A method for improving winter wheat mapping accuracy based on multi-temporal feature fusion and stacking ensemble learning Authors: Wancheng Tao, Yuting Shao, Shuxian Ren, Wei Su, Guanghua Lai, Shuai Yan Published: 2026 ABSTRACT AbstractWinter wheat is a strategic staple crop underpinning national food security in China, making large-scale and accurate remote sensing mapping essential for arable land management and agricultural regulation. However, in regions such as Jiangsu Province, characterized by highly heterogeneous and fragmented agricultural landscapes, conventional remote sensing classification methods are often limited by inadequate feature representation and weak discriminative capability, resulting in suboptimal mapping accuracy. To address these challenges, this study develops a high-accuracy winter wheat mapping framework that integrates multi-temporal feature fusion and stacked ensemble learning. The Sentinel-2 time-series imagery is employed as the primary data source. Temporal profiles are reconstructed using Savitzky–Golay filtering to suppress noise while preserving phenological dynamics. The multi-dimensional feature set is constructed by combining spectral band reflectance, spectral indices, and texture metrics to capture spatio-temporal crop growth patterns. Then, A stacked ensemble learning architecture is implemented, incorporating four base classifiers: Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Trees (CART), and Gradient Tree Boosting (GTB). Subsequently, the optimized meta-learner is applied to the outputs of these base classifiers to enhance generalization capacity and model robustness. Experimental results demonstrate that the integrated feature fusion strategy significantly improves classification performance compared to single-feature configurations. The optimized stacked model achieves an Overall Accuracy (OA) of 94.74% with a Kappa coefficient of 0.9283, substantially outperforming all individual classifiers. Winter wheat distribution maps for 2021–2023 show strong consistency with statistical yearbook data, with OA of 95.31%, 94.83%, and 94.74%, and Kappa coefficients of 0.9300, 0.9272, and 0.9283, respectively, confirming the temporal stability and transferability of our model. This study establishes a robust and scalable remote sensing identification framework suitable for complex agricultural landscapes, providing methodological support for regional crop monitoring, dynamic cultivated land management, and food security assessment.Keywords: Winter wheat, Time-series imagery, Multi-dimensional feature, Ensemble learningSubject terms: Ecology, Ecology, Environmental sciences, Mathematics and computing Introduction Accurate and timely monitoring of the spatial distribution of winter wheat is crucial for ensuring national food security, supporting agricultural decision-making, and optimizing resource allocation. As the largest wheat producer in the world, China relies on winter wheat for approximately 85% of its total wheat production, meaning its yield stability holds significant implications for global grain markets1. However, in intensive agricultural regions of eastern China, such as Jiangsu Province, high landscape heterogeneity and fragmented cropping patterns present substantial challenges for conventional remote sensing methods. Complex crop rotations and overlapping phenological stages often constrain feature discriminability, limit model generalization, and hinder fine-scale spatial delineation. Therefore, developing a robust and adaptable remote sensing framework capable of accurately identifying winter wheat under complex agricultural landscapes has emerged as a key scientific and practical challenge in the field of agricultural remote sensing. In recent years, remote sensing technology has become a fundamental tool for large-scale agricultural monitoring due to its high efficiency, extensive spatial coverage, and regular revisit capabilities2,3. Among publicly available data sources, Sentinel-2 optical imagery has emerged as particularly valuable, offering high spatial and temporal resolution coupled with rich spectral information, thus enabling fine-scale crop classification4. Furthermore, cloud-based platforms such as Google Earth Engine (GEE) have greatly enhanced the efficiency of processing time-series data and conducting large-scale geospatial analyses5. Despite these advancements, accurately mapping winter wheat in areas with complex cropping systems remains a considerable challenge. Existing methods often fail to fully utilize discriminative features; for example, some studies overlook the exploitation of phenological characteristics6 or depend solely on single classifiers7, which limits model adaptability and robustness8–12. Moreover, the effective selection and integration of multidimensional features to develop robust and generalizable classification models remain unresolved issues in the field. Currently, machine learning algorithms including Random Forest (RF)13, Support Vector Machine (SVM)14, Gradient Tree Boosting (GTB)15, and Classification and Regression Trees (CART)16 have been extensively adopted in remote sensing classification tasks, showing considerable predictive performance across various applications. However, in heterogeneous agricultural landscapes, individual classifiers frequently face challenges in balancing high accuracy with stable predictions, with their effectiveness often highly dependent on feature configuration. In response to these limitations, ensemble learning techniques, especially stacking strategies, have gained recognition for their ability to enhance overall performance and improve model generalization by integrating the complementary strengths of multiple base learners17,18. Such approaches have demonstrated notable success in domains such as land cover classification19 and vegetation mapping20. Nevertheless, research focused on systematically optimizing ensemble frameworks, particularly for high-resolution winter wheat identification within complex agricultural systems, remains relatively scarce. Accurate and timely monitoring of winter wheat spatial distribution is vital for national food security, agricultural decision-making, and resource optimization. To address these challenges, this study develops a high-accuracy winter wheat mapping framework that integrates multi-temporal feature fusion and stacked ensemble learning. The framework is implemented in Jiangsu Province, using Sentinel-2 time-series imagery from 2021 to 2023 as the primary data source. Temporal profiles were reconstructed by applying Savitzky–Golay filtering is applied to effectively preserve phenological dynamics while suppressing noise. Spectral band reflectance, spectral indices, and texture metrics are fused to construct a multidimensional feature space for classification. Then, a stacked ensemble architecture was designed to leverage the complementary strengths of four base classifiers, including RF, SVM, CART, and GTB. An optimized meta-learner was applied to synthesize their outputs for final prediction, thereby enhancing the robustness and generalization of our