Based on total nitrogen (TN) contents, spectral reflectance (R) of and their logarithm (lgR), logarithm first derivative ((lgR)'), reciprocal (1/R), reciprocal first derivative ((1/R)'), first derivative (R'), square root (√R) and reciprocal first derivative (1/(R)') transformations of 133 coastal wetland soil samples, the predicating models of soil TN contents were established by partial least squares regression (PLSR), random forest regression (RFR) and support vector regression (SVR). The results showed that:Correlations between soil TN contents and spectral forms from high to low were:(1/R)' > R' > (lgR)' > 1/R > lgR > 1/(R)' > √R > R. Correlations between soil TN contents and spectral transformations were higher than those of R, and Pearson correlation coefficient of (1/R)' was highest (0.746). R² of all models established by PLSR and SVR based on R', (1/R)', (lgR)' and 1/(R)' transformations and RFR method were greater than 0.732, indicating their applicable for soil TN content estimation, and SVR model based on 1/(R)' had the highest accuracy, with R² of 0.987, RMSE of 0.057 g/kg and MAE of 0.050 g/kg, which was the optimal model for accurately predicting TN content in coastal wetland soil.