MODELLING WEATHER DEGRADATION OF WOODEN FACADES USING NIR HYPERSPECTRAL IMAGING OF THIN WOOD SAMPLES

Transcription

1 WCTE 2016 World Conference on Timber Engineering August 22-25, 2016 Vienna, Austria MODELLING WEATHER DEGRADATION OF WOODEN FACADES USING NIR HYPERSPECTRAL IMAGING OF THIN WOOD SAMPLES Ingunn Burud 1, Knut Arne Smeland 2, Thomas Thiis 3, Lone Ross Gobakken 4, Anna Sandak 5, Jakub Sandak 6, Kristian Hovde Liland 7 ABSTRACT Untreated wooden surfaces degrade differently when exposed to varying doses of natural weathering. Theis study aims to assess the degradation of wooden surfaces caused by weathering and to determin a weather dose for the degradation. Several sets of very thin wood samples have been exposed to natural weathering and one set of samples was exposed to UV-radiation in a AtlasUV. The samples have been studied with multisensory techniques including near-infrared hyperspectral imaging. Spectra of earlywood and latewood could be extracted from the hyperspectral image cubes and changes in the spectra were modeled as a function of UV solar radiation to see if the weathering deterioration was reflected in the NIR spectra. A regression model was obtained using Tikhonov regression, an algorithm that yields robust prediction models when predicting new test data. The lignin and holocellulose content were estimated on selected samples separately for early- and latewood using a thermogravimetric analysis (TGA). The thermogravimetric results showed a clear correlation with the progress of weathering of the samples, for both earlywood and latewood and both for the outdoor samples and the samples is in the UV chamber exposed. This indicates that NIR spectroscopy can also be used to model lignin content in the wood. The result from this work is a first step towards fining a weather dose model determined by temperature and moisture content on the wooden surface in addition to the solar UV radiation. KEYWORDS: wood weathering, degradation kinetics, hyperspectral imaging, transmission 1 INTRODUCTION 12 The use of wood as a building material outdoors is widespread in various applications from ancient times. Untreated wooden surfaces are frequently utilized as cladding in modern buildings. Consequently, the importance of aesthetical character and appearance of wood materials and wood products used outdoors has increased. Wood subjected to weathering is degraded by various environmental agents such as solar radiation, cyclic wetting, atmospheric temperature and relative humidity changes, environmental pollutants and certain micro-organisms. The choice of wood is often related to wood/structure aesthetics and there is an increasing 1 Ingunn Burud, NorwegianUniversity of LifeSciences (NMBU), Ingunn.burud@nmbu.no 2 Thomas Thiis, NMBU, Thomas.thiis@nmbu.no 3 Knut Arne Smeland, NMBU, knut.arne.smeland@nmbu.no 4 Lone Ross Gobakken, Norwegian Institute of Bioeconomy Research (NIBIO), lone.ross.gobakken@nibio.no 5 Anna Sandak, CNR/IVALSA, anna.sandak@ivalsa.cnr.it 6 Jakub Sandak, CNR/IVALSA, sandak@ivalsa.cnr.it 7 Kristian Hovde Liland, NOFIMA, kliland@nofima.no demand for a better understanding of the deterioration mechanisms of wood during the outdoor exposure. While it is known that the characteristic grey patina visible after a few months of exposures is mostly caused by photodegradation of lignin in middle lamella by UV radiation [1], there are several other factors involved, such as moisture, temperature, biological growth, chemicals and mechanical abrasion [2]. The elements in façades made of wood can often cause non-uniform degradation patterns such as shown in Fig. 1. Various architectonical solutions and geometrical patterns and shapes are generally the main reasons for the heterogeneous appearance since these affect the microclimate on the surface. The microclimate on a wooden façade differs from the exterior climate and can be modelled using ray tracing to account for micro scale variations of the solar irradiance, temperature and moisture on the wall, as shown in Fig. 2 and presented in [3]. To be able to predict the aesthetical service life and the visual appearance of a wooden façade over time, well performing models of wood degradation as a function of the main climatic factors are crucial.

