[Submitted on 28 Oct 2026]
Abstract:Repeated airborne LiDAR surveys are costly, while stored 3D geometry can become outdated between observations. Existing change-detection methods compare observed epochs. However, they do not quantify the validity of geometry that has not been revisited. We distinguish observation, existence and geometric validity using a probabilistic filter over five geometric transitions: persistence, modification, growth, removal and replacement. The framework learns class-specific continuous-time transition priors from interval-censored observations and combines shape residuals with LiDAR free-space evidence. Functional anchors enable prior sharing across sites with different survey frequencies. In controlled experiments with misspecified dynamics and 60% visibility, our method reduces combined error from 36.9% to 16.5% and expected calibration error (ECE) from 0.19-0.32 to 0.017. Cross-site sharing reduces ECE at two-survey sites from 0.156 to 0.029. On Urb3DCD-v2, the method achieves 9.8% combined error and 0.016 ECE, outperforming baselines at 12.6% and at least 0.070, respectively. Re-survey prioritisation improves only modestly over random selection with two observations. Code and an evaluation protocol are released at this https URL.
Submission history
From: Anwaar Ulhaq Dr [view email]
[v1] Wed, 28 Oct 2026 03:45:09 UTC (460 KB)
