Resinous defects incorporated into exterior, treated, appearance products, such as barge boards and weatherboards, can cause unsightly resin bleed and resin show-through, which is a discolouration of the paint layer above the resin feature. Manufacturers, therefore, go to great lengths to eliminate resinous wood via manual grading operations and also as part of automated grading systems. Images from resinous and non-resinous wood were acquired using a near-infrared (NIR) hyperspectral camera which produced spectral data for each pixel (0.94 × 1.00 mm). NIR data from these images in the range from 977 to 1565 nm were used to generate a partial least squares (PLS) model that was able to detect resin features reliably. Wavelength regions with peaks at 1180 and 1370 nm showed a distinct difference between non-resinous and resinous wood. Using a multi-step classification process, it was possible to filter out other wood features, such as sapstain or knots that might interfere with the correct identification of resinous areas. It was possible to average the spectral information of the PLS model to 18 wavelengths and the spatial information to 10 mm/data point without significant loss of resin-prediction ability. The optimised, reduced-information model predicted resin features in 24 out of 30 shooks correctly (80%). Solely shooks with small or narrow resin features or minor show-through were predicted incorrectly, i.e. they were typically predicted as rejects even though no resin show-through feature was apparent in the painted shook.