【Objective】 The Northeast of China is one of the most important grain production base for China. In recent years, unreasonable use of cultivated land in this region has caused a decline in soil fertility, posing a severe threat to the nation’s food security. To ensure sustainable agricultural practices, it is important to develop a rapid and reliable method for monitoring variations in soil physicochemical properties. 【Method】This study focused on the Youyi Farm in the Sanjiang Plain as the research area, aiming to evaluate the feasibility of Sentinel-2 multi-temporal remote sensing imagery in predicting key soil fertility properties. A total of 103 surface soil samples in cropland were collected, and Sentinel-2 images acquired during the potential bare-soil periods (April, May, and June) from 2019 to 2023 were selected to build a Random Forest regression model for predicting soil organic matter, total nitrogen, total phosphorus, and total potassium. To investigate the temporal effects of images on the prediction performance and to identify optimal temporal combinations for high-precision prediction of these soil fertility properties, the images were organized along two temporal axes. First, the images were divided into seven-year groups (five single-year groups: 2019, 2020, 2021, 2022, and 2023, and two multi-year groups: 2020—2022 and 2019—2023). Then, within each year-group, the images were further divided into four month-groups (three single-month-groups: April, May, and June, and one multi-month-group: April—June). Finally, the combination of the seven year-groups and four month-groups produced 28 distinct year–month temporal groupings. For each grouping, all available Sentinel-2 images were synthesized by median compositing to produce 28 synthetic images that served as inputs to the Random Forest model. 【Result】The results indicate that soil organic matter was predicted with the highest accuracy among the four fertility properties, with an R² of 0.62 and an RMSE of 6.58 g·kg-1. The prediction accuracy of soil total nitrogen was similar to that of organic matter, with an R² of 0.58 and an RMSE of 0.34 g·kg-1. Total phosphorus predictions were not sufficiently accurate for practical applications, with the highest accuracy of an R² of 0.13 and RMSE of 0.01 g· kg-1. The total potassium achieved a relatively high prediction accuracy of an R² of 0.53 and RMSE of 1.55 g·kg-1. 【Conclusion】The results in different year-groups and month-groups indicated that (1) multi-year synthetic images outperformed single-year synthetic images in prediction accuracy, and (2) the synthetic images from May showed the highest prediction accuracy among the monthly groupings. These findings demonstrate that careful temporal selection and multi-temporal synthesis of Sentinel-2 imagery can improve the prediction accuracy of soil organic matter, total nitrogen, and total potassium in the cultivated land of Northeast China. In contrast, Sentinel-2 spectral bands alone are difficult to effectively predict total phosphorus content. Integrating auxiliary environmental variables (such as topography, climate, or cultivation management) or employing alternative remote sensing data may be necessary to achieve higher accuracy. Overall, this study provides methodological guidance and technical support for regional-scale soil fertility monitoring and mapping in the Sanjiang Plain.