Time-lapse and infrared thermographic defects/features classifications for ageing building envelopes and secondary landslide hazards

Abstract

Non-destructive testing (NDT) and remote sensing techniques in building and civil engineering have been widely adopted in defect/feature classification. Among various NDTs, research and implementation of Infrared Thermography (IRT) has increased rapidly since 2008. With its advantages of non-contact, remote from targets, effective data collection and rapid imaging of temperature distributions, IRT is commonly adopted for evaluating the conditions of building envelopes and secondary landslide hazards. For both building envelopes and secondary landslide hazards, the need for such quantitative evaluation is ever-increasing due to the ageing of buildings and extreme weather, respectively. But traditional snapshot IRT only provides qualitative results indicating the numbers and locations of the suspected defects, but not severity leading to potential hazards, such as falling tiles from height and hidden saturated water channels causing secondary landslides. The accuracy of IRT evaluation, and its usefulness are always highly in doubt. This thesis develops a novel time-lapse IRT approach on evaluating building envelopes and the hazard of secondary landslide by making use of the different rates of change of heat transfer in different materials during heat absorption (illuminated by the Sun) and dissipation (during sunset). Time-lapse measurements of thermal contrasts were conducted by a ground-based IR camera and a UAS-IR camera for building envelopes and potential secondary landslides, respectively. A novel physical approach (Thermal Decay Mapping) and a statistical approach (Supervised Machine Learning) were developed to conduct defects/features classification through a series of rigorous validation experiments. The defect classes of building envelopes were binary: debond or intact wall finishes, and defects/feature classes of slopes were rocks of different weathering states, visual/hidden cracks, moisture and vegetation. All these defects/features are of high engineering interest because ignorance would cause casualties which could be warned and prevented with the IRT approaches developed and validated in this thesis. In the physical approach (Thermal Decay Mapping), thermal images with different time stamps were processed, and linear regression was performed on the temperature curves of each pixel on a natural logarithmic scale. The values in the natural logarithmic base (for both temperature and time) were normalized to retrieve the gradient of thermal decay of each pixel and generate a thermal decay map of the target of interest (i.e. building envelopes and secondary landslide hazards). Two thresholding methods, adaptive thresholding and Otsu's thresholding, were performed to generate binary images for defect classification. The performance of these thresholding methods was evaluated by the calculation of a confusion matrix (true positive, true negative, false positive and false negative). The results showed that the defect/feature classifications were further enhanced by making use of the pixel's standard deviation (SD method). For building inspection, the results reported an average accuracy of 80%, which is considered satisfactory in NDT's world. Meanwhile, for the slope inspection, vegetation, suspected moisture and suspected cracks could be successfully segmented, although accuracy could not be evaluated due to a lack of ground truth data. In the statistical method (supervised machine learning), the classic logistic regression algorithm was adopted. The models were trained by datasets involving sample walls built with different wall finishing materials, yielding the highest, high and low emissivity values, and a rock slope with hexagonal rock columns. The results showed successful predictions on the target walls in the building datasets. For the binary images produced from the building datasets, the sizes of the suspected debonds can be estimated at an accuracy rate of 88%. This percentage can adequately predict the severity of those invisible debonds on building envelopes. But on the other hand, the prediction model for slope inspection could not return a favourable prediction. It was due to limitations like poorer spatial resolution measured by instantaneous field of view (IFOV), insufficient number of temperature data points in the training datasets, and complexity & mixture of various feature types, all of which were in comparison with datasets of building envelopes. In general, quantitative IRT analysis was proven scientifically sound, feasible and practical for defect/feature classification for evaluation of both building envelope and hazard of secondary landslides. Various experimental, environmental and material factors, including data acquisition and materials' colours and texture affecting the performance of the two physical and statistical methods were thoroughly evaluated in this thesis. The findings, analysis and conclusions clearly pave the way for the NDT industry to migrate from arbitrary and qualitative defect mapping to scientific and quantitative analysis of defect severity in building envelopes and secondary landslide hazards.