "Improving CRIS-HAZARD: Automated First-Pass Flood Image Screening via Phase-1 Hard Negative Mining" (Singh & Dixon, pending submission to Computers & Geosciences).

4,099 deduplicated (SHA-256 exact-match only, no perceptual dedup) street-level images, stratified 80/20 by fine-grained category, seed=42. Flood prevalence is 39.3% in both splits. No held-out test split — the paper reports all metrics on the val split.

Category	Train	Val	Total
street_major	640	156	796
street_moderate	245	56	301
street_minor	406	110	516
flood total	1291	322	1613
river	323	76	399
swimming_pool	70	28	98
park_walkway	326	82	408
street_clear	475	105	580
animal	233	67	300
building	232	55	287
vehicle	189	50	239
plant	141	34	175
non_flood total	1989	497	2486
grand total	3280	819	4099

Fields

• image: the image.
• label: 0 = flood, 1 = non_flood.
• category: fine-grained category (street_major, street_moderate, street_minor, river, swimming_pool, park_walkway, street_clear, animal, building, vehicle, plant).

Usage

from datasets import load_dataset
ds = load_dataset("zinnia82/flood-binary-hnm-benchmark")

River confounder imagery is sourced separately (RIWA dataset via Kaggle, franzwagner/river-water-segmentation-dataset) and capped at 400 images per the paper's methodology.

Changelog

• 2026-07-12: Replaced with the finalized, capped dataset matching the paper's reported Table 1 counts (previously an uncapped draft with ~1,600 river images across train+val). val is now a properly registered, loadable split. Dropped the source provenance column present in the previous version.