2 Figure 1: Unevenly exposed wooden cladding (Photo:Ingunn Burud) in Norway The goal of the present study was to investigate the kinetic of the degradation rate of thin wood samples exposed to short term weathering dose (20-30 days) and to define a weather dose for the degradation. The experiment is a part of the Round Robin test conducted within COST action FP1006 Bringing new functions to wood through surface modification. Experiments with approximately 100 µm thick wood samples have been carried out. The advantage of thin samples is that they can be treated both as surface and bulk for the analysis. Several sample-sets of thin wood were exposed outdoors and one sample-set was exposed in a UV radiation chamber (AtlasUV), and the study has been focused on the influence of UV radiation on the weathering of the samples. According to [6], a widely used practice has been to use total energy from the integrated solar spectrum as the timing variable for the amount of radiation a sample has been exposed to. This is obtained by integrating over the entire spectral power curve of sunlight, and multiplying by exposure time [1]. One aim was to model the kinetics by means of near infrared (NIR) hyperspectral imaging. Previous studies have shown that NIR spectroscopy and hyperspectral imaging are well suited scientific tools for rapid and non-destructive characterization of wood surfaces [4, 5]. Modeling the degradation as a function of UV radiation using multivariate regression techniques was therefore carried out to explore how the deterioration of the earlyand latewood can be predicted as a first step to the overall goal of defining a weather dose for the degradation. Figure 2: Annual mean of solar radiation on a wooden façade (above) and annual mean of temperature a wooden façade (below). The light blue areas are the windows in the façade. (from Thiis et al. [3]). 2 MATERIALS AND METHODS 2.1 EXPERIMENTAL SETUP Experimental samples were prepared from one piece of Norway spruce wood (Picea abies) on a slicing planner (Marunaka) to a thickness of ~100µm and an exposed surface of 30mm x 35mm. Sets of samples were exposed in 15 locations in Europe and were collected before exposition and after 1, 2, 4, 7, 9, 11, 14, 17, 21, 24 and 28 days of weathering. An additional set of 105 samples were exposed outdoors in Ås, Norway, facing South at 45 degrees as shown in Fig. 3. On the first day 21 samples were put outside for exposure, and then one sample was collected each day and stored at room temperature. After 7 days a new set of 21 samples were exposed, and the following days one sample was collected each day following the same procedure as for sample set 1. This procedure was followed for 5 sample sets of 21 samples each. After collecting all samples, they were stored in darkness in a climatic chamber with a constant temperature of 20 C and 65% relative humidity to obtain an acclimatized weight before further processing. Weather data for 10 of the locations has been obtained. For the samples from Ås, Norway a national weather station is located approximately 200m away from the exposure site. This weather station has a pyranometer that measures the solar UV radiation in the wavelength region ( ) nm, which was used as response values when modelling the wood degradation from the NIR spectra. Finally, a supplementary set of 30 samples were exposed in a UV chamber of type Atlas UVTest TM. The Atlas UVTest TM has eight UVA-340 lamps installed with an output of 0.89 W/m 2 /nm each. The samples were exposed in the UV chamber for 1 to 10 cycles of 2.5h

3 continuous UV radiation followed by 30 minutes water spraying. Figure 3: Exposure of samples facing south in Ås, Norway. Figure 4: Hyperspectral camera with custom made setup for transmission mode imaging. The samples is put on the glass plate and light shines through from a light bulb in the box. 2.2 HYPERSPECTRAL IMAGING The spectral imaging measurements were conducted in a laboratory setup with a pushbroom-type hyperspectral camera (Specim, Oulu, Finland), which has a Mercury Cadmium Telluride (MCT) detector, sensitive in the near infrared region ( nm) distributed over 256 channels. One dimension of the detector is used for the spectral separation and the other for imaging one of the two spatial directions so that one line is recorded each time with a spectrum in each pixel. The second spatial dimension is obtained by moving the camera over the sample using a translation stage. On