The export coefficient model has been applied worldwide to the estimation of non-point source (NPS) pollution. Determining the export coefficients (ECs) from each pollution source and different space–time progressions is problematic because of uncertainty in the ECs of nitrogen from different land-use patterns. Bayesian theory uses the prior probability distribution and likelihood data to generate a posterior probability distribution. The total nitrogen (TN) ECs and stream loss rates K (d−1) for five land-use patterns were estimated by combining published results with monthly data for ChangLe River system for 2004–08. After 104 iterations, the results had small Markov chain Monte Carlo errors and convergence was obtained. Average TN ECs for the entire watershed were 26.1 ± 8.8, 70.3 ± 9.4, 41.7 ± 6.9, 8.9 ± 1.6 and 6.2 ± 0.5 kg ha−1 yr−1 for paddy field, dry land, residential land, woodland and barren land with coefficients of variation (CVs) of 16.9, 6.31, 8.91, 13.3 and 27.9% among sub-catchments respectively. The average K value was 0.33 d−1 with a CV of 11.3%. Estimated ECs, K and the coupling water quality model were used to predict the years 2008 and 2009; the results validated the model. This Bayesian model can determine ECs using prior knowledge and monitored data, overcoming the problems of the regression model. The model facilitates explicit consideration of uncertainty in NPS management.