Soil salinization is a primary factor limiting agricultural development in arid regions. Accurately predicting soil salinity in croplands using remote sensing is advantageous for precise agricultural management and sustainable development; however, limited separability between non-salinized and mild-salinized croplands in remotesensing images, combined with the impacts of environmental conditions and agricultural practices, results in existing methods exhibiting low accuracy in predicting cropland soil salinity. This study develops a two-step strategy involving classification and retrieval. Initially, a multi-cycle classification method is established, dividing the original sample set into rested sample set (R) obtained through multiple sample deletion cycles and a deleted sample set (D). Separate classifiers were trained on the R and D sample sets, and rules were generated to distinguish areas where these classifiers are applicable. Subsequently, the soil salinity in salinized croplands was predicted using the optimal remote sensing retrieval features, with Pearson correlation coefficients (PCCs) and backward feature elimination. Then, machine learning techniques were applied to predict soil salinity. The proposed method achieved superior classification accuracy for salinized cropland (Fsalinized = 0.82, Fnon-salinized = 0.87) compared with traditional methods such as random forest (RF) (Fsalinized = 0.74, Fnon-salinized = 0.64) and 1D-CNN (Fsalinized = 0.76, Fnon-salinized = 0.65). The R2 value of the proposed method in soil salinity retrieval in salinized croplands was 0.66, outperforming the R2 for the salinity retrieval results without classification obtained using traditional methods (0.23). The two-step strategy proposed in this study provides a feasible and highly accurate approach for distinguishing mild-salinized and non-salinized croplands, thus enhancing cropland soil salinity content (SSC) retrieval accuracy.