A large-scale early winter wheat recognition method based on automatic extraction of training samples

Abstract

The present invention proposes a large-scale winter wheat early recognition method based on automatic extraction of training samples, and the steps of the method are as follows: obtain Sentinel‑1 and Sentinel‑2 satellite images in the wheat growing season through a remote sensing cloud platform; generate spatiotemporally continuous Sentinel‑1 and Sentinel‑2 time series; construct a winter crop index WCI; extract winter crop pixels; distinguish between winter wheat and winter rapeseed; construct a winter wheat classification model in combination with the Sentinel‑1 / 2 images of the current year; obtain the winter wheat classification product of the current year based on the classification model; migrate the training model to the target year, and clarify the earliest identifiable period of winter wheat by testing the influence of different time window phase characteristics on classification accuracy. The method can extract winter wheat in a timely and accurate manner, and has great application potential in winter wheat planting area monitoring, yield prediction, food security assessment, etc.

Description

Technical Field

[0001] The invention belongs to the field of agricultural remote sensing monitoring and is a large-scale winter wheat early recognition method based on automatic extraction of training samples. Background Art

[0002] Crop type remote sensing classification products can provide accurate crop spatial distribution and planting area information. They are the basic base map for regional-scale crop growth monitoring and productivity prediction, and are crucial for food security risk assessment. Due to factors such as fragmented fields and complex planting methods, these methods generally have problems such as low accuracy, low efficiency, and poor universality, and large-scale, high-precision crop classification products are still scarce. Existing crop classification products based on remote sensing technology mostly use machine learning methods, which rely heavily on ground truth data to train machine learning classifiers and build crop classification models. Generally speaking, the acquisition of ground truth data requires field point-by-point surveys, which are time-, manpower-, and material-costly and inefficient. Therefore, developing a set of automated training data extraction methods is crucial for the automated production of large-scale crop classification products. In addition, the public release time of most existing crop classification products is seriously delayed, which makes it difficult to meet the needs of accurate management of crop production in the current season. How to perform early high-precision identification of crop classification products based on limited images in a limited time is also an urgent problem to be solved. Summary of the invention

[0003] The technical problem solved by the present invention is to provide a large-scale winter wheat early identification method based on automatic extraction of training samples. Through the time-series optical Sentinel-2 data, the spectral differences between winter crops and other ground objects are clarified, the winter crop index is constructed to enhance the winter crop signal, and the winter crop pixels are automatically extracted using the Otsu threshold segmentation algorithm. Based on the polarization characteristics of Sentinel-1VH, winter wheat and winter rapeseed are further distinguished, so as to extract a large number of high-quality winter wheat training samples. Finally, based on the machine learning method, the automated production of Chinese winter wheat distribution products with a spatial resolution of 10 meters is realized, and the classification accuracy and the earliest identifiable period of winter wheat in each period are clarified based on the model migration method. This method overcomes the dependence of traditional classification methods on real surface data, has high accuracy, efficiency, and strong robustness, and can be widely used in the rapid production and release of large-scale, high-precision, and high-spatial-resolution winter wheat spatial distribution products.

[0004] The technical solution to achieve the purpose of the present invention is:

[0005] A large-scale winter wheat early identification method based on automatic extraction of training samples comprises the following steps:

[0006] Step 1: Obtain Sentinel-1 and Sentinel-2 satellite images during the winter wheat growing season in China’s major winter wheat producing provinces, pre-process the images on the remote sensing cloud platform, and generate spatiotemporally continuous Sentinel-1 and Sentinel-2 monthly time series.

[0007] Step 2: Construct the Winter Crop Index (WCI) based on the Sentinel-2 time series and winter crop phenology information, and extract winter crop pixels, including:

[0008] Step 2-1: Conduct ground surveys and literature searches to obtain information on land cover types and crop growth processes in the study area;

[0009] Step 2-2: Analyze the phenological differences between winter crops and non-winter crops based on the Sentinel-2 NDVI time series of various land cover types;

