Introduction

Accurate spatial distribution and planting information of crops constitutes the fundamental basemap for monitoring regional crop growth and predicting productivity1. Currently, factors like fragmented fields and complex planting methods contribute to low accuracy in existing crop classification and mapping methods, resulting in a scarcity of large-scale, high-precision crop classification products2. Additionally, the majority of existing crop classification products are released with significant delays, making it difficult to meet the demand for precise management of crop production in the current season3.

The development of satellite remote sensing technology provides a viable solution for the rapid and accurate extraction of large-scale crop planting information3. Common methods for crop mapping using remote sensing technology include: techniques that rely on a small amount of ground truth data for visual interpretation of high-resolution images4,5, techniques that depend on prior knowledge extracted from existing crop type maps6,7, and techniques that involve training data or model transfer8,9. The first method primarily relies on ground truth data to classify single-period remote sensing images, but it is susceptible to phenomena such as “same material, different spectra” and “different material, same spectra”, leading to mixed classification results and uncertain accuracy. The second method utilizes existing crop type map products with multiple crop planting structures or similar structures as training data to generate prior knowledge and then completes the mapping of crop planting structures in the study area. However, this method is influenced by the spatial distribution of existing crop planting structures. Additionally, since these methods typically rely on sampled data, the regional characteristics of samples can affect the accuracy of results due to the transfer between different regions or different times within the same region10. As mapping accuracy is easily affected by intra-class and inter-class changes in reflectance and phenology, training data or model transfer techniques still require original ground truth values to train supervised models and determine robust classification features, and they are susceptible to changes in land cover types6.

High-quality ground truth data collection requires significant time, manpower, and financial support. Mapping the wide-ranging and timely distribution of crop planting structures is constrained by the lack of representative training data and ground truth data11. Given these constraints, scientists are increasingly using crop phenological characteristics to map distribution structures, yielding promising results3,12. This approach takes into account the spectral differences of land cover at different growth stages to classify crops. Traditional extraction of crop phenological periods typically relies on experts recording field observations or using cameras to monitor phenological information in specific areas12. Previous studies have demonstrated that crop phenological characteristics are influenced by meteorological factors, soil properties, terrain features, etc., leading to significant spatiotemporal variations13. Time-series satellite remote sensing images can be used to record detailed phenological and spectral changes during the crop growth cycle, providing a novel method for large-scale monitoring of crop phenology11.

At present, there is a limited number of published studies on the distribution of winter wheat planting structures in China. The Global Change and Terrestrial Ecosystem Modeling Research Team at Sun Yat-sen University14 mapped winter wheat in China from 2016 to 2022 with a spatial resolution of 30 m by comparing the similarity between known winter wheat plots and unknown land cover NDVI time series curves. The Smart Agriculture Innovation Team at the College of Agriculture, Nanjing Agricultural University3, developed a classifier for winter wheat crops using ground truth data and a detailed analysis of the primary land cover types and their spectral variations in China. Utilizing machine learning and model transfer techniques, they generated China’s inaugural winter wheat planting distribution maps with a 10-m resolution for the period 2020–2022. This type of winter wheat extraction method requires extensive field survey data support. Meanwhile, resolutions of 30 m or coarser often suffer from serious mixed pixel problems due to the influence of plot size, leading to uncertainty in classification accuracy. Building upon prior research, this study aims to develop an automated mapping method for winter wheat that does not rely on ground truth data, has high model credibility, and strong applicability. It aims to automatically extract key phenological periods of winter wheat using long time-series remote sensing imagery and construct extraction models by analyzing the spectral specificity of winter wheat during these key phenological periods, thereby achieving automatic identification of winter wheat.

Materials and methods

Study areas

This study selected Yancheng City, Jiangsu Province, China, as the experimental site to validate the proposed method. Yancheng City is located in the central coastal area of Jiangsu Province. It is the largest prefecture-level city in terms of land area in the province. It lies within the agricultural ecological zone of the North China Plain and is a major agricultural city. Abundant agricultural resources and favorable climatic conditions facilitate crop growth. The principal crops are wheat, rice, and corn. Among them, wheat is one of the main food crops and also the main winter crop.

Data sources

Satellite data

The vegetation index can quantitatively provide information about crop growth and health, serving as a primary data source for the analysis of vegetation phenology, crop growth conditions, and monitoring of pests and diseases. Among numerous vegetation indices, the Normalized Difference Vegetation Index (NDVI) is the most frequently used and stable parameter15. Therefore, this study selected NDVI as the main parameter for extracting crop planting structure models.