average it takes 5 seconds to scan the sample. The spatial resolution of the setup was approximately 100 µm. The hyperspectral image acquisition was carried out in transmission mode using a custom setup: backside illumination with halogen lamps below a semi opaque glass plate and another transparent glass plate transmitting NIR radiation above the samples (see Fig. 4). At the end of the scan of each sample, a short image with the shutter closed is carried out and the mean signal from each line in this dark image is subtracted from each band in the hypercube bør kanskje definers?. This procedure is performed to remove the bias level and to correct for pixel-to-pixel variations in the detector. A white calibration has to be performed in order to remove the spectral signal from the lamps. A mosaic of 8 hyperspectral images (wavelength 990 nm is shown) of selected samples exposed for 0, 3, 4, 7, 10, 17, 19 and 21 days is shown in Fig. 5a. The earlywood and latewood parts can be identified on the hyperspectral images as can be seen in Fig. 5a. In order to automatically extract spectra from the early- and latewood separately, a masking algorithm based on Principal Component Analysis (PCA) was applied to the images. A Partial Least Square Discriminant Analysis (PLS-DA) was applied to such mosaic images in order to visualize the evolution of early and latewood [8]. The algorithm of PLS-DA uses the manually selected classes of spectra and predicts corresponding classes to the remaining spectra in the dataset. Two classes (earlywood and latewood) were defined on a non-weathered sample when setting-up the PLS-DA model. An additional class was selected for pixels corresponding to the material deficiency due to cracks. The earlywood/latewoods were clearly discriminated on the samples exposed to weathering for very short time periods. On the other hand, most latewood zones on samples exposed to longer-time weathering (more than 10 days) were clearly misclassified by PLS-DA and were considered as similar to early wood. This indicates that chemical changes to wood due to weathering (as recorded in NIR spectra) homogenize the surface along woody polymers degradation. Therefore even if PLS-DA does not seem as a perfectly suitable method to classify the overall mosaic of samples with different weather exposure, it can be used as a method to demonstrate the spectral differences along the degradation progress. Figure 5: a) Hyperspectral image (wavelength 990 nm) of 8 samples exposed for periods from 0 days (upper left corner) to 28 days (lower right corner). The white areas correspond to cracks. b) Result from a PLS-DA classification where earlywood (dark blue) and latewood (light blue) classes were determined on the non exposed sample. Cracks in the samples were also classified (shown as yellow).

4 2.3 REGRESSION ANALYSIS Regression analysis was carried out for the samples exposed in Ås, Norway. The mean spectra of early and latewood separately were used as predictor variables and total cumulated UV solar radiation for each sample was used as response variable. The set of predictor variables and the response variable are all continuous, and so the point of departure is Ordinary Least Squares Regression, of the familiar form: y = β 0 + β 1 x β n x n + ε. Here β 0 is the expected value when all x = 0, β 1,, β n are the coefficients associated with each variable, and ε is the error term representing variation not accounted for by the model. In linear algebra notation the terms can be arranged as: Xβ = y where the rows of X correspond to observations and the columns correspond to variables, with a vector of ones added to account for the constant term. β is a vector containing the regression coefficients and y is the response vector. The aim of the regression is to derive a model that predicts new observations accurately and reliably, given an observation with response y from a population with an unknown mean µ, an associated explanatory variable x, and an estimate of y given by some function y = f x. A challenge with all regression models is the trade-off between increasing the model complexity or the model robustness. A model with increased complexity may predict well for the set on which it was trained, but it might not generalize well to new data. By relaxing the requirement that the model should be completely unbiased to the training data, it is possible to obtain a model that has stronger predictive power. The regression method chosen