[0010] Step 2-3: Based on the phenological differences between winter crops and non-winter crops, winter crops are divided into the early growth stage (EGS) and the late growth stage (LGS), where the early growth stage is from the sowing period to the heading period, and the late growth stage is from the heading period to the harvest period. Further, the sowing period, heading period and harvest period of winter crops are retrieved from the NDVI time series, and the NDVI of the corresponding growth period is extracted, which are recorded as NDVI min1 , NDVI max and NDVI min2 , and then construct the winter crop index WCI based on these indicators; in step 2-3, the specific formula of the winter crop index WCI is: WCI = ΔNDVI EGS ×ΔNDVI LGS =(NDVI max -NDVI min1 )×(NDVI max -NDVI min2 )

[0011] Among them, ΔNDVI EGS and ΔNDVI LGS are the increase and decrease of NDVI in the early and late growth stages of winter crops, respectively. min1 , NDVI max and NDVI min2 Represent the NDVI values ​​of winter crops at the sowing, heading and harvesting stages respectively.

[0012] Step 2-4: Divide the study area into 1°×1° grids, and based on the Otsu threshold segmentation algorithm, perform adaptive threshold segmentation on the WCI grid by grid to automatically extract winter crop pixels. In Step 2-4, the study area is divided into 1°×1° grids. Based on the Otsu adaptive threshold segmentation algorithm, the optimal segmentation threshold OT is found, and the WCI is segmented grid by grid based on OT, so as to automatically extract winter crop pixels. The specific formula for finding OT is:

[0013] σ 2 (t) = P0(μ0 - μ T ) 2 + P1(μ1 - μ T ) 2

[0014]

[0015] where σ 2 (t) is the between-class variance generated by dividing winter crops and other land covers based on the threshold t, P is the probability that undefined pixels belong to winter crops or other land covers, and μ0, μ1, and μ T are the means of winter crops, other land covers, and all land covers, respectively.

[0016] Step 2-5: Further combine the cultivated land layer of the land cover product (FROM_GLC10) to mask the extracted winter crop pixels and eliminate the errors existing in the extracted winter crop pixels.

[0017] Step 3: Based on the differences in spectral indices and backscattering of different winter crops in the Sentinel-1 and Sentinel-2 time series, distinguish winter wheat and winter rapeseed in winter crops, specifically including:

[0018] Step 3-1: According to the ground survey data, extract the spectral indices and backscattering coefficients of winter wheat and winter rapeseed based on the Sentinel-1 and Sentinel-2 time series;

[0019] Step 3-2: Analyze the differences in each spectral index and polarization band of winter wheat and winter rapeseed, and determine the sensitive bands and phenological periods, denoted as SP t ;

[0020] Step 3-3: Distinguish winter wheat and winter rapeseed according to the SP t threshold. If SP t < T, it is considered to be a winter wheat pixel in winter crops, otherwise it is a winter rapeseed pixel.

[0021] Step 4: Winter rapeseed and non-winter crop pixels are merged to define non-winter wheat pixels, and then stratified sampling is performed on winter wheat and non-winter wheat pixels grid by grid, and 1000 pixels are randomly selected from each grid as candidate training samples. Finally, based on the spatial neighborhood analysis, samples with insufficient representation in the candidate samples are removed, and the remaining samples are the final training samples.

[0022] Step 5: Input the training samples obtained in step 4 and the individual Sentinel-1, Sentinel-2, and the fused time series features of Sentinel-1 and Sentinel-2 into the Random Forest (RF) classifier, build the RF classification model, and classify winter wheat in 2020 and 2021, so as to verify the reliability of the automatic extraction of training samples and the quality of the training samples, and clarify the contribution of optical and radar features to winter wheat classification. Combined with the automatically generated training data, the optimal features and the RF classifier, the 10-meter spatial resolution winter wheat classification products for the study area in 2020 and 2021 were generated.

[0023] Step 6: Migrate the RF classification model under the optimal feature combination in step 5 to the target year (2021 and 2022). First, generate the winter wheat classification product with a spatial resolution of 10 meters for the target year based on the same features, and evaluate the reliability of the model migration method according to the accuracy of the winter wheat classification product. Then, further combine the image features of different periods of the target year to classify winter wheat, evaluate the classification accuracy of winter wheat in different periods, and finally determine the earliest identifiable period of winter wheat.

[0024] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0025] 1. The large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention integrates phenological characteristics and multi-source remote sensing data to extract training samples. The method is highly efficient and has strong universality;

[0026] 2. The large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention adopts automatic extraction of training data and machine learning methods on the remote sensing cloud platform to produce large-scale high spatial resolution winter wheat spatial distribution products. The method is simple to operate, efficient and effective.