This research selects the European Space Agency's Sentinel-2 optical remote sensing satellites as the primary remote sensing data source. The combination of Sentinel-2A and Sentinel-2B satellites provides multispectral data with the highest resolution of 10 m and a revisit period of 5 days, offering great potential for extracting agricultural phenological characteristics and crop identification3. Throughout the entire growth phenological period of winter wheat, a total of 386 optical images with cloud cover below 30% from October 2022 to June 2023 were extracted. Subsequently, high-quality time-series vegetation index datasets were constructed through processes including radiometric and atmospheric correction, cloud masking, spectral index calculation, daily scale median synthesis, 10-day scale maximum value synthesis, and filling of missing data regions.

Reference data

A land use change dataset in the research area was used as reference data to support the selection of different land cover samples in S–G filtering and the determination of farmland patches in the research area.

The MODIS Land Cover Dynamics Product MCD12Q2 dataset, along with historical agricultural meteorological decadal reports from Jiangsu Province, serves as a reference for the phenological history of winter wheat growth, was used to verify the extracted critical phenological growth time points of winter wheat. MCD12Q2 has a spatial resolution of 500 m and provides four phenological period nodes, including start of growing season (green-up period), length of growing season, peak of growing season (heading period), and end of growing season (maturity period)16.

The actual irrigated area of winter wheat as reported in the agricultural basic information of Yancheng City in 2023 was used to validate the accuracy of the model-extracted results. A prior knowledge database of winter wheat fixed planting areas within the cultivated land scope, extracted using the 30-m spatial resolution winter wheat mapping results for China from 2016 to 202214 and the 10-m resolution winter wheat planting distribution maps for China from 2020 to 20223, was used to verify the positional accuracy of winter wheat planting structure.

Methodology

Reconstruction of NDVI series data

Due to the influence of atmospheric conditions, sensor observation angles, underlying surface structures, and other factors, time-series NDVI data often contain residual noise, which limits the further utilization of the data and may lead to incorrect conclusions17. Consequently, it is necessary to denoise and reconstruct the time-series NDVI data before application. Among numerous time-series data reconstruction algorithms, the Savitzky–Golay (S–G) filtering algorithm has been recognized as particularly robust and effective18. The S–G filtering algorithm, initially proposed by Savitzky and Golay, utilizes local polynomial least squares fitting to achieve data smoothing and denoising through weighted averaging within a moving window. The calculation formula is as follows:

$$ NDVI_{i}^{*} = \frac{{\mathop \sum \nolimits_{j = - m}^{m} C_{i} \cdot NDVI_{{j + {\text{i}}}} }}{N} $$
(1)

where \(NDVI_{{j + {\text{i}}}}\) represents the original value of the i-th NDVI at time j; \(NDVI_{i}^{*}\) represents the filtered NDVI value; m denotes the half-width of the moving window; N represents the window size, \(N = 2{*}m + 1\); \(C_{i}\) stands for the filtering coefficient of the i-th NDVI value.

The S–G filtering involves two important parameters, including the half-width of the moving window (m) and the polynomial coefficients (d). Typically, setting m too small may “overfit” data points, making it difficult to capture long-term trends, whereas a larger m may ignore some important changes in the NDVI time series. Meanwhile, lower values of d may yield a smoother effect yet introduce a filtering bias, while higher values can mitigate this bias but risk overfitting19. Chen et al.19 suggested that a suitable range for m is 2–7, and for d is 2–4. This paper employed a fitting effect index19 to objectively determine the optimal combination of m and d, and thus complete the reconstruction of long-term NDVI datasets. The formula for calculating the fitting effect index is as follows:

$$ F_{k} = \mathop \sum \limits_{i = 1}^{n} \left( {\left| {NDVI_{i}^{k + 1} - NDVI_{i}^{0} } \right|{*}W_{i} } \right) $$
(2)

where \(F_{k}\) represents the fitting effect index for a kth fitting time with different m and d combination. The combination (m, d) corresponding to the minimum \(F_{k}\) value is taken as the final S–G filtering parameter. n is the total amount of data in the dataset. \(NDVI_{i}^{k + 1}\) is the NDVI value corresponding to the ith data point of the kth fitted time series, and \(NDVI_{i}^{0}\) is the original NDVI value corresponding to the ith data point. \(W_{i}\) is the weight coefficient of the ith data point, calculated by the following formula:

$$ W_{i} = \left\langle {\begin{array}{*{20}l} 1 \hfill & {\quad NDVI_{i}^{0} \ge NDVI_{i}^{k + 1} } \hfill \\ {1 - {{d_{i} } \mathord{\left/ {\vphantom {{d_{i} } {d_{max} }}} \right. \kern-0pt} {d_{max} }}} \hfill & {\quad NDVI_{i}^{0} < NDVI_{i}^{k + 1} } \hfill \\ \end{array} } \right. $$
(3)

where \(d_{i} = \left| {NDVI_{i}^{0} - NDVI_{i}^{k + 1} } \right|\) represents the absolute difference of \(NDVI_{i}^{0}\) and \(NDVI_{i}^{k + 1}\); \(d_{max}\) is the maximum value of \(d_{i}\).

Phenological periods extraction

The main rule for extracting crop phenology using remote sensing monitoring technology is based on the morphological characteristics of the crop growth cycle reflected by time-series vegetation indices, and determined the key extreme points of curve changes through specific methods. Common methods include moving average method, threshold method, and derivative method20. The relative threshold method, a method that is widely utilized for phenological assessment, mainly operates at the pixel scale. It determines the key phenological stages of vegetation growth by setting dynamic thresholds for time-series vegetation index data, thereby mitigating the subjectivity involved in threshold selection20. In this study, based on the reconstruction of vegetation index time-series data using S–G filtering, the dynamic threshold method was adopted to extract key phenological periods of winter wheat growth at the pixel scale. The calculation formula is as follows21:

$$ NDVI_{lim} = \left( {NDVI_{max} - NDVI_{min} } \right)/{\text{C}} $$
(4)

where \(NDVI_{lim} ,\;NDVI_{max} ,\;NDVI_{min}\) are the dynamic threshold, maximum NDVI, and minimum NDVI within the crop growth cycle, respectively; C is the coefficient, generally set to 20%21.

Figure 1 depicted the time series dynamics of NDVI in winter wheat during the growing season. The time node at which the maximum NDVI appears was defined as the heading stage of winter wheat, while the time nodes at which NDVI reaches \(NDVI_{lim}\) during the rising and falling processes respectively represented the greening and ripening stages of winter wheat. The time node at which the minimum NDVI appears during the rising phase of the NDVI curve represented the seeding stage of winter wheat, while the time node at which the minimum NDVI appears during the falling phase of the NDVI curve represented the harvesting stage of winter wheat.

Figure 1
figure 1

The time series dynamics of NDVI in winter wheat during the growing season.

Full size image

Winter wheat planting structure extraction

Within a shorter growth cycle, the growth status of winter wheat and corresponding NDVI values exhibit unique temporal and spatial patterns. During the seeding period (Fig. 2a), winter wheat is either absent or just beginning to grow, characterized by exposed soil and scant vegetation cover. The NDVI typically reflects the optical properties of the soil, and remote sensing monitoring shows the lowest NDVI values during the winter wheat growth cycle. During the heading stage (Fig. 2b), winter wheat growth reaches its peak, with remote sensing monitoring showing high NDVI values and vegetation greenness maintained at a high level. During the harvesting period (Fig. 2c), the characteristics of remote sensing monitoring NDVI mainly reflect the reduction in vegetation coverage due to aging, shedding, and harvesting activities in winter wheat fields, with NDVI values typically significantly lower than those of surrounding vegetation. Therefore, the characteristics of NDVI during these three key phenological periods can be used to extract the planting structure of winter wheat.

Figure 2
figure 2

Growth stage diagram of winter wheat in key phenological phases.

Full size image

In large-scale agricultural planting systems, the phenomenon of homogeneity in different varieties of the same crop and heterogeneity in varieties of different crops can lead to misidentification of winter wheat. Hierarchical classification models can significantly reduce the mutual influence between different land cover types3. Therefore, before extracting the structure of winter wheat planting, a hierarchical classification model was constructed using the maximum NDVI values during the winter wheat growing season to classify all land cover types into winter crops and non-winter crops. The model is as follows:

$$ NDVI_{max} > threshold\_1 $$
(5)

where \(threshold\_1\) represents the model threshold. Pixels with NDVI maximum values greater than this threshold were classified as winter crops, otherwise were classified as non-winter crops. Zhang et al.22 suggested that a relatively loose threshold of 0.5 can be used to construct masks for winter crops and eliminate non-cultivated lands. Therefore, we set \(threshold\_1{ }\) to 0.5 in our research.

