Based on the measured reflectance spectra of soil samples in the laboratory, the relationship between soil magnesium content and reflectance spectra was analyzed, and the methods of principal component regression (PCR) and partial least squares regression (PLSR) were compared with support vector machine regression (SVMR) analysis for soil magnesium content estimation, then the estimation models of magnesium content between the reflectance spectra and its transforms were established to provide the basis for soil magnesium content estimation by hyper-spectra. The results showed that the average prediction accuracy of SVMR model reached 80.96%, higher than PCR and PLSR models by 6.16% and 4.20%, respectively. The first derivative of the reflectance spectra obtained the best outcome in the different mathematical transforms with an average estimation accuracy of 80.76%, higher by 4.95% and 4.61% than reflectivity and reciprocal logarithmic transforms, respectively. Therefore, SVMR model of first derivative transform was optimal to estimate the total magnesium content with an accuracy of 84.04%.