An urban canopy’s humidity significantly affects thermal comfort, public health, and building energy efficiency; however, it remains insufficiently understood. This study employed 3-year (2020–2022) fixed measurements from Guangzhou to investigate the temporal patterns of relative humidity (RH), vapor pressure (Ea), and vapor pressure deficit (VPD) across eight local climatic zones (LCZs). Clear and distinct patterns in the humidity characteristics among the LCZs were revealed on multiple timescales. The RH and VPD of each zone were higher in summer than in winter, with peak RH observed in LCZ A (83.45%) and peak VPD in LCZ 3 (13.6 hPa). Furthermore, a significant daytime urban dry island (UDI) effect in the summer and a nighttime urban moisture island (UMI) effect in the winter were observed in terms of the Ea difference between urban and rural areas. The strongest UMI occurred during winter nights in LCZ 8, with a peak intensity of 0.8 hPa, while the UDI was more frequent during summer days in LCZ 1, with a maximum value of −1.2 hPa; meanwhile, compact areas had a slightly higher frequency of UDI than open areas. Finally, the effects of the urban heat island (UHI) and wind speed (V) on UMI were analyzed. During the daytime, a weak correlation was observed between the UHI and UMI. Wind enhanced the intensity of the nighttime UMI. This research offers further insights into the canopy humidity characteristics in low-latitude subtropical cities, thereby contributing to the establishment of a universal model to quantify the differences in moisture between urban and rural areas.

Daylight glare probability (DGP) is the control variable for the operation and control of dayighting and solar energy in buildings, the use of AI to derive this metric warrants further research. Conventional DGP assessment relies on viewpoint-specific high-dynamic-range (HDR) imaging, limiting its application for continuous room-scale monitoring. This study proposes a non-intrusive coordinate-conditioned deep learning framework for spatially resolved DGP prediction. Unlike conventional image-to-value prediction, the proposed method formulates glare estimation as a spatially conditional learning problem by jointly incorporating HDR scene information and occupant location coordinates. A photometry-informed HDR transfer learning strategy is developed to improve prediction efficiency under limited DGP-labelled data. The image encoder is first adapted to the HDR domain using public indoor HDR panorama datasets through luminance-based proxy learning and then fine-tuned using measured ceiling HDR observations and eye-level HDR-derived DGP values. Four complementary HDR representations, including tone-mapped appearance, log-luminance, high-luminance regions, and multi-region luminance features, are integrated to capture glare-relevant spatial and photometric characteristics. The proposed framework was validated using 400 independent daylight scenes and 4800 spatially indexed DGP samples collected in a full-scale office-like environment. The model achieved an MAE of 0.0322, RMSE of 0.0368, R² of 0.9083, and Pearson correlation coefficient of 0.9536 for continuous DGP prediction. It also achieved 89.86% accuracy for four-class glare severity classification and F1 scores above 96% for threshold-based glare-event detection. The proposed approach enables room-scale DGP reconstruction from a single fixed HDR sensor, providing a practical solution for occupant-centric visual comfort monitoring and intelligent daylight-responsive building control.

Visual comfort assessment uses different methods and physical metrics for daylight and artificial lighting. Assessment under mixed lighting requires a common framework that relates the combined luminous scene to occupants’ overall comfort. This study developed a mixed-light visual comfort probability model (mVCP) that integrates four questionnaire-defined perceptual domains. The model estimates the probability of each of seven ordered comfort levels and the corresponding expected overall comfort rating. A repeated-measures experiment comprised 720 trials from 30 participants, including 240 trials reserved for testing. Brightness neutrality, colour-appearance suitability, perceived spatial uniformity and glare disturbance were predicted separately, then combined through cross-fitted non-negative aggregation and a proportional-odds model. Under joint participant–condition blocking, mVCP achieved a mean absolute error (MAE) of 1.087 points on the seven-point overall comfort scale and a normalised ranked probability score (RPS7) of 0.113. Relative to direct prediction using Extra Trees, the multidomain stack reduced MAE by 0.327 and RPS7 by 0.029. Comparisons with matched auxiliary supervision and a brightness–glare model supported domain-specific predictor assignment and the inclusion of all four domains, respectively. On the separate mixed-lighting test set, MAE was 1.012 and RPS7 was 0.116. The results support the predictive value of the specified perceptual representation within the sampled setting. Category probabilities, expected comfort and domain profiles provide complementary information for comparing lighting scenes and a basis for future occupant-centred control that evaluates comfort alongside energy demand.

The obstructive effect of urban context on the daylighting performance of a target building cannot be overlooked. Daylighting simulation software addresses this factor through surrounding context modelling. By contrast, AI-driven predictive tools require custom algorithms to account for such effects. This study developed a probabilistic framework based on prior conditions for predicting the attenuation of daylighting metrics relative to an unobstructed baseline. A Directional Obstruction Signature (DOS) was computed from aperture-based directional visibility and first-hit distance responses and summarized as an externally assigned Type16 prior defined by four severity strata and four directional patterns. A type-conditioned conditional variational autoencoder was trained using unobstructed baseline descriptors and the supplied Type16 prior to predict the median attenuation statistic together with stochastic prediction intervals. On the held-out test set, the model achieved an empirical coverage of 0.821, a mean interval width of 0.456, and MAE of 0.151. An external plausibility check showed that when each case was queried under externally supplied Type16 priors, the observed attenuation values were bracketed by limited and severity-consistent subsets of prior-conditioned intervals. These results position the approach as a scenario-based attenuation risk interface for early-stage design with externally assigned or exploratory obstruction priors.