Although fertilizer plays an important role in grain yield increase, many negative effects will be brought by the overuse.Information about the soil fertility within a field is essential to the decision-making process on fertilization.Ion-selective electrode (ISE), showing obvious advantages in rapid response, inexpensive cost, simple use and easy integration, is a good method for soil macronutrients detection.However, interference of coexisting ions and external circumstances has a significant impact on measurement result and is the main obstacle for ISE in practical applications.In this paper, a ANN model was developed to reduce the interference of Cl- and temperature on NO3-Ion selective Electrode.For the artificial neural network model, three layers were decided.The output of the model was NO3--N concentration and three input parameters were temperature, response potential of ISE-NO3-and of ISE-Cl- respectively.As determination of various parameters associated with artificial neural networks and finding the optimum topology were a very time-consuming process, the model parameters were optimized by response surface methodology (RSM).In the optimizing process, the mean absolute error (MAE) of the model output was chosen as response to decide significant parameters effecting the model.The influence parameters on the MAE were evaluated by using multiple comparison analysis based on single factor experiment.Five topological parameters training percentage, the neuron number in the hidden layer, GDM algorithm parameters of training epoch number, step size, and momentum coefficient were selected as the main effect factors.The influences of training percentage (60-80),the neuron number in the hidden layer (2-30), GDM algorithm parameters of training epoch number (50-10000), step size (0.05-0.5), and momentum coefficient (0.4-0.8) were investigated using a 3-level, 5-factor face-centered central composite design (CCD) response surface experiment, which consisted of 52 runs.10 repetitions at the center point were included and used for variance analysis.The relationships between the MAE and tested parameters were quantitatively described by a ternary quadratic equation then the optimum number of hidden neurons, momentum coefficient, training epoch, step size, and training runs were found.The quadratic equation was reliable(R2=0.9698), significant(P<0.01) and lack of fit is not significant(P=0.6453 >0.05), which means the equation can be used to determine relative parameters associated with artificial neural networks and find the optimum topological parameters.Considering the time cost and generation ability together, the optimal topological parameters for ANN model were obtained as follows: number of neurons =22, training epoch =7670, step size =0.28, training percentage =65% and momentum coefficient =0.85, under which the MAE was 1.0 correspondingly.At last, a multi-layer feed-forward (MLFF) network with one hidden layer and using gradient descent with momentum (GDM) as learning algorithm was used to develop the error correct model.A verification test was implemented to test the model and the results were not different from output of the model significantly (P=0.1738 > 0.05).The mean absolute error (MAE) of verification test was less than 6.3mg/kg.In range of 10 to 40℃, the best ANN model can correct interference of Cl- within 250mg/kg while the main ion concentration ranging from 5 to 250mg/kg.For practical soil nutrient detection, the MAE could reach 8.6mg/kg and relative standard deviation was lower than 7.5% compared with 20% of no model correction, which indicated the accuracy can meet the requirement.