In view of the fragmented planting landscape and complex planting structure of Chinese farmland, achieving high accuracy identification of target crops is of great importance for subsequent crop yield estimation, grain policy adjustment and national food security guarantee. Based on Google Earth Engine (GEE) remote sensing data processing cloud platform, this study proposes a planting structure extraction method applicable to different fertility stages of winter wheat. The method adopts multi-temporal Sentinel-2 images covering key fertility stages of winter wheat in 2021 as the data source, and integrating multi-dimensional feature variables, including spectral band features, index features, texture features and topographic features. In this study, the GBDT (gradient boosting decision tree) classifier was employed to extract the planting area and spatial distribution information of winter wheat at different fertility stages at the field scale. The best fertility period for winter wheat identification was discussed. In addition, the crop recognition performance of different common classification models at the field scale was compared and analyzed. The experiments were conducted in Chengu Town, Henan Province, China, and the experimental results showed that the accuracy of planting area recognition was relatively high in the standing and jointing stage (3.11-4.10) of winter wheat, with an overall classification accuracy of 94.61% and a Kappa coefficient of 92.68%. The highest recognition accuracy was achieved in the heading and flowering stage (4.11-5.10), with an overall classification accuracy of 97.01% and a Kappa coefficient was 95.52%; however, the classification accuracy was low in grain-filling and milky stage (5.11-6.10), with an overall classification accuracy of 86.23% and a Kappa coefficient of 81.33%. The results showed that the GBDT classifier could effectively extract land cover information under field-scale conditions and achieve better feature classification recognition during winter wheat’s heading and flowering stage. In addition, this study compared GBDT with traditional classifiers such as Random Forest (RF), CART (classification and regression tree) and Naive Bayesian (NB). The results show that the GBDT classifier has the best performance in feature recognition, with an overall classification accuracy of 1.20 and 5.99 percentage points higher than the RF and CART classifiers, respectively, and a Kappa coefficient of 1.61 and 8.04 percentage points higher than the RF and CART classifiers, respectively. Moreover, the NB classifier has the worst recognition precision, with an overall classification accuracy and a Kappa coefficient of 84.43% and 78.69%, respectively. The results of this study can provide effective technical support for fine-grained crop extraction at the field scale.