Vacant houses increasingly challenge regions undergoing demographic change, requiring effective and scalable survey methods. While UAV-based 3D modeling has been applied to vacant house surveys, existing approaches rely mainly on daytime imagery and manual visual interpretation. Nighttime UAV imagery could exploit indoor lighting as an occupancy indicator; however, its application is constrained by noise and low contrast. To address these challenges, we proposed a UAV-based framework that integrates nighttime imagery, low-light image enhancement (LLIE), and 3D reconstruction, and evaluated its feasibility through a joint analysis of data characteristics, model design, and enhancement outcomes. Experiments in Omuta and Kashiwa-no-ha reveal a distinct distributional shift in UAV imagery over the daytime–nighttime transition, from skewed-normal patterns during the daytime to gamma- and Poisson-dominated distributions as illumination decreases. Because pixel-wise paired day–night UAV image pairs are rarely available, experiments primarily relied on publicly available LLIE architectures with pre-trained weights, with only CycleGAN and the self-calibrated illumination (SCI) model trained on UAV data. These pre-trained models are largely learned from gamma-distributed benchmark datasets, resulting in a domain gap when applied to real UAV imagery with Poisson-dominated statistics. Under this constraint, SCI consistently produced high-contrast, low-noise enhancements across both study areas, achieved the highest frequency of top-three rankings across perceptual metrics among all models (e.g., in Kashiwa-no-ha: CLIP 88.7, 1st; LPIPS 0.43, 2nd; SSIM 0.33, 3rd). These results demonstrate the practical feasibility of integrating nighttime UAV surveys and LLIE for vacant house assessment, supporting future urban monitoring and digital twin applications. • LLIE framework for Nighttime UAV vacant house surveys. • Gamma-to-Poisson shift explains benchmark/UAV domain gap. • SCI model is robust for extreme low-light UAV imagery. • Self-supervised SCI enables adaptation without paired data.
Yu et al. (Sun,) studied this question.