model. Specifically, this study aims to: (1) quantify the contribution of multi-temporal feature combinations to winter wheat discrimination, (2) evaluate the performance gains of the proposed stacked ensemble model compared to individual classifiers, and (3) generate and validate high-resolution spatial distribution maps of winter wheat in Jiangsu Province from 2021 to 2023 against statistical yearbook data. In summary, these objectives establish a methodological framework with strong spatiotemporal stability to support precise crop monitoring, agricultural resource management, and food security assessment in complex agricultural landscapes. Data sources and preprocessing Overview of the study area Jiangsu Province is situated between 30°45′N and 35°20′N latitude, and 116°18′E and 121°57′E longitude, located in the lower reaches of the Yangtze River in eastern China. It is located in the lower reaches of the Yangtze River along the eastern coast, thereby forming a core area of the Yangtze River Delta. The topography is depicted in the true color map (Fig. 1) and covers an area of 107,200 square kilometers, characterized by alluvial plains and a dense network of rivers and lakes. The region has a subtropical climate that transitions into the warm temperate zone of East Asia, experiencing four distinct seasons and annual precipitation ranging from 800 to 1200 mm, thereby creating favorable conditions for wheat growth. Winter wheat in Jiangsu Province is typically sown from October to November, enters the overwintering stage from December to February, revives in March, joints in April, heads from April to May, and is harvested in June. In addition to winter wheat, other major crops in the region include rice, which is planted in summer and harvested in autumn, and corn, which is planted in spring or summer. These crops occupy the fields during different seasons than winter wheat. Jiangsu Province is an important region for agricultural production in China, with an arable land area of 45,800 square kilometers that accounts for 3.4% of the national total21. The winter wheat planting area in Jiangsu was 2,357.87 thousand hectares in 2021 (https://www.stats.gov.cn/), which is the primary focus of this study. Study area and sample distribution. For the purpose of spatial analysis, Jiangsu Province is divided into three sub-regions based on geographic location and agricultural characteristics. Northern Jiangsu includes Xuzhou, Lianyungang, Suqian, Huai’an, and Yancheng. Central Jiangsu includes Yangzhou, Taizhou, and Nantong. Southern Jiangsu includes Nanjing, Zhenjiang, Changzhou, Wuxi, and Suzhou. Data souces Sample point data In this study, 590 sample points were collected across the three study years (2021–2023), including 257 winter wheat points and 333 points from other land cover types. The samples were obtained from both visual interpretation of high-resolution imagery and field surveys. A stratified random sampling strategy was adopted to ensure adequate representation of major land cover categories. The study area was first stratified into four broad land cover types: winter wheat, impervious surfaces, water bodies, and other vegetation. Sample points were then randomly allocated within each stratum based on its proportional coverage in the study area. The final sample composition consisted of 257 winter wheat points, 180 impervious surfaces, 90 water bodies, and 63 other vegetation points. This distribution reflects the actual land cover characteristics of Jiangsu Province, where winter wheat is the dominant crop and impervious surfaces represent a significant land cover type. The distribution of sample points across the three sub-regions is as follows. Northern Jiangsu contained 313 sample points, including 154 winter wheat points and 159 non-wheat points. Central Jiangsu contained 161 sample points, including 77 winter wheat points and 84 non-wheat points. Southern Jiangsu contained 116 sample points, including 26 winter wheat points and 90 non-wheat points. This distribution reflects the actual land cover characteristics of each sub-region, with northern Jiangsu having a higher proportion of winter wheat due to its dominance in agricultural production. Sample points were collected independently for each year, ensuring that reference labels reflect the actual land cover for that specific year. For each year, the corresponding samples were randomly partitioned into training and testing sets following a stratified sampling strategy. The testing set, which was not used in model training, served to construct confusion matrices and calculate OA and Kappa coefficients. Sentinel-2 optical imagery Sentinel-2 is an Earth observation mission jointly led by the European Commission and the European Space Agency (ESA). The mission consists of two satellites, Sentinel-2A and Sentinel-2B, which together provide a five-day revisit cycle, thereby ensuring continuous time-series observations. Then, Sentinel-2 imagery was selected as the remote sensing data source for this study. They offer multispectral coverage across the visible, near-infrared, and shortwave infrared regions at spatial resolutions of 10 m, 20 m, and 60 m, making them well-suited for detecting winter wheat phenology and for fine-scale classification of small agricultural fields3. Compared to Landsat imagery, Sentinel-2 exhibits clear advantages in temporal frequency, spatial resolution, and classification performance4. These advantages are reflected in three main aspects. First, Sentinel-2 provides richer spectral information through 13 spectral bands, which enhances the separability between winter wheat and other land-cover types. Second, its higher spatial resolution, with key bands at 10 m, enables more accurate discrimination of small agricultural parcels and reduces mixed-pixel effects. Third, the shorter five-day revisit cycle supports the acquisition of denser and more complete time-series data over the crop growing season. Sentinel-2 Level-2A surface reflectance products from January 1, 2021, to December 31, 2023, were used in this study. These products have undergone atmospheric correction and are suitable for quantitative remote sensing analysis and land-cover classification. Google Earth Engine cloud platform Google Earth Engine is a sophisticated cloud-based geospatial analysis platform that is recognized for its petabyte-scale multi-source remote sensing data storage capacity, robust parallel computing capabilities, and collaborative open environment22. As the premier cloud computing platform for remote sensing globally, GEE has demonstrated considerable advantages across various research domains, including disaster monitoring, agricultural remote sensing, climate change studies, water resource management, urban development, and biodiversity conservation. In this study, the GEE platform is leveraged to process Sentinel-2 time series image data (European Union/ESA/Copernicus/S2_HARMONIZED), efficiently completing essential steps such as image preprocessing, feature extraction, and classification model training through its cloud-based data storage and parallel computing capabilities23. This