Building upon the differentiation between winter and non-winter crops, this paper utilized NDVI values during three key phenological stages of winter wheat, including seeding, heading, and harvesting, to construct a differential model for extracting the planting structure of winter wheat. The extraction model is as follows:

$$ \left\{ {\begin{array}{*{20}l} {{\text{NDVI}}_{max} - {\text{NDVI}}_{1} > threshold\_2} \hfill \\ {{\text{NDVI}}_{max} - {\text{NDVI}}_{2} > threshold\_3} \hfill \\ \end{array} } \right. $$
(6)

where \(NDVI_{max} ,\;NDVI_{1} ,\;NDVI_{2}\) respectively represent the NDVI values of heading stage, seeding stage, and harvesting stage; \(threshold\_2,\;threshold\_3\) are the corresponding dynamic thresholds.

Validation samples determination based on surface prior knowledge

Authenticity verification of remote sensing product mainly relies on ground truth data to evaluate the uncertainty and provide precise evaluation indicators23. In cases where ground truth data is lacking, verifying the accuracy of remote sensing products becomes challenging, leading to significant uncertainty regarding their capability to accurately depict the Earth’s surface conditions24. Prior knowledge can be used to describe surface information under normal conditions. With the development of remote sensing technology, prior knowledge has been widely applied in remote sensing model construction and quantitative parameter inversion25,26. Similarly, prior knowledge can be used for the authenticity verification of remote sensing products. This paper utilized the winter wheat planting pattern distribution results from prior research as prior knowledge, and identified pure winter wheat pixels within the cultivated area as reference data to evaluate the spatial accuracy of the extracted results in this paper. The schematic diagram was shown in Fig. 3, and the specific steps were as follows:

Figure 3
figure 3

Schematic diagram of the validation sample determination.

Full size image

Step 1 Extract the overlapping regions of crop planting areas determined by prior knowledge from different sources within the cultivated area;

Step 2 Ascertain the center points of the aforementioned overlapping crop planting areas;

Step 3 Use the central point as the focal point of the sample, and delineate an n*n (where n represents the pixel matrix dimensions) area based on the pixel size of the to-be-verified results as the base pixel dimension. If this area falls completely within the overlapping crop planting areas identified in Step 1, it qualifies as a verification sample; if not, it is to be excluded.

Step 4 Based on the remote sensing imagery of the winter wheat study period, visually interpret the determined verification samples, remove samples with surface type changes, and form the final prior sample dataset.

Ethical statement

The plant collection and use were in accordance with all the relevant guidelines.

Results and analysis

Reconstruction of NDVI series data

Utilizing the land use change dataset as a reference, random sample points within the study area were selected in proportion to the various land use types to determine the optimal (m, d) combination for fitting. Table 1 displayed the fitting effect indices calculated under different combinations of m and d. It was evident that with a constant polynomial degree d, the fitting effect index initially raised with an increase in the half-width of the filtering window (m), only to subsequently experience a rapid decline. Conversely, when m remains constant, an increase in d leaded to a decrease in the fitting effect index, followed by a gradual enhancement in the fitting quality. The experimental results indicated that a smaller m coupled with a larger d markedly enhances the fitting performance. The (m, d) pairing of (3, 4) emerged as the optimal combination for the S–G filter throughout the fitting iteration sequence.

Table 1 Fitting-effect index within different combination of m and d.
Full size table

To verify the accuracy of the fitting parameters, three different land surface substrates, including vegetation, hardened surfaces, and water bodies, were selected, and the Pearson correlation coefficient and standard deviation before and after filtering were employed to evaluate the filtering parameters and filtering effects of the S–G filter. Table 2 illustrated that, across the three typical land surface substrates, the (3, 4) combination of (m, d) achieved the maximum Pearson correlation coefficient and minimum standard deviation before and after filtering, indicating that this combination can maximize the filtering effect while preserving the original vegetation growth characteristics to the greatest extent.

Table 2 Comparison of S–G fitting effects within different combination of m and d upon various land covers.
Full size table

Winter wheat phenological periods extraction

Based on the reconstruction of NDVI time-series dataset using S–G filtering, the extraction of crop phenological stages in field plots were conducted using the difference model and dynamic threshold method. Figure 4 showed the three key phenological nodes of crop growth, namely seeding, heading, and harvesting periods, in different cultivated land plots. The phenological distribution map revealed that the seeding period of crops in cultivated land (Fig. 4a) mainly concentrates in the middle to late October, with a small number of plots distributed in the early to middle November; the heading period (Fig. 4b) mainly concentrates in the middle to late April; and the harvesting period (Fig. 4c) mainly occurs in the middle to late June. The extraction results were consistent with the historical statistical data on the phenology of winter wheat growth in Jiangsu Province.