[0027] 3. The large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention produces the spatial distribution products of winter wheat in the target year based on the historical classification model and model migration. Based on the accuracy of winter wheat classification products in different periods, it is clarified that the earliest identifiable period of winter wheat is the heading period. This method has high reliability, strong timeliness and is easy to promote. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the technical roadmap of the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention;

[0029] Figure 2 It is a schematic diagram of winter crop index construction of a large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention;

[0030] Figure 3 It is a graph showing the change of different optical and radar time series characteristics of the large-scale winter wheat early recognition method based on automatic extraction of training samples of the present invention, wherein (A), (B), (C) and (D) are NDVI, NDYI, VV and VH characteristics respectively;

[0031] Figure 4 3. It is a comparison chart of winter wheat classification accuracy in various provinces in 2020 and 2021 based on different classification features and classifiers of the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention, wherein (A) & (E), (B) & (F), (C) & (G) and (D) & (H) represent OA, F1, UA and PA, respectively, and green and orange represent the winter wheat classification accuracy based on Sentinel-2 features and RF classifier and based on Sentinel-1 and Sentinel-2 fusion features and RF classifier, respectively;

[0032] Figure 5 This is a comparison chart of winter wheat classification accuracy in various provinces in 2021 and 2022 based on the RF model migration of the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention, wherein (A) & (E), (B) & (F), (C) & (G) and (D) & (H) represent OA, F1, UA and PA, respectively, and green and orange represent the winter wheat classification accuracy based on Sentinel-2 features and RF classifier and based on Sentinel-1 and Sentinel-2 fusion features and RF classifier, respectively;

[0033] Figure 6 2020-2022 is a schematic diagram of the consistency verification results of the remote sensing estimated area of ​​winter wheat and agricultural statistical data in 2020-2022 based on the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention, wherein (A) and (B) are the consistency verification results of the remote sensing estimated area of ​​winter wheat and agricultural statistical data in 2020 and 2021 based on the automatically extracted samples at the provincial scale, (C) and (D) are the consistency verification results of the remote sensing estimated area of ​​winter wheat and agricultural statistical data in 2021 and 2022 based on the model migration at the provincial scale, (E) and (F), (G) and (H) are the corresponding consistency verification results at the municipal scale;

[0034] Figure 7 It is a comparison chart of the winter wheat classification results of the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention and other similar products, wherein the first column is the Sentinel-2 image, the second column is the result of the present invention (ChinaWheat10), and the third, fourth and fifth columns are the results of other similar products (TWDTW, PTDTW, ChinaCP) respectively;

[0035] Figure 8 This is a comparison chart of the classification accuracy of winter wheat at different periods under different features of the large-scale winter wheat early identification method based on automatic extraction of training samples of the present invention, wherein the blue broken line is the corresponding accuracy of the single-phase Sentinel-2 feature (S2), the green broken line is the corresponding accuracy of the single-phase Sentinel-2 and Sentinel-1 fusion feature (S2+S1), the red broken line is the corresponding accuracy of the multi-phase Sentinel-2 and Sentinel-1 fusion feature (S2+S1+T), and the light blue background frame represents the earliest identifiable period of winter wheat. DETAILED DESCRIPTION

[0036] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0037] The implementation of the present invention is based on Sentinel-1 and Sentinel-2 satellite images, and the image parameters and features used are shown in Table 1.

[0038] Table 1 Satellite image parameters and related characteristics used in the study

[0039]

[0040] This example was carried out in the main winter wheat producing provinces in China (Huanghuaihai: Jiangsu, Anhui, Hubei, Henan, Shandong, Hebei, Shanxi, Shaanxi; Northwest China: Gansu, Xinjiang; Sichuan Basin: Sichuan), covering an area of ​​about 4 million km 2 , the proportion of winter wheat planting accounts for more than 96% of the national winter wheat planting area. Moreover, the automatic extraction of training samples was repeated and verified in 2020 and 2021, and the early identification of winter wheat was repeated and verified in 2021 and 2022, which can effectively verify the advancement, universality and robustness of the present invention.