cloud-based computing approach effectively mitigates the bottlenecks associated with data downloading and local computation in traditional remote sensing analysis, thereby significantly enhancing research efficiency. Statistical yearbook of prefecture-level cities in Jiangsu province The municipal winter wheat cultivation dataset for the study area from 2021 to 2023 was obtained from the China County Statistical Yearbook. This dataset served as a reference for evaluating the winter wheat planting area estimated by the model. The effectiveness of identifying winter wheat planting areas in Jiangsu Province using time-series remote sensing images was assessed by comparing the collected statistics with the model output. Data preprocessing Cloud masking and radiometric calibration In remote sensing image analysis, atmospheric conditions, such as cloud cover and aerosols, often distort the spectral information of features, thereby affecting classification accuracy. To enhance data quality and to improve the accuracy of subsequent identification results, this study implements a series of optimization measures to mitigate atmospheric interference during the data preprocessing stage. First, the QA60 quality control band in the Sentinel-2 image is utilized to extract cloud and high-level cloud information identified by the 10th and 11th bits through bitwise operations. This enables cloud masking at the image element level, effectively identifying and excluding regions affected by clouds24. Secondly, to standardize the data scale, the pixel values of all image bands are divided by 10,000 to fall within the interval of 0 to 1, thus facilitating subsequent feature extraction and classification algorithm processing. Additionally, the radiometrically corrected Sentinel-2 HARMONIZED surface reflectance product was selected for this study, which eliminates interference from atmospheric scattering, inter-sensor differences, and variations in solar altitude angle. Preprocessing steps, including de-clouding, splicing, and cropping, were performed in GEE to ensure the comparability and consistency of multi-temporal and multi-scene images in the time dimension. These preprocessing steps are commonly applied to mitigate potential data quality issues, and ensure consistency in the time-series imagery for subsequent winter wheat identification. Reconstruction of time-series vegetation indices The Savitzky-Golay filter is a time-series smoothing method based on local polynomial least-squares fitting. It effectively suppresses high-frequency noise while preserving the overall trend and key phenological features of the data25–27. In this study, the SG filter was applied to reconstruct time series from Sentinel-2 imagery. The calculation formula is shown in Eq. (1). where, is the filtered pixel value at time i; i is the observed value in the original sequence; denotes the polynomial fitting coefficient. k is the polynomial degree (set to 3 in this study). And m is the half-window size, defining a window containing 2 m + 1 observations. The Savitzky-Golay filter was implemented within the Google Earth Engine environment using custom JavaScript code. Prior to filtering, a linear interpolation step was applied to generate a regular time series with a 5-day temporal resolution. A polynomial order of 3 was used, and the sliding window width was set to 15 days, covering 7 time points given the 5-day temporal resolution, to balance noise removal and preservation of phenological dynamics. For each target pixel, a local window was constructed centered on each time point, and polynomial least-squares fitting was performed within the window to obtain the fitting coefficients. The fitted value was then used as the smoothed reconstruction result for that time point. This method effectively suppresses high-frequency noise, defined as temporal fluctuations caused by residual cloud contamination, atmospheric variability, and sensor artifacts that do not reflect actual crop growth patterns, while preserving key phenological features. After cloud masking and interpolation, a total of 1384 Sentinel-2 images were retained for 2021, 1393 for 2022, and 1218 for 2023, covering the winter wheat growing season from November 1 to July 1 of the following year. These images were used to reconstruct a regular time series with a 5-day temporal resolution for each feature. The Savitzky-Golay filter was then applied to this regular time series to suppress high-frequency noise while preserving phenological dynamics. Research methodology In this study, a multi-feature fusion stacked integrated model was developed to delineate the framework for identifying winter wheat planting areas, as illustrated in Fig. 2. This process involves five main steps: first, the GEE platform was utilized to acquire multi-temporal Sentinel-2 images. The QA band facilitated cloud masking, followed by the application of the SG filter to reconstruct the time series of vegetation indices, including Photochemical Reflectance Index (PSRI), B8A, and Soil Adjusted Vegetation Index (SAVI). Second, spectral band reflectance features, spectral indices features, and texture features were extracted to construct the classification feature set9. Next, classification models including RF, SVM, CART, and GTB were constructed, and their performance was optimized through an integration strategy28. The classification results were quantitatively evaluated based on the optimal model. Finally, utilizing the optimal model identification strategy, the spatial distribution mapping of winter wheat planting areas for 2021–2023 was completed. The classification results were then analyzed and verified against the statistical data from the yearbook. Technical flowchart. Feature set construction Spectral band reflectance features The spectral reflectance features were derived from ten Sentinel-2 bands, specifically B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, which span the visible to short-wave infrared spectrum29. These bands capture the distinct reflectance properties of various land cover types and collectively constitute the spectral feature set utilized in this study. Spectral index features Spectral indices are calculated through the combination of multiple spectral bands, providing multi-dimensional indicators of vegetation physiological status, biomass levels, and growth vigor30–32. For this study, a set of representative indices was selected, the Normalized Difference Vegetation Index (NDVI)33, Green Chlorophyll Vegetation Index (GCVI)30,34, PSRI35, Normalized Difference Water Index (NDWI)36, SAVI37, Index-Based Built-Up Index (IBI)38, Ratio Vegetation Index (RVI), and Difference Vegetation Index (DVI)39. Their corresponding formulas are listed in Table 1. The synergistic application of these indices captures the diverse spectral signatures of various land cover types, thereby supplying essential multi-dimensional feature inputs for subsequent computational modeling and analysis. Table 1Spectral indices and their formulas.IndicesFormulationNormalized difference vegetation indexGreen chlorophyll vegetation indexPlant senescence reflectance