Aiming at the problem of complex intensive aquaculture environment and difficult on-line detection of ammonia nitrogen concentration, a hybrid modeling method based on generalized additive model and support vector regression (SVR) is proposed. First, according to the mechanism analysis, water temperature, dissolved oxygen, pH and conductivity are selected as auxiliary variables. Second, the data is analyzed to determine the distribution of each variable, and  log function is used as the connection function to establish a generalized additive model of anomia nitrogen concentration. In order to improve the modeling accuracy, it is combined with BP neural network, random configuration network and SVR modeling methods, the hybrid soft measurement of ammonia nitrogen concentration in aquaculture water is carried out, and the results are compared. Results show that the generalized additive model can quickly find the relationship between ammonia nitrogen concentration and auxiliary variables and has higher interpretability. The hybrid model based on SVR compensation has higher precision, and the change trend of ammonia nitrogen concentration can better track the changes in laboratory test values. Compared to using the generalized additive model alone, the root mean square error of the hybrid model is reduced by 0.013; compared to using SVR alone, the root mean square error of the hybrid model is reduced by 0.005. The experimental results prove the effectiveness of the hybrid model, which provides a new method for the detection of ammonia nitrogen concentration in aquaculture.