Damage classifier v7 — HCI hard negatives

Generated 2026-07-01T01:23:39.489564Z

Motivation: Of 2,668 v6f-flagged buildings, 82.5 % are MS==0. Audit shows MS==0 alone has ~23 % label noise (hidden hand-pos), but with num_obs>=2 AND uncertainty==0, noise drops to ~8 %. This filtered pool (HCI, n=6,246) contains 553 v6f-flagged polygons — using it as hard negatives should cut FPs directly.

Comparison table (sorted by CV LR PR-AUC)

run desc n_train n_pos n_neg cv_auc_lr cv_auc_gb cv_ap_lr cv_ap_gb
v7c_hci_w5 HCI w=5, LC w=1, hand-neg w=5 15028 86 14942 0.8357 0.7633 0.8766 0.8100
v7b_hci_w3 HCI w=3, LC w=1, hand-neg w=5 15028 86 14942 0.8361 0.8071 0.8764 0.8255
v7a_hci_w2 HCI w=2, LC w=1, hand-neg w=5 15028 86 14942 0.8361 0.7869 0.8758 0.8275
v7d_hci_only HCI-only (LC dropped), hand-neg w=5 6397 86 6311 0.8332 0.8208 0.8758 0.8223
v6f_repro hand-neg w=5 + all MS==0 w=1 (v6f) 15028 86 14942 0.8363 0.7765 0.8751 0.7943

Best: v7c_hci_w5 · LR threshold: 0.543 · GBT threshold: 0.000

Flagged: LR 4008 · GBT 7418

Plots

ROC (LR)

PR (LR)

Score distribution (v7 best)