indexNormalized difference water indexSoil adjusted vegetation indexIndex-based built-up indexRatio vegetation indexDifference vegetation indexDVI=NIR-Red Spectral indices and their formulas. Texture features Texture features are primarily used to capture the spatial structure information of various features within the image, which is crucial for distinguishing similar features40–42. In this study, the Gray Level Co-occurrence Matrix (GLCM) is employed for extracting texture features, including Contrast, Variance, Single Average Gray, and Difference Variance. Contrast quantifies the difference between gray levels within an image, as illustrated in Eq. (2): where, P(i,j) denotes the probability of co-occurrence between gray level i and gray level j in the gray level co-occurrence matrix, while are gray level indices that typically range from 0 to 255, and N refers to the total number of gray levels. The variance measures the degree of gray level variation in the image, as shown in Eq. (3): The Mean Gray Scale refers to the average gray value of all pixels in the image, calculated as illustrated in Eq. (4): where, μ represents the mean gray value, i denotes the gray value, and indicates the probability from the co-occurrence matrix. Difference Variance is employed to quantify the uniformity of pixel gray-level distributions within an image, as expressed in Eq. (5). Difference Variance where, |i-j| denotes the absolute value of gray level difference;N denotes the total number of pixel pairs within the neighborhood. Texture features are incorporated in this study to enhance the differentiation of features exhibiting similar spectral characteristics. All extracted features were combined into a unified multi-temporal feature set that served as the common input for all base classifiers. The feature set comprised three categories: spectral band reflectance from ten Sentinel-2 bands, eight spectral indices (NDVI, GCVI, PSRI, NDWI, SAVI, IBI, RVI, and DVI), and texture features (contrast, variance, mean gray, and difference variance) extracted from the B8A band using the gray level co-occurrence matrix.All features were derived from multi-temporal Sentinel-2 imagery and reconstructed using the Savitzky-Golay filter to preserve phenological dynamics. Temporal profiles across the winter wheat growing season were retained for each feature, enabling classifiers to leverage phenological patterns as key discriminative information. Classifier construction Random forest Random Forest is an ensemble learning algorithm that constructs multiple decision trees using bootstrap sampling and random feature selection13. It effectively handles high-dimensional data and provides robust classification performance by reducing overfitting through model averaging. In the training process, each tree is grown on a bootstrap sample drawn with replacement from the original training set. At each node of the tree, a random subset of features is selected as candidates for splitting, which decorrelates individual trees and enhances model diversity. The final classification result is determined by majority voting across all trees. Feature importance is evaluated through the mean decrease in Gini index across all trees. For each feature, the total reduction in node impurity contributed by splits using that feature is aggregated and normalized, providing a quantitative measure of its contribution to classification accuracy43. Support vector machine Support Vector Machine is a supervised learning method that constructs an optimal hyperplane in the feature space to separate different classes14. It maximizes the margin between the hyperplane and the closest points of each class, thereby improving generalization ability. For nonlinear classification problems, SVM employs kernel functions to map the original data into a higher-dimensional space where a linear separation becomes feasible. A radial basis function kernel was adopted in this study, with hyperparameters optimized via grid search. SVM is well suited for high-dimensional remote sensing data due to its effectiveness in handling nonlinear relationships and limited sample sizes. Classification regression tree The Classification and Regression Trees are nonparametric decision tree algorithm that constructs a binary tree model by recursively partitioning the feature space, making it widely utilized for both classification and regression tasks. In classification tasks, the Gini index is minimized as the classification criterion, while in regression tasks, the mean of the leaf node samples serves as the output. The model is trained without assuming a specific data distribution, which enables it to effectively handle nonlinear relationships. To mitigate overfitting, pruning strategies are commonly employed to simplify the tree structure44. In this study, CART utilizes a multi-class probability output to enhance the stability and interpretability of the classification results, making it suitable for classification and recognition tasks involving high-dimensional remote sensing data. Gradient tree boosting Gradient Tree Boosting is an ensemble learning algorithm in which multiple weak learners are constructed iteratively to progressively fit the residuals of the previous model iteration, thereby enhancing overall prediction performance. The core concept of GTB involves using gradient descent to optimize the objective function, allowing each new tree to correct errors along the negative gradient direction and enhancing the model’s ability to express nonlinear relationships and feature interactions. GTB balances model complexity and generalization ability by controlling parameters such as tree depth and learning rate, which makes it highly adaptable15. However, it is more sensitive to hyperparameter settings, and the training overhead is higher, requiring careful fine-tuning to achieve optimal performance. Stacking ensemble model The Stacking model is an ensemble learning strategy in machine learning that combines the prediction results of multiple base learners, thereby enhancing the overall prediction performance and generalization ability of the model18,45. This method employs a hierarchical architecture consisting of multiple base models and a meta-model46. The adoption of a two-stage stacking strategy, rather than a single-stage classification approach, is motivated by the complementary inductive biases of diverse base learners. Different machine learning algorithms exhibit distinct strengths: RF excels in handling high-dimensional features through bootstrap aggregation, SVM maximizes class separation margins in high-dimensional spaces, and GTB captures complex nonlinear interactions via sequential error correction47. From the perspective of bias-variance trade-off, stacking reduces both bias and variance simultaneously: the meta-learner corrects systematic errors (bias) from individual base classifiers, while the combination of diverse models reduces prediction variance by avoiding overfitting to specific training samples. This theoretical property makes stacking particularly suitable for heterogeneous agricultural landscapes, where no single classifier