Figure 4
figure 4

Spatial patterns of winter wheat in various key phenological phase. (Note: The number in phenological phases represents different time note, such as, “101” represents early October, “102” represents mid-October, and “103” represents late October). Model output; maps generated using ArcGIS 10.7.1 (https://www.esri.com).

Full size image

To verify the accuracy of phenological extraction, ten plots were systematically selected within the study area, and the collected phenological information of crops were used as a reference to evaluate the extracted results. The data sourced from historical agricultural meteorological decade reports of Jiangsu Province were used as the phenological reference data for the validation of seeding period, while the reference data for heading and harvesting periods were from MODIS Land Cover Dynamics Product MCD12Q2 data. The comparative results in Table 3 showed that the distribution of key phenological stages of winter wheat growth extracted in this study were in good agreement with the reference data, which can accurately reflect the phenological changes in winter wheat growth in the study area.

Table 3 Accuracy of phenological phase with reference data.
Full size table

Winter wheat planting structure extraction

Taking the phenological period into account, using the extraction method for winter wheat planting structures proposed in this study, we identified winter wheat based on time-series NDVI features, as depicted in Fig. 5. Table 4 presented the statistical accuracy evaluation of the winter wheat planting structure extraction, utilizing the 2023 Statistical Yearbook data for Yancheng City. The accuracy of the extracted winter wheat planting areas in each county exceeded 91.00%, with an overall accuracy rate of 94.92%, indicated that the proposed extracting algorithm for winter wheat planting structures is highly accurate.

Figure 5
figure 5

Planting distribution map of winter wheat. Model output; maps generated using ArcGIS 10.7.1 (https://www.esri.com).

Full size image
Table 4 Accuracy assessment based on statistical yearbook of Yancheng, Jiangsu.
Full size table

To further validate the spatial details of the generated winter wheat planting distribution, this study used the aforementioned sample determination method based on surface prior knowledge to determine the winter wheat distribution samples in the study area (Fig. 6), and then conducted a spatial distribution accuracy evaluation of the extracted winter wheat results. As shown in Table 5, the positional accuracy of the winter wheat planting structure varied between 89.09 and 98.12% among different counties, with an overall accuracy of 93.26% and Kappa coefficients ranging from 0.8792 to 0.9809, averaging at 0.9228. This indicated that the winter wheat sample classification of the winter wheat planting structure extraction algorithm proposed in this study has high accuracy and consistency.

Figure 6
figure 6

Schematic diagram of the validation sample distribution defined by prior knowledge. Model output; maps generated using ArcGIS 10.7.1 (https://www.esri.com).

Full size image
Table 5 Accuracy assessment of winter wheat spatial distribution in Yancheng, Jiangsu.
Full size table

Discussion

Winter wheat is one of the principal grain crops in China. Timely and accurate mastery of its refined spatial distribution holds significant importance for guiding agricultural production and ensuring food security. In our research, winter wheat extracting using remote sensing technology involves multiple steps, including the selecting and processing of remote sensing data sources, the identification of key temporal images during crop growth, and the construction of crop planting structure extraction models27.

Crop phenology refers to the natural rhythm phenomenon formed by the influence of environmental factors and genetic factors, which plays an important role in guiding and managing agricultural production, ecosystem protection, and restoration. Traditional crop phenology information mainly relies on field observations. However, this method cannot achieve large-scale phenological observations, and the results are easily affected by subjective human factors20. Remote sensing technology can achieve continuous observation of large-scale crop phenological periods through two steps: the reconstruction of remote sensing time series data and the extraction of phenological parameter20. This study first used S–G filtering technology to filter and reconstruct NDVI time series data, and then used the vegetation index threshold method to extract the phenological periods of winter wheat. During the data filtering and reconstruction process, the optimal S–G filtering parameter combination was automatically derived using the filtering effect index, thereby addressing the issue of subjective bias in the selection of window size and polynomial coefficients. By evaluating the filtering effect on various land surface underlays, it was ascertained that this method can identify the most robust filtering parameter combination. Based on the filtered and reconstructed NDVI time series data, the threshold method was employed to extract the key phenological periods of crops. In addition, the phenological nodes commonly used in the extraction of winter wheat area are the greening stage, heading stage, and maturity stage2,3,28. This study selected three phenological stages of winter wheat planting structure, including seeding period, heading stage, and harvesting stage. NDVI time-series curves indicated that the selected three phenological stages can discern winter wheat from other winter crops. Verification using traditional ground observation data and other remote sensing phenology products showed that the extracted phenological periods in this study has a high accuracy and can truly reflect the key phenological nodes of winter wheat growth.

