🏠 Buildings v.07-06 — per-location benchmark

This page details the validation of the 🏠 Buildings v.07-06 segmentation model (Global domain, 0.3 m / z19) on a set of 7 areas of interest (AOI), compared against manually annotated ground truth. For reference, each AOI is also evaluated for the previous candidate v.02-06 and the current production 🏠 Buildings model.

All metrics are object-wise: a predicted footprint counts as a true positive (TP) when it matches a ground-truth footprint, otherwise it is a false positive (FP); unmatched ground-truth footprints are false negatives (FN). Evaluation run: 2026-06-09.

Note

In the overlay images, predicted footprints are compared visually against the ground-truth annotation for each AOI (combined view exported from the test pipeline).

United States

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.945	0.923	0.968	60 / 5 / 2
v.07-06	0.960	0.952	0.968	60 / 3 / 2
🏠 Buildings (current)	0.916	0.870	0.968	60 / 9 / 2

United States — prediction vs ground-truth overlay.

Canada

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.826	0.784	0.874	76 / 21 / 11
v.07-06	0.809	0.771	0.851	74 / 22 / 13
🏠 Buildings (current)	0.809	0.752	0.874	76 / 25 / 11

Canada — prediction vs ground-truth overlay.

South Africa

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.779	0.917	0.677	44 / 4 / 21
v.07-06	0.739	0.891	0.631	41 / 5 / 24
🏠 Buildings (current)	0.782	0.956	0.662	43 / 2 / 22

South Africa — prediction vs ground-truth overlay.

New Zealand

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.756	0.716	0.800	48 / 19 / 12
v.07-06	0.734	0.691	0.783	47 / 21 / 13
🏠 Buildings (current)	0.788	0.722	0.867	52 / 20 / 8

New Zealand — prediction vs ground-truth overlay.

Côte d’Ivoire

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.802	0.859	0.753	67 / 11 / 22
v.07-06	0.717	0.814	0.640	57 / 13 / 32
🏠 Buildings (current)	0.778	0.863	0.708	63 / 10 / 26

Côte d’Ivoire — prediction vs ground-truth overlay.

United Kingdom

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.650	0.565	0.765	39 / 30 / 12
v.07-06	0.703	0.650	0.765	39 / 21 / 12
🏠 Buildings (current)	0.646	0.539	0.804	41 / 35 / 10

United Kingdom — prediction vs ground-truth overlay.

Australia

Model	F1	Precision	Recall	TP / FP / FN
v.02-06	0.721	0.617	0.866	71 / 44 / 11
v.07-06	0.674	0.625	0.732	60 / 36 / 22
🏠 Buildings (current)	0.579	0.525	0.646	53 / 48 / 29

Australia — prediction vs ground-truth overlay.

Summary

Compared with the previous version (the current production 🏠 Buildings model), v.07-06 lifts the mean object-wise F1 from 0.757 to 0.762 (+0.005), with mean precision up (+0.024) and recall slightly down (−0.023). The gains are largest on the weaker AOIs — Australia (F1 +0.095), United Kingdom (+0.057) and the United States (+0.044) — while Canada holds parity. The trade-off is lower recall on dense / informal patterns: New Zealand (−0.054), Côte d’Ivoire (−0.061) and South Africa (−0.043). Overall the update raises precision and the floor on the hardest areas at a small cost to recall elsewhere. The tables above also list the earlier candidate v.02-06 for reference.