in this study named Tikhonov [7] regularization has a regularization term added to the least squares estimator in order to increase the model robustness. To check the validity of the models, the datasets were split in two, with 2/3 of each dataset assigned as training data and the remaining 1/3 as test data. The observations were assigned at random, and the model fitted to the training set was used to predict for the test set. 2.4 CHEMICAL ANALYSIS A subset of the samples from each exposure group (outdoor exposed samples and UV chamber exposed samples) were analyzed with a Simultaneous Thermal Analyzer coupled with Fourier Transform InfraRed spectrometer (STA FF9 F1 Jupiter, NETSCH, Germany) for an estimate of the lignin and holocellulose (cellulose + hemicellulose) content. The samples on which this additional measurement was performed had all been placed outside on the same day, and the samples were chosen such that the whole range from 1 to 21 days of exposure were covered, with equal intervals. For each sample earlywood and latewood were separated, using a scalpel, in order to determine the lignin and holocellulose content for each wood type separately. The samples in its individual containers in the STA are shown in Fig. 6. Figure 6: The samples cut into pieces of earlywood and latewood in the containers prior to the chemical analysis. Photo: Monica Fongen (NIBIO). 2.5 WEATHER DEGRADATION In addition to the supervised PLS-DA classification that visualised the weathering effects on the early and latewood parts of the samples, the weather degradation was modelled using various regression techniques such as Ridge regression and PLS regression on the segmented spectra of early and latewood from the PCA based masking method. The regression models were developed with two response variables: the number of days of outdoor exposure and amount of UV radiation. MATLAB (Mathworks) was used as software platform for the multivariate data analysis. Local weather records for the exposed periods were collected including sun radiation (spectrally resolved), temperature and relative humidity (RH). On that base the surface temperature and wood moisture content on the experimental samples surface were estimated using the simulation tools reported by Thiis et al. [3]. 3 RESULTS AND DISCUSSION 3.1 VISUAL APPEARANCE The appearance of the wood samples changed after only one day of outdoor exposure of natural weathering. Fig. 7 shows four of the samples exposed for 0, 2, 5 and 20 days. It can be clearly seen that the colour turns more yellow (confirmed by additional CIE L*a*b* measurement) and fibres were consequently removed from the surface with the progress of degradation. The drastic change to the chemical structure of the measured samples already after one day of exposure was more dominant for the zone of early wood, as reported in [8]. Erosion rate has been previously investigated by several authors ([9,10,11,12]) however regarding longer periods of time. It has been reported that during the first several

5 years of natural weathering, early wood eroded much more quickly than late wood [9]. A first suggestion for computing a weathering coefficient was presented in [13] where a custom algorithm for multi-sensory data fusion and computation of weathering indicator was developed. Raw data parameters obtained by different sensors were summarized considering various importances (weight). It was found that the computed weather index followed a similar trend to that of the subjective expert quality indicators, but with reduced scatter [13]. seen in Fig. 8 where the cumulated solar radiation is shown for various location. The UV radiation was therefore used as response variable when modelling the degradation by means of diverse regression techniques [15]. It was found that Ridge regression algorithm was relatively faster and more efficient than PLS [15]. The regression results of the mean spectra for early and latewood of the outdoor exposed samples as a function of total UV radiation (showed in Fig. 9 and Fig. 10) yielded at good R 2 = 0.92 and R 2 = 0.87 on the validation set for early and latewood respectively. The corresponding results for the samples exposed in the UV chamber are shown in Figs 11 and 12. Also here there are good regression performances with R 2 = 0.79 and R 2 = 0.91 on the validation set for early and latewood respectively. This means that NIR spectroscopy can be used as an indicator on the weather deterioration of the wood surface. Moreover, the hyperspectral imaging yields the spatial information in addition to the spectra, so that the models could be obtained separately for early and latewood. However, it should be noted that UV radiation alone is not enough to predict weathering degradation since it is known that the effect of a certain amount of UV energy will also depend on the moisture and temperature conditions of the surface [6]. Figure 7: Four of the samples exposed for 0, 2, 5 and 20 days (left to right). 