Abstract: The net ecosystem carbon exchange (NEE) of coastal wetland is affected by various environmental factors. The selection of parameters is very important for estimating and modeling the NEE of coastal wetland. How to reasonably select the input parameters affects not only the accuracy of the estimation results, but also the applicability of the prediction model. Four correlation coefficients were used, including the Pearson correlation coefficient, the Spearman correlation coefficient, the distance correlation coefficient, and the correlation coefficient of maximum mutual information, to calculate the correlation between various environmental factors and NEE, according to which the best combination of input parameters was chosen. Using the measurement data of the Yancheng salt marsh wetland in Jiangsu Province, eight parameter combinations with the highest correlation were selected, then eight factors were input into the convolutional neural network for model training, and finally four prediction models obtained. The root mean square error and mean absolute error were used to verify the accuracy of the model. After calculation, the root mean square errors of the four models were 0.0134, 0.0092, 0.0109, 0.0051, and the absolute errors were 0.064, 0.068, 0.0574, 0.0439, respectively. This study shows that: 1) to model the NEE of coastal wetland with the parameter combination based on the maximum mutual information coefficient, the photosynthetic effective radiation, surface radiation, net radiation, photosynthetic effective radiation, soil albedo, air temperature, relative humidity, surface temperature, and soil moisture content, the accuracy of the modeling is the best and the error is the smallest. 2) Among 15 parameters used, the net photosynthetic effective radiation, net radiation, surface radiation, and NEE are strongly correlated in the four correlation coefficients, which showed that radiation parameter had a greater impact on wetland carbon cycle than other parameters. 3) Relationship between each parameter and NEE included both linear and nonlinear relationships. The conventional single linear analysis method cannot completely and accurately reflect the response relationship between each environmental parameter and NEE. In the future works, we shall not only study the linear relationship among variables but also pay more attention to the nonlinear mutual relationship. 4) The accuracy of the coastal wetland NEE prediction model based on convolutional neural network was better than other similar models’, which shows that the model is applicate in NEE prediction modeling. This study provided a reference for NEE prediction modeling and analysis of coastal wetland in the future.