Abstract:
Forest products companies in the U.S. face vigorous competition from other wood
producers around the world and other industries (steel, aluminum, plastics,
composites). To be competitive, forest companies need to control costs, sort and
allocate logs to the most appropriate markets, and recover more value at time of
harvest. Interest in log sorting based on internal wood properties is increasing.
Wood properties, such as stiffness and density, are now being considered by log
buyers. Assessing these properties in-forest and in real-time will be a challenge for
log supply managers. The utility of near infrared (NIR) technology for measuring
wood density is showing promise in laboratory conditions. The rationale behind
this study was to evaluate NIR under conditions that are similar to
field harvesting operations to estimate log density. Douglas-fir wood samples
(110 disks) were collected from the McDonald-Dunn forest and processed in the
OSU Oak Creek laboratories. Processing conditions were organized to simulate a
harvester processor environment by using a chainsaw, and then channeling the
chips with a chute to concentrate chips to move past an NIR sensor. This apparatus
was intended to mimic a sensor system fitted to a harvester head.. A rugged
Prospectra D² NIR sensor was used to collect spectral data.
The generated spectra were analyzed in two forms, as raw data (without any
transformations) and a transformed data (2nd derivative). Then, four types of
calibration models were applied to predict log density: (1) models that used tree
parameters only as a predictor (the simple model), (2) models that used NIR
absorbance data and Partial Least Squares (PLS) analysis procedures , (3) models
that used NIR absorbance data and Multiple Linear Regression (MLR) analysis
procedures, and (4) models that used a mix of NIR absorbance data and tree
parameter data and MLR analysis procedures. The goal of the models was to use
the NIR data to predict the density of the log that has been cut.
Model results were also obtained for validation (full cross validation) and
calibration sets. Data analysis suggests that correlations for calibration sets (R)
were high, but when validation was applied there were large drops in R values.
The best fit model was the simple model, the model that did not include NIR data
as predictors.
Our interpretation of why the simple model was the best fit is that there is great
variability of wood characteristics across the stem section, that there was
morphological problems associated with how we presented the samples, and that
we used a narrower spectral range of NIR compared to the range used in earlier
studies.