3.2 NIR SPECTRA Spectra corresponding to the early and latewood zones at varying degradation stages were successfully extracted from the hyperspectral images. Changes in the spectra were assessed separately (for early- and latewood) due to significant differences to the morphological, chemical and physical structures of these woody components. The PLS-DA results from classes of early- and latewood selected on the non-exposed sample show that NIR spectra changes with the degradation process of the samples [8]. These differences are partly due to changes in the thickness of the samples (erosion) and partly due to changes in chemical composition occurring during exposure. The changes due to varying thickness were corrected by pre-processing the images using the Extended Multiplicative Scatter Correction (EMSC) [14]. Figure 8: Cumulated solar radiation for 28 days at 10 different locations in Europe. 3.3 REGRESSION ANALYSIS The deterioration kinetics due to weathering was modelled both with number of exposed days and with total UV radiation as response variable. The total UV radiation varies much for the samples exposed at different locations or at different time intervals as can be Figure 9: Ridge regression of earlywood, predicted versus measured UV radiation [14]

6 3.4 LIGNIN DEGRADATION Figure 10: Ridge regression of latewood, predicted versus measured UV radiation [14]. The averaged spectra corresponding to pixels representing early and late wood before weathering and at the final stage of the process were studied in [8]. The apparent differences were in the range 1830nm to 1907nm; related to -OH groups of cellulose and in -C=O of hemicelluloses respectively [8]. Other bands assigned to carbohydrates, such as 2291nm and 2328nm (related to C-O, -OH and C-H stretching) were also clearly identified. Surprisingly, spectra variations in lignin peaks (1685nm, 1672nm and 2267nm) were not as apparent as for other woody polymers, even if a track of such lignin modification is visible especially in case of early wood [8]. Thermogravimetric analysis curves have previously been usedto determine the degradation kinetics of wood as the devolatilization characteristics during the heating process reflects the chemical composition in the sample [16]. The thermogravimetric measurements of the thin samples in this experiment clearly indicate that there is a change in the wood kinetics with the weathering [17]. This is illustrated by a section of the time derivative curve of the mass loss as a function of temperature shown in Figs. 13 and 14. The section of the figures show the mass loss peak at approximately 320 C, shifting in temperature value with exposure time. Figure 11: Measured and predicted UV radiation on samples (earlywood) exposed to UV in chamber, using the regression model determined on the outdoor samples [15] Figure 13. Mass loss for earlywood relative to the non exposed sample. The plots are 1 st derivatives of the thermogravimetric curves. Each color represents different exposure time for the sample, from 1 to 20 days. Figure 12: Measured and predicted UV radiation on samples (latewood) exposed to UV in chamber, using the regression model determined on the outdoor samples [15].

7 [1] Williams, R. S. Weathering of wood. Handbook of Wood Chemistry and Wood Composites, edited by Rowell, R. M. CRC Press 2005 [2] Gobakken, L.R., Surface mould growth on painted and unpainted wood; influencing factors, modelling and aesthetic service life, (PhD) thesis, Norwegian University of Life Sciences 2009 Figure 12.. Mass loss for latewood relative to the non exposed sample. The plots are 1 st derivatives of the thermogravimetric curves. Each color represents different exposure time for the sample, from 1 to 20 days. CONCLUSION Thin samples exposed outdoors or to UV radiation showed visible degradation after very short exposure times, both for samples exposed outdoors and samples exposed to UV radiation in artificial