The effect of thermal modification (TM) on the color of western hemlock wood and its physical and mechanical properties were investigated. The focus of this study was the prediction of material properties of thermally modified wood based on the color change via the “group method of data handling (GMDH)” neural network (NN). The NN was trained by color parameters for predicting the equilibrium moisture content (EMC), density, porosity, water absorption (WA), swelling coefficient, dynamic modulus of elasticity (MOE_dyn) and hardness. The color parameters showed a significant correlation with temperature and are well correlated with the heat treatment (HT) intensity. Color parameters combined with the GMDH-type NN successfully predicted the physical properties of the material. The best correlation was achieved with the swelling coefficient, EMC and WA. All these properties were significantly influenced by HT. The color parameters did not seem suitable for predicting the wood hardness and MOE_dyn. The GMDH NN shows a higher model accuracy than the multivariate linear and partial least squares (PLS) regression models.