Abstract: Soybean is one of the most important oil crops, and Argentina is the third-largest soybean producer in the world, accounting for 17% of the global soybean yield. Timely and accurate information on soybean spatial distribution is critical for ensuring global food security. Sentinel-2 multispectral data and machine learning classification models are used to investigate the potential of soybean identification in the early stage of the growing season in Argentina, with the help of Google Earth Engine (GEE). The earliest time window and optimal feature set for soybean identification are explored. Results are as follows: 1) the random forest (RF) classification model demonstrated the highest level of classification accuracy compared to the backpropagation neural network (BPNN), support vector machine (SVM), and naive Bayes (NB) models; 2) Soybean can be accurately identified as early as the end of February (filling stage), which is approximately one month before harvest; 3) The optimal feature-subset can reduce the amount of input data by 80% while maintaining high classification accuracy. The overall accuracy (OA) of the RF classification model is 85.87%, and the relative error between the estimated soybean planting area and the agricultural statistics is 3.45%. This study provided a high-precision method for early-season identification of soybeans over large scales. The results can provide a data support for early futures trading and agricultural insurance, as well as a reference for policy-making to ensure global soybean food security.