[0041] A large-scale winter wheat early identification method based on automatic extraction of training samples, such as Figure 1As shown in the figure, it mainly includes image preprocessing, surface real data collection, automatic extraction of training samples, winter wheat mapping and verification, and earliest identifiable period monitoring, which specifically includes the following steps:

[0042] Step 1: Acquire Sentinel-2 and Sentinel-1 satellite images during the winter wheat growing season in China's main winter wheat producing provinces. In 2020, 2021, and 2022, 65766, 65465, and 65305 Sentinel-2 images and 4955, 4826, and 4477 Sentinel-1 images were obtained, respectively. Preprocess the Sentinel-2 and Sentinel-1 images on the remote sensing cloud platform. The preprocessing steps of Sentinel-2 include cloud removal, vegetation index calculation, monthly median synthesis, and missing value filling, while the preprocessing steps of Sentinel-1 mainly include denoising, polarization index calculation, and monthly median synthesis. The Sentinel-2 cloud probability dataset can characterize the probability that each pixel in the Sentinel-2 image is contaminated by clouds. The present invention first removes clouds from the original Sentinel-2 surface reflectance image based on this dataset to obtain a cloud-free Sentinel-2 image. Then, the present invention calculates three commonly used vegetation indices, NDVI, EVI and LSWI, based on the original bands of Sentinel-2 to increase the sensitive classification characteristics of winter wheat. Furthermore, the present invention screens all Sentinel-2 images month by month, and fuses the Sentinel-2 images of each month using a median synthesis algorithm to obtain a monthly image. Finally, a linear interpolation method is used to fill in the missing pixels on the monthly image to obtain a spatiotemporally continuous Sentinel-2 monthly time series. For Sentinel-1, the present invention first performs denoising on each Sentinel-1 image based on the Refine Lee algorithm; then calculates the two polarization features of Ratio and Diff; finally, similar to the Sentinel-2 processing method, the median of the Sentinel-1 images of each month is synthesized into one period to generate the Sentinel-1 monthly time series.

[0043] Step 2: Construct the Winter Crop Index (WCI) based on the Sentinel-2 time series and winter crop phenology information, and extract winter crop pixels, including:

[0044] Step 2-1: Conduct ground surveys and literature searches to obtain information on land cover types and crop growth processes in the study area. In winter, the main types of land features in the study area include winter crops, forests (evergreen and deciduous), buildings, water bodies, and spring crops. Among them, winter crops mainly include winter wheat (~91%) and winter rapeseed (~8%), as well as a small amount of garlic, barley, etc. (<1%).

[0045] Step 2-2: Analyze the phenological differences between winter crops and non-winter crops based on the Sentinel-2 NDVI time series of various land cover types. For non-vegetated land features such as water bodies and buildings, their NDVI remains basically stable during the winter crop growing season, and the NDVI value is relatively low. Although the NDVI value of evergreen forest is higher, it also remains basically stable. The deciduous forest will show a trend of first decreasing and then increasing during the winter crop growing season. The NDVI of spring crops will show a significant upward trend at the end of the winter crop growing season, while the winter crops will show a trend of first increasing and then rapidly decreasing throughout the growing season.

[0046] Step 2-3: Based on the phenological differences between winter crops and non-winter crops, winter crops are divided into the early growth stage (EGS) and the late growth stage (LGS), where the early growth stage is defined as the period from sowing to heading, and the late growth stage is defined as the period from heading to harvest ( Figure 2 ).from Figure 2 It can be seen that the NDVI changes of winter crops in the early and late growth stages are significantly higher than those of other landforms.

[0047] The product of the NDVI variation range of EGS and LGS can effectively enhance the difference between winter crops and other landforms. The NDVI of winter crops during the sowing period, heading period and harvest period are at the local minimum or maximum, respectively. By retrieving the NDVI time series, the NDVI of the corresponding growth period can be extracted and recorded as NDVI min1 ,NDVI max and NDVI min2 , and then calculate the winter crop index WCI based on these indicators.

[0048] Step 2-4: Since the study area covers a large area with nearly 40 billion pixels, it is difficult to process this massive data set at one time even with cloud computing technology. Therefore, the present invention divides the study area into 1°×1° grids, and performs adaptive threshold segmentation on WCI grid by grid based on the Otsu threshold segmentation algorithm to automatically extract winter crop pixels.

