The Sanjiang Plain is an important grain production area in the black soil region of Northeast China. However, since its reclamation, the soil fertility of cultivated lands has declined significantly. Traditional chemical measurement methods are inefficient and difficult to meet the needs of rapid and accurate monitoring of soil fertility attributes. Spectral technology has the potential to predict soil fertility. Still, few existing studies have targeted multiple soilfertility attributes simultaneously, and the prediction accuracy of some soil fertility attributes is relatively low. Therefore, this study took the typical cropland area of the Sanjiang Plain, Youyi Farm, as the study area. We utilized visible and near-infrared spectroscopy, combined with four spectral preprocessing methods, including SG (Savitzky-Golay) spectral smoothing, first-order derivation, standard normal variate transformation, and multiplicative scatter correction, as well as the competitive adaptive reweighted sampling (CARS) band selection algorithm. The partial least squares regression model was employed to simultaneously predict four key soil fertility attributes: soil organic matter (SOM), total nitrogen (TN), total phosphorus (TP), and total potassium (TK). The study aimed to explore the potential of spectral prediction for multiple soil fertility attributes and investigate the role of variable selection in improving prediction accuracy. The results showed that: (1) When using the full spectral range (400~2 400 nm) without variable selection, the prediction accuracy of SOM and TN was relatively high, with cross-validation R² values ranging from 0.85 to 0.89 and 0.86 to 0.89, respectively. The prediction accuracy of TK was also relatively high, with R² ranging from 0.63 to 0.72, but the prediction accuracy of TP was lower, with R² ranging from 0.08to 0.34. (2) After CARS band selection, the prediction accuracy of all four soil fertility attributes improved, with the largest improvement found in TP. The optimal cross-validation R² was 0.97, 0.96, 0.82, and 0.92 for SOM, TN, TP, and TK, respectively. (3) The CARS variable selection method identified the spectral bands corresponding to the characteristic functional groups related to SOM and TN. The prediction of TN utilized both its relationship with SOM and its intrinsic characteristic bands. The prediction of TP mainly relied on soil spectral information, while the prediction of TK utilized both soil spectral information and its relationship with SOM and TN. This study demonstrated the potential of spectral technology for simultaneously predicting multiple key soil fertility attributes in the typical cropland area of the Sanjiang plain and found that variable selection can significantly improve the prediction accuracy of soil attributes (TP) that do not have obvious spectral characteristics, providing a methodological reference for rapid monitoring of soil fertility in black soil regions.