Buildings’ sustainability is one of the crucial parts for achieving urban sustainability. Applied to buildings, life-cycle assessment encompasses the analysis and assessment of the environmental effects of building materials, components and assemblies throughout the entire life of the building construction, use and demolition. Estimate of carbon emissions is essential and crucial for an accurate and reasonable life-cycle assessment. Addressing the need for more research into integrating analysis of real-time and automatic recording of key indicators for a more accurate calculation and comparison, this paper aims to design a real-time recording model of these crucial indicators concerning the calculation and estimation of energy use and carbon emissions of buildings based on a Radio Frequency Identification (RFID)-based system. The architecture of the RFID-based carbon emission recording/tracking system, which contains four functional layers including data record layer, data collection/update layer, data aggregation layer and data sharing/backup layer, is presented. Each of these layers is formed by RFID or network devices and sub-systems that operate at a specific level. In the end, a proof-of-concept system is developed to illustrate the implementation of the proposed architecture and demonstrate the feasibility of the design. This study would provide the technical solution for real-time recording system of building carbon emissions and thus is of great significance and importance to improve urban sustainability.

Glass curtain walls dominate the envelope of contemporary high-rise commercial buildings, yet their long-term optical, thermal, and aesthetic performance is highly sensitive to surface cleanliness. Manual cleanliness inspection of high-rise facades is hazardous, labour-intensive, and inconsistent, motivating the development of unmanned aerial vehicle (UAV)-based assessment frameworks. However, existing UAV-based studies primarily focus on structural defect detection, neglect the operational conditions under which visual cleanliness assessment remains physically reliable, and seldom relate learned feature representations to interpretable optical degradation mechanisms. This paper presents a hybrid deep-learning and physics-informed framework for UAV-based cleanliness assessment of glass curtain walls, deployed on a QX-90A tethered cleaning UAV (UEV, Nanjing, China) and validated on GCW-429, a newly curated in situ dataset of 429 expert-labelled facade tiles acquired from a commercial office building in Nanjing, China. A ResNet-50 binary classifier distinguishing acceptable from cleaning-needed states achieves a test-set ROC-AUC of 0.620, an accuracy of 0.645, and an F1 of 0.571 for the cleaning-needed class. Critically, stratified evaluation reveals a strong dependence of inference reliability on capture conditions: AUC reaches 0.681 under Cloudy illumination but drops to 0.394 under Sunny direct illumination, a regime in which specular reflection from the low-emissivity coating dominates the sensor signal. Spearman correlation between CNN penultimate-layer principal components and five physics-informed optical indicators (Otsu pollution coverage, K-means dispersion, Sobel edge density, Gabor texture energy, SSIM clarity) yields nine pairs with rho ≥ 0.30 (p < 0.001), demonstrating that the network learns physically meaningful representations rather than spurious cues. Grad-CAM analyses corroborate these findings and explicitly localise the failure mode of Sunny-illumination inference to specular high-light regions. The framework closes an inspect-assess-clean loop on a single UAV platform and yields concrete operational protocol recommendations for facade-cleaning campaigns.