Abstract: Winter wheat is one of the most important staple food crops in China. It is of great significance to accurately and rapidly extract its planting area and spatial distribution at an early stage. Most current studies have focused mainly on the remote sensing images from a specific growth stage or the entire growth cycle. However, it is relatively scarce in the mapping of the early winter wheat. This study aims to accurately extract the spatial distribution of winter wheat at an early stage, particularly for decision-making and yield prediction. Feature selection was combined to extract the winter wheat using shapley additive explanations (SHAP) and hyperparameter optimization of a random forest (RF) model with a agenetic algorithm (GA). The study area was also taken as Jiaozuo City, Henan Province, China. Spectral, texture, polarization, and vegetation index features were constructed using multi-source remote sensing data (Landsat-8, Sentinel-1, and Sentinel-2) from 2021. According to the typical growth stages of the winter wheat, the growing season was divided into eight temporal stages. Features of each stage were ranked and then selected using SHAP values. Key RF hyperparameters (number of decision trees, number of features considered for splitting, and minimum samples per leaf node) were optimized using the GA. A comparison was then made on the classification accuracy, visual classification effectiveness, and extracted area versus statistical area across different stages. The earliest feasible time node was determined for the accurate extraction. Finally, the generalization of the model was validated in the entire Henan Province. The results showed that: (1) The optimal algorithm significantly improved the classification accuracy in the multiple stages (seedling, tillering, overwintering, jointing, and maturing period), compared with the standard RF algorithm. The accuracy increased in the range from 0.92 to 3.02 percentage oints. Furthermore, the classification also exhibited the finer visual details. (2) In stage 1 (seedling period), the classification accuracy reached 92.41% with a relative error of 16.69%, compared with the official statistical area and an R² of 0.907. While in the second stage (tillering period), the classification accuracy was improved to 96.51%, the relative error decreased to 8.81%, and the R² increased to 0.955. (3) A comparison was made on the classification accuracy and spatial distribution between early extractions (stages 1, 2) and the optimal stage (stage 5). The spatial distribution of the winter wheat was then identified as early as late November (about 7 months before harvest). The more accurate spatial distribution was obtained by late December (about 6 months before harvest). (4) The generalization of the model was evaluated in the rest 17 cities in Henan Province. The independent accuracy exceeded 92% for these cities. The total extraction area in Henan Province was 55 182.0 km², with an R² of 0.973 and an RMSE of 384.49 km², compared with the statistical area (56 907.4 km²). The high reliability and spatial generalization were achieved for the large-scale applications. The high-precision extraction of the winter wheat distribution was obtained during the tillering stage. The timeliness of the winter wheat mapping significantly promoted the decision-making and resource allocation. The finding can also provide a strong reference for the rapid response scenarios, such as the yield estimation, early warning of pests and diseases.