consistently maintains optimal performance across varying feature subspaces and phenological stages. The stacking process consists of two stages. In the first stage, the unified multi-temporal feature set described in Section "Feature set construction", comprising spectral band reflectance, spectral indices, and texture features, served as the input to each base classifier (RF, SVM, CART, and GTB). Each base classifier then output class probability vectors for every sample. In the second stage, these probability vectors from all base classifiers were concatenated to form a new feature matrix, which served as the sole input to the meta classifier for final prediction. By integrating the structures and learning mechanisms of different models, stacking effectively reduces the bias introduced by individual models, mitigates the risk of overfitting, and enhances adaptability to complex data distributions. Quantitative evaluation metrics In this study, Overall Accuracy (OA) and the Kappa coefficient are utilized to quantitatively assess the classification accuracy. Overall Accuracy denotes the ratio of correctly classified pixels to the total number of pixels. This ratio reflects the overall correctness of the classified image, and the corresponding formula is provided in Eq. (6): where, represents the number of image elements in the confusion matrix that are accurately classified for the category, K denotes the total number of categories, and N indicates the total number of samples. The OA value ranges from 0 to 1, with a higher value implying greater accuracy in categorization. The Kappa coefficient is a statistical measure that quantifies the difference between classification results and random consistency. Unlike OA, the Kappa coefficient accounts for chance agreement and provides a more objective representation of classification accuracy. The formula is presented in Eq. (7): where, refers to the probability of random consistency, as shown in Eq. (8): where, denotes the sum of the row in the confusion matrix, representing the number of image elements classified as the class; denotes the sum of the column, indicating the number of image elements that actually belong to the class. In addition to OA and Kappa, producer’s accuracy and user’s accuracy are reported to provide a more comprehensive assessment. Producer’s accuracy, also referred to as omission error, is defined as the proportion of reference pixels correctly classified for a given class. User’s accuracy, also referred to as commission error, is defined as the proportion of classified pixels that match the ground truth. These metrics are derived from the confusion matrix and follow good practices recommended in the remote sensing community48. Experimental design and comparison strategies All experiments were conducted on the GEE platform. A total of 590 sample points (257 winter wheat and 333 other land cover types) were collected across the three study years (2021–2023). For each year, the corresponding samples from that year were randomly partitioned into training (70%) and testing (30%) sets using stratified random sampling, with five repetitions to ensure result stability. Separate models were trained independently for each year to account for interannual variability in phenology and environmental conditions. Hyperparameters for each base classifier were optimized using grid search with five-fold cross-validation. The search ranges were as follows: for RF, the number of trees ranged from 50 to 200; for SVM, the regularization parameter C ranged from 0.1 to 10 and the kernel coefficient gamma from 0.01 to 1; for CART, the maximum tree depth ranged from 5 to 20; for GTB, the learning rate ranged from 0.01 to 0.3, the number of estimators from 50 to 200, and the maximum tree depth from 3 to 10. Three comparison strategies were designed. First, seven feature configurations were compared to evaluate feature contributions: spectral bands only (M1), spectral indices only (M2), texture features only (M3), spectral bands with spectral indices (M4), spectral bands with texture features (M5), spectral indices with texture features (M6), and the full feature set (M7). These configurations are summarized in Table 2. Second, individual base classifiers were compared against the stacking ensemble to quantify performance gains. Third, one-stage classification (individual classifiers) was compared with two-stage stacking classification (class probability outputs concatenated as inputs to a meta classifier) to demonstrate the effectiveness of the hierarchical architecture. OA and Kappa coefficient were used for quantitative evaluation. Table 2Feature configurations for classification comparison.ConfigurationFeature compositionM1Spectral band reflectance onlyM2Spectral indices onlyM3Texture features onlyM4Spectral band reflectance + Spectral indicesM5Spectral band reflectance + Texture featuresM6Spectral indices + Texture featuresM7Spectral bands + Spectral indices + Texture features Feature configurations for classification comparison. Experimental results and analysis Feature importance assessment The feature importance assessment results based on the Gini index are shown in Fig. 3. Among the 22 constructed features, which encompass spectral band reflectance features, spectral indices features, and texture features, the PSRI exhibited the highest importance score. It contributed the most to the average decrease in the Gini index, indicating its critical driving role in the winter wheat classification decision-making process. This result provides a quantitative basis for the optimization of model feature selection and architecture. Distribution of feature importance based on gini index. Notes: B8A-Con represents the contrast feature of texture; B8A-Var denotes the variance of the texture feature; B8A-Savg refers to the average gray level of the texture feature; B8A-Dvar indicates the differential variance of the texture feature. This study selected the three features with the highest contribution scores, namely PSRI, B8A, and SAVI, to further analyze their temporal variation patterns. As shown in Fig. 4, the SG filtered time-series curves effectively suppressed noise and clearly reflected the complete growth cycle of winter wheat. Temporal analysis reveals that during the early growth stage from January to February, all indices remained at low levels. As growth progressed, both B8A and SAVI peaked during the heading stage from April to May, while PSRI maintained higher values during the overwintering period from December to February, declined in spring, and rose again during the grain-filling stage in May to June. Original and SG-reconstructed time series of PSRI, B8A, and SAVI for a representative winter wheat sample point. Among these temporal patterns, the heading stage from April to May exhibited the most pronounced spectral differences between winter wheat and other land cover types, as shown in Fig. 5. During this period, B8A and SAVI reached their peak values, while PSRI showed a distinct decline following an early-season peak. These temporal patterns correspond to critical phenological stages of