[0049] Step 2-5: Since the reflectivity of some pixels may be abnormal due to other factors, they may be misidentified as winter crop pixels. The present invention further obtains the global 10-meter spatial resolution land cover product FROM_GLC10, extracts the cultivated land layer from the product, and then masks the extracted winter crop pixels based on the cultivated land layer to eliminate possible errors in the extracted winter crop pixels.

[0050] Step 3: Differentiate between winter wheat and winter rapeseed in winter crops based on the differences in spectral indices and backscattering of different winter crops in the Sentinel-1 and Sentinel-2 time series, including:

[0051] Step 3-1: Extract the spectral index and backscatter coefficient of winter wheat and winter rapeseed based on the Sentinel-1 and Sentinel-2 time series according to the ground survey data. The present invention extracts the most commonly used Normalized Difference Vegetation Index (NDVI) from the optical data. Figure 3 A), and the Normalized Yellow Vegetation Index (NDYI) which is sensitive to yellow. Figure 3 B), extracting all the polarization features VV( Figure 3 C) and VH( Figure 3 D).

[0052] Step 3-2: Based on t-test statistical analysis, the differences in spectral indexes and polarization bands between winter wheat and winter rapeseed were analyzed. It was found that the VH polarization difference between winter wheat and winter rapeseed was most significant in the late flowering period of rapeseed (April). Figure 3 This trend can also be seen in D.

[0053] Step 3-3: Distinguish winter wheat and winter rapeseed according to the VH polarization threshold. After analyzing 11734 and 11675 pure winter wheat and winter rapeseed pixels, it was found that VH = -16 can most accurately distinguish the two. VH < -16 is considered to be a winter wheat pixel in the winter crop, otherwise it is a winter rapeseed pixel.

[0054] Step 4: Winter rapeseed and non-winter crop pixels are merged and defined as non-winter wheat pixels. Then, winter wheat and non-winter wheat pixels are stratified and sampled grid by grid. 1000 pixels are randomly selected in each grid as candidate training samples. Finally, based on the spatial neighborhood analysis technology, it is further determined whether the candidate sample and its surrounding pixels belong to the same type. If they are the same, the sample is considered representative. Otherwise, it is deleted. Finally, high-quality and representative training samples are retained. Finally, 189,650 (winter wheat: 45,313, others: 144,337) and 190,192 (winter wheat: 46,148, others: 144,044) training samples were obtained in 2020 and 2021, respectively.

[0055] Step 5: Input the training samples obtained in step 4 and the individual Sentinel-1, Sentinel-2, and Sentinel-1 and Sentinel-2 fusion time series features into the Random Forest (RF) classifier, build the RF classification model, and classify winter wheat in 2020 and 2021, so as to verify the reliability of the automatic extraction method of training samples proposed in the present invention and the quality of training samples, and clarify the contribution of optical and radar features to winter wheat classification. Figure 4 It can be seen that the accuracy of the fusion of Sentinel-1 and Sentinel-2 data is significantly better than that based on Sentinel-2 data alone, whether in 2020 or 2021. Combining the automatically generated training data with the optimal features and RF classifier, the 10-meter spatial resolution winter wheat classification products for the study area in 2020 and 2021 were generated. The overall accuracy of the winter wheat classification products produced was above 94% ( Figure 4 A and E), and the correlation with statistical data at the provincial and municipal levels is above 0.96 and 0.92 respectively ( Figure 6 A, B, E and F), accurate wheat field recognition ( Figure 7 The second column) and significantly better than other similar products ( Figure 7 3rd, 4th and 5th columns).

[0056] Step 6: Migrate the RF classification model trained with the optimal features (Sentinel-2 and Sentinel-1 fusion features) in step 5 to the target years (2021 and 2022). First, generate winter wheat classification products with a 10-meter spatial resolution for the target years based on the same features. The reliability of the model migration method is evaluated based on the accuracy of the winter wheat classification products. The overall accuracy of the first set of 10-meter spatial resolution winter wheat spatial distribution products in China in 2021 and 2022 is above 93% ( Figure 5 A and E), and the correlation with statistical data at the provincial and municipal levels is above 0.95 and 0.91 respectively ( Figure 6 C, D, G and H). Then, winter wheat classification was further performed based on the image features of different periods in the target year, the classification accuracy of winter wheat at different periods was evaluated, and the earliest identifiable period of winter wheat was determined based on the classification accuracy saturation point. Figure 8 It shows that the accuracy of winter wheat is above 80% after the wintering period (February), above 85% in the jointing period (March), and the highest accuracy (>90%) can be achieved in the heading period (April). The earliest identifiable period of winter wheat is the heading period.