chambers. The degradation due to weathering was successfully modeled using NIR hyperspectral imaging, which shows that this non-destructive technique is highly performant for these purposes. Regression models with good predictive abilities were obtained using Tikhonov regression with the total amount of UV exposure as response variable. Thermogravimetric analysis of a selection of the weathered samples demonstrated a clear trend in the mass loss curves with the amount of UV radiation, indicating that the lignin degradation of weathered samples could also be modeled from NIR spectroscopy. However, further thermogravimetric studies combined with density measurements of the samples needs to be carried out to confirm this. The result from the study is a first step towards a weather dose response model determined by temperature and moisture content on the wooden surface in addition to the solar radiation. Such a model is expected to be essential for predicting the future performance of wooden façades elements. Further analysis of the weather data along with improvement of the degradation assessment will be carried out in order to link the degradation progress with the defined weather doses. ACKNOWLEDGEMENT Part of this work was conducted within the project WoodBeBetter (The Research Council of Norway, Project: Increased use of wood in urban areas, pr.nr /E40) and BIO4ever (RBSI14Y7Y4), funded within a call SIR 2014 by MIUR (Italy). [3] Thiis, T, Burud, I., Kraniotis, D., Gobakken, L.R.The role of transient wetting on mould growth on wooden claddings. 6th International Building Physics Conference, Energy Procedia, 2015 [4] Sandak, A., Sandak, J., Riggio, M. Assessment of wood structural members degradation by means of infrared spectroscopy: an overview. 23, 3, DOI: /stc.1777, 2016a [5] Agresti, G., Bonifazi, G., Calienno, L., Capobianco, G., Lo Monaco, A., Pelosi, C., Picchio, R. Serranti, S. Surface Investigation of Photo-Degraded Wood by Colour Monitoring, Infrared Spectroscopy, and Hyperspectral Imaging, Journal of Spectroscopy, Article ID , 13 pages 2013 [6] Grossman, D. Errors caused by using joules to time laboratory and outdoor tests. In: ASTM Special Technical Publication 1202, p [7] Kalivas, J. H. Overview of two-norm (L2) and one-norm (L1) Tikhonov regularization variants for full wavelength or sparse spectral multivariate calibration models or maintenance. Journal of Chemometrics 26.6, pp issn : X. doi : /cem.2429, 2012 [8] Sandak, A., Burud, I., Flø, A., Thiis, T., Gobakken, L.R., Sandak, J. Hyperspectral imaging of weathered wood samples in transmission mode. International Wood Products Journal, special issue FP1101, submitted and under review 2016b [9] Williams, R. S.; Knaebe, M. T.; Sotos, P. G.; Feist, W. C. Erosion rates of wood during natural weathering. Part I, Effects of grain angle and surface texture, Wood and fiber science, 33, 1, 31-42, 2001a [10] Williams, R. S.; Knaebe, M. T.; Feist, W. C. Erosion rates of wood during natural weathering. Part II, Earlywood and latewood erosion rates, Wood and fiber science, 33, 1, 43-49, 2001b [11] Williams, R. S.; Knaebe, M. T.; Evans, J. W., Feist, W. C. Erosion rates of wood during natural weathering. Part III, Effect of exposure angle on erosion rate, Wood and fiber science, 33, 1, 50-57, 2001c [12] Sandberg, D., Söderström, O. Crack formation due to weathering of radial and tangential sections of pine and spruce, Wood Material Science & Engineering, 1, 1, 12-20, 2006 [13] Sandak, A., Sandak, J., Burud, J., Gobakken, L.R., Noel, M. Expert versus multi-sensor evaluation of wood samples after short term weathering. 47th IRG Annual Meeting, Lisbon Portugal 2016 REFERENCES [14] Kohler, A, Zimonja, M., Segtnan V. & Martens, H.

8 Standard Normal Variate,Multiplicative Signal Correction and Extended Multiplicative Signal Correction Preprocessingin Biospectroscopy. Comprehensive Chemometrics.. 2, pp , 2009 [15] Smeland, K.A., Liland, K.H., Sandak, J., Sandak, A., Gobakken, L.R., Thiis, T., Flø, A., Burud, I. NIR hyperspectral imaging in transmission mode : Assessing the weathering of thin wood samples. Submitted to Journal of NIR Spectroscopy, 2016 [16] Burud, I., Smeland, K.A., Liland, K.H., Thiis, T., Sandak, J., Sandak, A., Gobakken, L.R. Weather degradation of thin wood samples. 47th IRG Annual Meeting, Lisbon Portugal 2016 [17] Grønli, M. G., Varhegyi, G., Di Blasi, C., Thermogravitemric Analysis and Devolatiliation Kinetics of Wood. Ind. Eng. Chem. Res. 41, , 2002