Multi-spectral remote thermal imaging for surface emissivity and estimation of roof R-values using physics-based and data mining models

Date of Award


Degree Name

Ph.D. in Mechanical Engineering


Department of Mechanical and Aerospace Engineering


Advisor: K. P Hallinan


Remote thermal imaging of buildings is notable for providing interesting but generally qualitative images of buildings. A recent study showed that if accurate measurements of exterior surface temperatures could be obtained from single-point-in-time-imaging, then it would be possible to infer envelope R-values and thermal capacitances with reasonable accuracy. This research seeks to answer the question, How can we make possible reasonably accurate measurements of the external temperatures from at-scale remote imaging?" Without knowledge of the emissivity of the exterior surfaces, accurate thermal assessment is seemingly impossible. Here, we exploit the unique spectral characteristics of the most common exterior building surfaces using multi-spectral imaging. Four to five images of exterior surfaces in the 1-5-micron range, where the spectral emissivity of different building materials changes most, is posed. The pattern of the emission can be correlated to various envelope component surface spectral emissivities. A neural network pattern matching algorithm is used to "find" the surface type. Then, with known emissivity, the surface temperature can be inferred from the magnitude of the emission. Theoretical results indicate that temperature error in measuring the surface temperature in using this approach can be less than ±1°C. This error is sufficient for identifying envelope R-values based upon the research posed by Salahaldin and Hallinan [1]. Most exciting is the prospect of this technique for effectively measuring building R-values at scale via fly-over or drive by imaging. Conventional residential building energy auditing needed to identify opportunities for energy savings is expensive and time consuming. On-site energy audits require quantification of envelope R-values, air and duct leakage, and heating and cooling system efficiencies. There is a need to advance lower cost automated approaches, which could include aerial and drive-by thermal imaging at-scale in an effort to measure the building R-value. However, single-point in time thermal images are generally qualitative, subject to errors stemming from building dynamics, background radiation, wind speed variation, night sky thermal radiation, and error in extracting temperature estimates from thermal images from surfaces with generally unknown emissivity. This work proposes two alternative approaches for estimating roof R-values from thermal imaging, one a physics based approach and the other a data-mining based approach. Both approaches employ aerial visual imagery to estimate the roof emissivity based on the color and type of roofing material, from which the temperature of the envelope can be estimated. The physics-based approach employs a dynamic energy model of the envelope with unknown R-value and thermal capacitance. These are tuned in order to predict the measured surface temperature at the time of the imaging, given the transient weather conditions prior to the imaging. The data-mining approach integrates the inferred temperature measurement, historical utility data, and easily accessible or potentially easily accessible housing data. A data mining regression model, trained from this data using residences with known R-values, is used to predict the roof R-value in the unknown houses. The data mining approach was shown to be a far superior approach, demonstrating an ability to estimate attic/roof R-value with an r-squared value of greater than 0.88 using as few as nine training houses. The implication of this research is significant, offering the possibility of auditing residences remotely at-scale via aerial and drive-by thermal imaging coupled with utility analysis."


Buildings Thermography, Roofs Thermal properties Remote sensing, Infrared imaging, Exterior walls Thermal properties, Mechanical Engineering, Thermal Imaging, Multispectral, R-value, Data mining

Rights Statement

Copyright 2017, author