winter wheat and provide the strongest discriminative signals for classification. In contrast, the overwintering stage from December to February showed less distinct spectral separation, reflecting reduced vegetation activity during this period. Figure 5 presents the SG filtered time-series curves of PSRI, B8A, and SAVI for five land cover categories from November of each year to June of the following year. The temporal characteristics displayed in the three subfigures reveal that winter wheat exhibits distinct patterns in index fluctuation, peak timing, and amplitude throughout its growth cycle compared to other land cover types such as grassland and forest. These differences in the temporal characteristics across multiple features are utilized to enhance the identification accuracy of winter wheat. PSRI, B8A, and SAVI time series for five types of land cover. Classification of different feature combinations The performance of various feature combinations is compared and analyzed regarding remote sensing image classification accuracy in this study. The classification accuracy results of different feature combination models are presented in the radar map shown in Fig. 6, while the local visualization results are illustrated in Fig. 7. Classification accuracy of remote sensing images with different feature combinations. Notes: The specific feature configuration is shown in Table 2. Classification results for different feature combinations with true-color sentinel-2 background. As shown in Fig. 6, the full-feature combination model, which integrates spectral band reflectance, spectral indices, and texture features, performs optimally in classifying winter wheat farmland, achieving an OA of 93.10% and a Kappa coefficient of 0.8724. In contrast, the classification accuracy obtained by using only a single texture feature is the lowest, with an OA of 73.95% and a Kappa coefficient of 0.7031. Notably, the recognition accuracy of the single vegetation index feature reaches an OA of 89.66% and a Kappa coefficient of 0.8686, significantly contributing to the accuracy improvement of the full-feature combination. The advantage of the full-feature combination primarily stems from the synergistic and complementary mechanisms of multiple features. By retaining the original spectral information, key discriminative features are enhanced by the vegetation index, while spatial structural information is introduced to improve classification accuracy. A comparison of the subplots in Fig. 7 shows that the full‑features image displays reduced fragmentation and sharper boundaries. This result highlights its strong robustness in complex agricultural landscapes and its ability to effectively overcome the limitations of using only a single type of feature. Classification by different combinations of classifiers The performance of four machine learning algorithms—RF, SVM, CART, and GTB—is systematically evaluated for land cover classification of Sentinel-2 remote sensing imagery using the GEE platform and a stacked integration strategy for meta-model selection. As shown in Table 3, GTB exhibits the best classification performance among the four algorithms, achieving an overall classification accuracy of 94.19% and a Kappa coefficient of 0.8747. Based on these results, GTB is selected as the meta-model in the stacked integration model of this study to achieve high-accuracy identification of the spatial distribution of winter wheat cultivation in the study area. Table 3Comparison of classification accuracy of four baseline machine learning algorithms used as base learners in the stacking ensemble framework.ClassifierOAKappaRF93.10%0.8724SVM91.67%0.7886CART92.68%0.8216GTB94.19%0.8747 Comparison of classification accuracy of four baseline machine learning algorithms used as base learners in the stacking ensemble framework. The performance of multiple classifier combinations for remote sensing image classification is compared in constructing the base model, with classification results presented in Fig. 8 and quantitative evaluation metrics in Fig. 9. Experimental results demonstrate that the combination of RF, SVM, and GTB achieves the highest classification accuracy for winter wheat farmland, with an overall accuracy (OA) of 94.74% and a Kappa coefficient of 0.9283. Notably, the full combination incorporating all four base classifiers yields inferior performance compared to the RF + SVM + GTB configuration. Classification results of remote sensing images using different model combinations. Quantitative evaluation results of stacking mode. The selection of the optimal base classifier combination is guided by principles of model diversity and complementarity. CART, being structurally similar to RF as a tree-based algorithm, introduces redundant predictive information that does not enhance the ensemble’s discriminative capacity. In contrast, the RF + SVM + GTB combination captures distinct and complementary decision boundaries: RF ensures ensemble stability through bootstrap aggregation, SVM enforces margin-based class separation in high-dimensional space, and GTB sequentially optimizes residual errors to capture complex nonlinear patterns. This diversity enables the meta-learner to synthesize more robust and generalizable predictions, as reflected in the superior performance of the RF + SVM + GTB configuration over the full combination. Spatiotemporal distribution and quantitative evaluation of winter wheat Based on the experimental results presented in Sects. 4.1, 4.2 and 4.2, the complete feature set consisting of spectral band reflectance, spectral indices, and texture were selected, and the optimal classifier combination of RF, SVM, and GTB constructed using the stacked integration strategy was applied to identify winter wheat planting areas in Jiangsu Province from 2021 to 2023, generating a 10-m spatial resolution winter wheat distribution product. The model accuracy evaluation results indicate the OA of 95.31% and a Kappa coefficient of 0.9300 in 2021, 94.83% and 0.9272 in 2022, and reaching 94.74% and 0.9283 in 2023. These results demonstrate good temporal stability and adaptability to varying data conditions, as shown in Fig. 10. To provide a more comprehensive assessment of classification performance, confusion matrices and per-class accuracy metrics for each year are presented in Tables 4. Producer’s accuracy and user’s accuracy for winter wheat consistently exceeded 92% across all three years, with misclassifications primarily limited to confusion with other vegetation. Based on the analysis of local classification results shown in Fig. 11, the study area over the three years indicates that the winter wheat planting area changed smoothly, exhibiting a stable spatial pattern and strong continuity in planting. Remote sensing classification accuracy from 2021 to 2023. Table 4Confusion matrix and per-class accuracy from 2021 to 2023.True classPredicted classWinte wheatImpervious surfacesWater bodiesOther vegetationUA2021winte wheat3200196.97%Impervious surfaces01500100%Water bodies00110100%Other vegetation100480%PA96.97%100%100%80%2022Winte wheat2600092.86%Impervious