Abstract: A long-term in situ measurement cannot fully meet the requirements of soil nitrate measurement in recent years, due to the high cost and the inability. In this study, a measurement method was proposed to collect the soil solution using the titanium sintering filter cartridge, in order to detect the nitrate concentration in the soil solution under near-infrared spectroscopy. Correspondingly, the detection device was also designed in this case. Therefore, it was demonstrated that the nitrate was suitable for the analysis at 1 250 to 1 860 nm. The commercially available Near-Infrared (NIR) Light-Emitting Diode (LED) products were combined with the 940, 1 050, 1 200, 1 310, 1 350, 1 450, and 1 550 nm NIR LED. Simulated experiments of soil solution collection were carried out using self-made soil columns. Titanium sintering filter cartridge was selected as the soil solution sampler to collect the soil solution. The results show that the soil solution collection was positively correlated with the water content of the soil, but negatively correlated with the burial depth and the clay content of the soil. It infers that the titanium sintered cartridge was suitable for sandy or clayey soils with low clay content. Once the soil moisture content exceeded 20 %, the collection time was selected as 30 min. But the collection time needed to be extended, when the moisture content was low. A systematic evaluation was made to consider the volume of solution required for the test chamber and the vertical distribution characteristics of soil nitrogen. More importantly, the optimal collection time of the soil solution was required over 60 min, where the burial depth was 10 to 40 cm. In addition, the comparative tests confirmed that the titanium sintered cartridge presented a greater capacity for soil solution collection than the LLYQ-P02 soil solution sampler, suitable for lower soil moisture contents. Finally, the BP neural network was trained using a Bayesian regularisation, with the voltage data at different wavelengths as the input layer and solution nitrate-nitrogen content as the output layer. Specifically, 70 % of all data were randomly used as the training set and the remaining 30 % as the validation set to complete the training of the nitrate-nitrogen prediction model. The feasibility and stability of the detection device were tested through the performance tests (including stability, resolution, dynamic response, and field comparison tests). The results show that the voltage tended to increase with the increasing nitrate concentration, while the voltage was lower at 1 310, 1 450 and 1 550 nm. Therefore, the presence of a large absorbance confirmed that the absorption at 1 310 and 1 550 nm was influenced by the nitrate and at 1 450 nm by water. The Pearson's correlation coefficients were 0.997 and 0.995 for the training and test set of the nitrate nitrogen prediction model, respectively. The root mean square error of prediction was 3.43 during this time. The improved prediction model of nitrate nitrogen performed the best prediction than before. The relative deviations of the measuring device from the ion electrode and the colourimetric were 5.9% max. and the absolute deviation was 9 mg/L max., respectively, fully meeting the requirements for the in-situ monitoring of nitrate nitrogen in soil solution. The new device can serve as fully automatic, non-destructive (no interference with the original soil structure), long-term, and in-situ, not available with the current nitrate ion electrodes and soil nutrient quick testers. An optimal combination of operational performance was achieved, where the standard deviation of the output voltage was 0.006, and the dynamic response time for a single wavelength was 1.4 s during the long-term operation of the detection device. The finding can provide a feasible solution for the long-term automatic measurement of soil nitrate nitrogen and the construction of a water-fertilizer integration system.