With the development of computer and remote sensing technology, extraction of crop planting areas using time series remote sensing images combined with phenological characteristics of crop growth has been more widely applied and promoted2,3,7,14. Time series data can provide continuous and dynamic observation of crop growth, capturing subtle changes in crop growth processes. Combined with crop phenological information, it can be used to division of crop growth stages accurately, and then to establish perfect crop information extraction model. Various crop information can be extracted through appropriate threshold settings, thereby providing more data references for agricultural management and decision-making. However, the absence of concurrent ground sample data presents a significant challenge to the evaluation of model accuracy23. Previous studies have demonstrated that high-precision ground-truth data can be used to characterize the spatial distribution of vegetation growth in the study area29,30,31. In this research, based on existing high-precision winter wheat planting area distribution data covering the study area, regions with consistent winter wheat planting structures over multiple years were selected. A certain sample generation strategy was used to automatically generate a prior sample library. Then, samples were screened through visual interpretation by utilizing high-resolution remote sensing imagery from the study period to assess the accuracy of the spatial distribution of the extracted winter wheat planting areas. Both statistical evaluation of planting area and evaluation of spatial distribution of planting structure based on prior sample libraries suggested that the proposed winter wheat planting structure extracting model in this study, has a high accuracy.

Some potential uncertainties could affect the accuracy of winter wheat planting structure extraction. Firstly, the quantity of cloud-free satellite data substantially influences the identification quality14. To minimize the impact of clouds, we utilized all available data to construct the high-quality time-series vegetation index datasets based on daily scale median synthesis, 10-day scale maximum value synthesis, and method for filling of missing data regions. However, obtaining cloud-free time-series optical remote sensing images for small areas is relatively straightforward, but it remains challenging for large areas and in cloudy and rainy regions22. To solve those problems caused by cloud contamination, multi-source satellite data can be used to generate high-quality images for long-term monitoring. Secondly, empirical thresholds are generally determined through experience and practical application, carrying a certain degree of subjectivity and lacking rigorous theoretical proof21. In our research, in the process of phenological periods extraction and the process of constructing masks for winter crops and non-winter crops, we used empirical thresholds previously published by other researchers. The uncertainty surrounding these selected empirical thresholds may affect the credibility of our findings. In future research, we will introduce machine learning methods32 to optimize these empirical values, thereby improving the scientific and reliability of decision making. Thirdly, we innovatively propose a strategy for construct a prior sample library using published historical data to validate the results. This approach can compensate for the challenges brought by the lack of ground truth data. However, the uncertainty of historical data and changes in land cover types can lead to the unreliability of validation results23. In our future work, we will verify the accuracy of historical results using ground validation data to ensure the reliability and availability of data products from previous generations.

Conclusions

This study proposed a method for extracting winter wheat planting structure by considering crop phenology, involving three main steps: filtering and reconstructing NDVI time series data, identifying key growth phenological periods of winter wheat, and developing a model for extracting winter wheat planting structure. The method proposed in this study had the following advantages:

  1. 1.

    The optimal S–G filtering parameter combination was determined automatically using the fitting effect index to achieve filtering and reconstruction of NDVI time series data, avoiding the subjective bias of introduced by manual selection and enhancing the filtering efficacy.

  2. 2.

    Winter wheat growth phenological phases were identified automatically using a threshold method, avoiding the uncertainty in product accuracy caused by regional variations in ground-based observation.

  3. 3.

    A scheme for constructing a surface prior knowledge library based on prior knowledge was proposed to verify remote sensing products, solving the problem of lacking ground-based verifying data.

  4. 4.

    A winter wheat extracting model was developed based on the key phenological phases and hierarchical classification principles for extracting winter wheat plant structure automatically. Model thresholds was determined through a dynamic thresholding technique, thereby reducing the uncertainties associated with manual threshold setting in the model’s extraction results.

In summary, the winter wheat planting structure extraction model proposed in this study fully considered the phenological specificity of winter wheat growth, significantly improving the accuracy of extracting result. Furthermore, this method does not require the involvement of training samples, does not rely on manual experience, and exhibits strong model transferability. Future work will involve employing recent ground-truth data to validate the proposed method.