surfaces01700100%Water bodies0060100%Other vegetation2007100%PA100%100%100%77.78%2023Winte wheat2300195.83%Impervious surfaces01900100%Water bodies0090100%Other vegetation100480%PA95.83%100%100%80% Confusion matrix and per-class accuracy from 2021 to 2023. Remote sensing images of certain areas in Huai’an city from 2021 to 2023. Wheat planting distribution maps for Jiangsu Province from 2021 to 2023 were generated and are shown in Fig. 12. Northern Jiangsu, which serves as the core wheat-producing region of the province, exhibited a markedly larger planting area than central and southern Jiangsu. The comparison with statistical yearbook data for the same period is presented in Fig. 13, showing that the minimum difference among the three years was approximately 270,000 hectares, indicating relatively stable interannual variation in winter wheat planting area. The relative errors between the remotely sensed estimated areas and the statistical areas from 2021 to 2023 were 16.83%, 12.17%, and 11.33%, respectively. These area-based comparisons provide a complementary perspective to the pixel-level accuracy metrics. Thematic map of wheat planting distribution in Jiangsu Province from 2021 to 2023. Comparison of wheat area between Jiangsu provincial statistical yearbook and remote sensing classification algorithm. In addition, a comparative analysis was conducted at the prefecture-level city scale. The left panel of Fig. 14 shows the choropleth representation of officially reported winter wheat areas across prefecture-level cities, while the right panel presents the corresponding stratified statistical visualization. These comparisons reveal spatial variations in estimation consistency. For instance, in 2023, the estimated winter wheat planting area in Huai’an differed from the statistical yearbook data by approximately 30,000 hectares, while in 2022, a difference of approximately 24,000 hectares was observed in Nanjing. Such discrepancies reflect the influence of spatial heterogeneity and data quality on model performance, particularly in southern Jiangsu where frequent cloudy and rainy conditions constrain optical imagery availability. Nevertheless, at the provincial scale, the relative errors remained consistently below 16.83%, confirming the overall reliability of the proposed framework for regional crop monitoring. Comparison of statistical and remote sensing-based wheat area in jiangsu prefectural cities over three years. Discussion Based on multi-temporal Sentinel-2 imagery acquired between 2021 and 2023, a multi-dimensional feature set was constructed by integrating spectral, index, and texture features, and an RF + SVM + GTB stacked ensemble classification strategy was adopted to achieve accurate and efficient identification of winter wheat planting areas in Jiangsu Province. It is important to acknowledge the simplified nature of the classification scheme adopted in this study. The primary objective was to distinguish winter wheat from other land cover types, including impervious surfaces, water bodies, and other vegetation. Notably, winter wheat occupies the field during its growing season (typically November to June), a period when other major crops such as rice and corn are not present in the study area. This temporal separation reduces spectral confusion with summer crops. However, other winter crops or non-crop vegetation may still pose classification challenges. Extension to multi-crop classification, particularly separating winter wheat from other winter crops, represents an important direction for future research. Spatiotemporal stability verification of model performance During three consecutive years of experimental validation in Jiangsu Province, the stacked ensemble classification model developed in this study exhibited strong spatiotemporal stability. From the temporal perspective, the OA reached 95.31% with a Kappa coefficient of 0.9300 in 2021, increased to 94.83% and 0.9272 in 2022, and remained at 94.74% and 0.9283 in 2023. High accuracy levels were maintained over the three years with low interannual variability, indicating strong adaptability of the model to time-series classification. From the spatial perspective, independent validation was conducted across southern, central, and northern Jiangsu Province, where stable classification performance was observed in all regions, as presented in Table 5. From 2021 to 2023, OA exceeded 87% in all regions, with Kappa coefficients consistently higher than 0.81. In particular, northern and central Jiangsu Province, characterized by contiguous farmland and flat terrain, exhibited superior performance, with OA consistently exceeding 92%, whereas the more heterogeneous southern Jiangsu Province also achieved accuracy levels above 87%. Table 5Comparison of independent validation accuracy across southern, central, and northern Jiangsu province.Region202120222023OAKappaOAKappaOAKappaSouthern Jiangsu87.10%0.816390.48%0.875092.00%0.8986Central Jiangsu96.08%0.931892.00%0.881695.56%0.9327Northern Jiangsu95.00%0.921991.00%0.849697.37%0.9603 Comparison of independent validation accuracy across southern, central, and northern Jiangsu province. Several factors may contribute to this difference. Northern Jiangsu features contiguous farmland and flat terrain, which facilitate accurate classification. Central Jiangsu, despite having a lower proportion of winter wheat samples, also benefits from relatively homogeneous agricultural landscapes. In contrast, southern Jiangsu is characterized by more heterogeneous land cover, with fragmented agricultural fields interspersed with urban areas, which naturally increases classification difficulty. However, the sample distribution across sub-regions should also be considered as a potential factor. Northern Jiangsu had a relatively balanced class distribution, with 154 winter wheat and 159 non-wheat samples. Central Jiangsu had 77 winter wheat and 84 non-wheat samples, also maintaining a near-balanced distribution. Southern Jiangsu, in contrast, had a highly imbalanced distribution, with only 26 winter wheat samples compared to 90 non-wheat samples. This imbalance may lead to a model that is less sensitive to winter wheat in the south, potentially contributing to the lower accuracy observed. Therefore, the accuracy differences across sub-regions likely reflect a combination of landscape characteristics and sample distribution effects. Future studies with more balanced sampling across sub-regions would help disentangle these factors. These results demonstrate that the multi-feature fusion and stacked ensemble strategy adopted in this study exhibits strong spatiotemporal stability and regional adaptability. Crop rotation, particularly the alternation between winter wheat and summer crops such as corn or rice, is a common practice in the study area and may introduce temporal inconsistency in land cover labels across years. In this study, sample points were collected independently for each year, ensuring that reference labels reflect the actual land cover for that specific year. While crop rotation could potentially affect pixel-level classification accuracy, the strong