Color is an important soil property, and is frequently used to infer soil properties such as total iron (Fe). However, soil water affects the soil color and severely limits the estimation accuracy for Fe. To improve the Fe estimation accuracy, the water-absorption-peak-based color reconstructing machine (WCRM) method was proposed and investigated in this study. The rapid dewatering functions of the WCRM method were designed based on machine learning concepts and the two water absorption troughs of visible and near-infrared light (around 1400 and 1900 nm). Color and water data for the soil samples were obtained using an ASD FieldSpec 3 spectrometer under laboratory conditions. The optimal Fe model (1900 group; i = 0.5, j = 4.5, k = 2.5) for the WCRM method was obtained to determine the final WCRM model, which produced better results ( R 2 = 0.603, MREv = 6.6%, RMSEv = 2.914 g/kg, RPDv = 1.483, and RPIQv = 2.492) than the adaptive neuro-fuzzy inference system model. The WCRM method represents a new way of accurately estimating Fe and other soil properties. • The water-absorption-peak-based color reconstructing machine (WCRM) method was proposed. • The two water absorption peaks around 1400 and 1900 nm was used for color reconstructing. • The WCRM model was established based on reconstructed L∗a∗b∗ values. • This method can quickly reduce the influence of soil water during Fe multicolor modeling.