Compression wood is formed by the living tree to compensate for external loads. It creates wood fibers with properties undesirable in sawn products. Automatic detection of compression wood can lead to production advantages. A wood surface was scanned with a spectrometer, and compression wood was detected by analyzing the spectral composition of light reflected from the wood surface within the visible spectrum. Linear prediction models for compression wood in Norway spruce (Picea abies) were produced using multivariate analysis and regression methods. The resulting prediction coefficients were implemented in a scanning system using the MAPP2200 smart image sensor combined with an imaging spectrograph. This scanning system is capable of making a pixelwise classification of a wood surface in real time. Classification of one spruce plank was compared with analysis by scanning electron microscopy, showing that the automatic classification was correct in 11 of 14 cases.