agreement between estimated areas and statistical yearbook data (relative errors below 16.83%) suggests that quantity estimation remains robust. Nevertheless, the impact of crop rotation on spatial classification accuracy warrants further investigation using temporally consistent sample sets. Performance advantages of the stacked ensemble model The stacked ensemble model exhibits superior classification performance compared with individual models in complex agricultural landscapes. As shown in Table 3, the stacked model integrating RF, SVM, and GTB achieves the OA of 94.74% and a Kappa coefficient of 0.9283, corresponding to an improvement of approximately 3%−5% over the best-performing individual model. This advantage arises from the synergistic effects of multiple models, in which the stability of RF in handling high-dimensional features, the discriminative capability of SVM in high-dimensional spaces, and the capacity of GTB to capture complex patterns effectively complement one another. Through optimized integration via a meta-learner, a more robust decision-making mechanism is established. In confronting spectral variations during key growth stages of winter wheat, the stacked model effectively integrates the complementary recognition strengths of its base learners. This integration enhances discriminative capability in regions with ambiguous classification boundaries, thereby validating the applicability of the ensemble strategy for multi-temporal crop classification. Regional applicability of the proposed crop mapping framework The application and validation of the proposed technical framework in Jiangsu Province establish a solid foundation for regional extension. Based on intelligent processing implemented on the GEE cloud platform, winter wheat area estimates from 2021 to 2023 were compared with official statistical yearbook data, yielding relative errors consistently below 16.83%, which confirms the reliability and timeliness of the model at the provincial scale. Importantly, the proposed framework is not designed specifically for irrigation systems. Instead, it is built on crop phenological characteristics and spectral-spatial features. These elements are broadly applicable across diverse agricultural landscapes. Jiangsu Province, situated in the transitional zone of major winter wheat production areas in China, shares agro-ecological conditions representative of the broader Huang–Huai–Hai wheat region, thereby providing a favorable context for evaluating framework performance49. While the core feature system and ensemble architecture are designed for general use, application to other regions would require localized parameter optimization, including the collection of region-specific sample points. This remains an important direction for future research. Conclusions Multi-temporal Sentinel-2 remote sensing images from 2021 to 2023 were utilized to extract multiple features for dataset construction. Individual machine learning approaches were combined with a stacked integration model to identify and analyze the spatio-temporal dynamics of winter wheat planting areas in Jiangsu Province. The main conclusions are summarized in three key points. First, regarding feature combinations, the fusion scheme of spectral band reflectance features, spectral indices features, and texture features yields the best performance, achieving an OA of 93.10% and a Kappa coefficient of 0.8724, surpassing that of any single feature set. The synergistic integration of multidimensional features compensates for the limitations of individual features in characterizing information, thereby significantly enhancing classification accuracy. Second, the comparative analysis of different stacking models indicates that the combination of RF, SVM, and GTB achieves the highest accuracy, with an OA of 94.74% and a Kappa coefficient of 0.9283. In contrast, the multi-model stacking architecture that integrates RF, SVM, CART, and GTB yields lower classification performance, indicating that an excess of models may introduce redundant information interference and diminish overall recognition effectiveness. Finally, based on the comprehensive feature dataset, the spatial distribution of winter wheat from 2021 to 2023, generated using the optimal model, demonstrates stable classification accuracy. The remote sensing recognition results closely align with the official statistics, verifying the robustness and applicability of the proposed method. In summary, a large-scale agricultural remote sensing monitoring framework with strong robustness was developed, enabling high-accuracy classification of winter wheat in Jiangsu Province. The resulting products provide effective spatial data support for regional agricultural policymaking, dynamic cultivated land resource management, and food security monitoring. Future research may further integrate multi-source features and optimize machine learning approaches to explore their performance potential in time-series remote sensing classification, while extending the framework to broader geographic regions to enhance its generalization capability and practical applicability. Author contributions Conceptualization, Tao. W. and Shao. Y.; data collection and processing, Shao. Y., Ren. S., Lai. G.; data analysis, Tao. W., Su. W, Yan. S; writing—original draft preparation, Tao. W. and Shao. Y.; writing—review and editing, Tao. W., Su. W, Yan. S; visualization, Shao. Y., Ren. S., Lai. G.. All authors have read and agreed to the published version of the manuscript. Funding This study was supported by the Intergovernmental International Scientific and Technological Innovation Cooperation Project of National Key R&D Program (No. 2025YFE0102000), the Jiangsu Province Industry-Academia-Research Collaboration Program (No.BY20251504), the Suqian Sci&Tech Program (No.K202537), the Agricultural Science and Technology Innovation Project of Shandong Academy of Agricultural Sciences (No. CXGC2025G03), as well as the Talent Introduction Research Start-up Fund Project of Suqian University (No. XG2024XRC015). Data availability All Sentinel-2 datasets and the municipal winter wheat cultivation dataset used in this study are publicly available. Sentinel-2 data were sourced from Google Earth Engine (https://earth-engine.google.com/). The municipal winter wheat cultivation dataset is available at https://www.shujuku.org/china-county-statistical-yearbook.html. The sample point data collected for this study are available from the corresponding author upon reasonable request, as they are currently being used for ongoing research. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All Sentinel-2 datasets and the municipal winter wheat cultivation dataset used in this study are publicly available. Sentinel-2 data were sourced from Google Earth Engine (https://earth-engine.google.com/). The municipal winter wheat cultivation dataset is available at https://www.shujuku.org/china-county-statistical-yearbook.html. The sample point data collected for this study are available from the corresponding author upon reasonable request, as they are currently being used for ongoing research.