Comparison between neural networks and partial least squares for intra-growth ring wood density measurement with hyperspectral imaging

• Hyperspectral imaging allows measuring wood density at 79 μm spatial resolution.
• Partial least squares or neural networks transform hyperspectral data to density.
• Neural networks provide better results than partial least squares.
• The mean absolute percentage error for neural networks is 6.49%.
• Our method may substitute X-ray microdensitometry measurements.

In this paper, a procedure for transforming hyperspectral imaging information into intra-growth ring wood densities is presented. Particular focus was given to comparing the neural network and Partial Least Squares Regression (PLSR) processing methods. The hyperspectral measurements were performed in a wavelength range of 380–1028 nm, with a spatial separation of 79 μm. The study employed 34 samples from the same number of Pinus pinea tree samples. Density values were analyzed at a total of 34,093 positions in the samples. For neural networks, the mean absolute percentage error (MAPE) and standard deviation of absolute percentage error (StdAPE) values were 6.49% and 5.43%, respectively. For the PLSR method the MAPE and StdAPE were 6.87% and 5.70%, respectively. The neural networks allow reducing the percentage of sample positions with large errors. The proposed method for density measurement can be used for dendrochronology and dendroclimatology.