In soil nitrate-nitrogen (NO-3-N) detection based on ion-selective electrode (ISE), coexisting chloride ion (Cl-) is the primary interference factor. Currently most detection devices detect only single ion and are off-line. In order to solve these problems, this paper aimed at developing a soil NO-3-N detection device with multiple parameters based on ISE. The device embedded a back propagation (BP) neural network model and achieved on-line and real-time detection. Five methods were adopted to improve the BP neural network model due to its shortcomings of slow convergence rate and easily falling into local minimum. Two correction methods were used to calibrate detection results of the device. Several judgement programs were applied to improve the stability of electric potential acquisition. Standard solution tests were conducted to validate the accuracy of device. The experiments of NO-3-N detection using 20 soil samples was conducted, and the detection results were compared with that of linear regression model and optical detection to validate the effect of reducing the interference of Cl- and the soil NO-3-N detection accuracy. The results showed that the deviation between the detection results of the device and that of an ion meter was less than 1.0mV, meeting the soil NO-3-N detection accuracy requirement. The average relative error between the soil NO-3-N detection results of the device and the results detected by optical method was 8.83%, while the average relative error between the results of linear regression model and the results detected by optical method was 12.17%. The fitting coefficients R2 were both greater than 0.97. It indicated that the device could effectively reduce the interference of Cl-, had high accuracy and could be used for on-line detecting soil NO-3-N.