Near-infrared (NIR) spectroscopy technique was applied directly to the detection of pesticide phoxim residues. A sample pretreatment method was introduced. Samples were mixed with silica gel. Silica gel as a sorbent was employed to extract and enrich the low-concentration samples. Subsequently, diffuse reflection spectrum was measured on silica gel. Calibration models were developed using partial least square regression (PLSR) algorithm. Leave-one-out cross-validation was used to evaluate and compare the models. Two experiments were carried out, and the results show that 21 samples with the concentration gradient of 0.5 mg·L-1 exhibited a high correlation coefficient of cross-validation of 0.958, and a root mean square error of cross validation (RMSECV) of 0.872 mg·L-1, while 41 samples with the concentration gradient of 0.25 mg·L-1 gave a correlation coefficient of cross-validation of 0.924 and RMSECV of 1.15 mg·L-1. It is indicated that with the reduction in concentration gradient, the prediction capacity of models dropped, but there still existed a high correlation coefficient with the concentration of phoxim in the samples. The experiments proved that the sample pretreatment method with the introduction of silica gel as an absorber to enrich low concentration analyte was effective. The method was able to lower the detection limit of NIR. The developed technique has a potential application in low-concentration sample detection by NIR spectroscopy, such as pesticide residues.