[0057] Step 7: Verify winter wheat classification accuracy based on different methods:

[0058] Method 1: In 2020-2022, a total of 1148 winter wheat samples and 581 non-winter wheat samples were obtained through ground surveys, and 1775 winter wheat samples and 3994 non-winter wheat samples were obtained through visual interpretation of high-resolution images. A confusion matrix was constructed using ground survey data and visual interpretation data. The overall accuracy generated based on the confusion matrix was

[0059] The position accuracy of winter wheat classification products was quantitatively evaluated by using Overall Accuracy, OA, F1-score (F1), User's Accuracy (UA) and Producer's Accuracy (PA) indicators.

[0060] Method 2: The present invention further obtains winter wheat planting area statistical data at the provincial and municipal scales from the national and provincial statistical bureaus, and compares the winter wheat area in the statistical data with the classification results based on the coefficient of determination (R 2 ) and relative root mean square error (RMSE) were used to quantitatively evaluate the area accuracy of winter wheat classification products.

[0061] Method 3: The present invention downloaded the winter wheat classification products of PTDTW, TWDTW and ChinaCP, and compared with these winter wheat classification products to qualitatively evaluate the recognition effect of the method proposed in the present invention on winter wheat from the classification details.

[0062] The above descriptions are only some embodiments of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principles of the present invention. These improvements should be regarded as within the protection scope of the present invention.

Claims

1. A large-scale winter wheat early identification method based on automatic extraction of training samples, characterized in that: The following steps are involved: Step 1: Obtain Sentinel-1 and Sentinel-2 satellite images during the winter wheat growing season in China's main winter wheat producing areas, pre-process the images on the remote sensing cloud platform, and generate spatiotemporally continuous Sentinel-1 and Sentinel-2 monthly time series; Step 2: Construct the winter crop index WCI based on the Sentinel-2 time series and winter crop growth and development characteristics, and extract winter crop pixels; Step 3: Distinguish winter wheat and winter rapeseed among winter crops based on the differences in spectral indices and backscattering of different winter crops in Sentinel-1 and Sentinel-2 time series; Step 4: Winter rapeseed and non-winter crop pixels are merged to define non-winter wheat pixels. Then, stratified sampling is performed on winter wheat and non-winter wheat pixels on a grid-by-grid basis. Multiple pixels are randomly selected from each grid as candidate training samples. Finally, based on spatial neighborhood analysis, samples with insufficient representation in the candidate samples are removed, and the remaining samples are determined as the final training samples. Step 5: Input the training samples obtained in step 4 and the individual Sentinel-1, Sentinel-2, and Sentinel-1 and Sentinel-2 fusion time series features into the random forest RF classifier to build an RF classification model and classify winter wheat in the historical period, thereby verifying the reliability of the automatic extraction of training samples and the quality of the training samples, and clarifying the contribution of optical and radar features to winter wheat classification; Combining the automatically generated training data with the optimal features and RF classifier, a 10-m spatial resolution winter wheat classification product was generated for the historical period of the study area; Step 6: Migrate the RF classification model under the optimal feature combination in step 5 to the target year to perform early identification of winter wheat in the target year: first generate a winter wheat classification product with a spatial resolution of 10 meters for the target year, then classify winter wheat based on the image features of different periods of the target year, evaluate the classification accuracy of winter wheat in different periods, and finally determine the earliest identifiable period of winter wheat.

2. A large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 1, characterized in that: Step 2 specifically includes: Step 2-1: Obtain information on land cover types and crop growth processes in the study area; Step 2-2: Analyze the phenological differences between winter crops and non-winter crops based on the Sentinel-2 Normalized Difference Vegetation Index (NDVI) time series of various land cover types; Step 2-3: Based on the phenological differences between winter crops and non-winter crops, winter crops are divided into the early growth stage EGS and the late growth stage LGS. The early growth stage is from the sowing period to the heading period, and the late growth stage is from the heading period to the harvest period. The sowing period, heading period and harvest period of winter crops are retrieved from the NDVI time series, and the NDVI of the corresponding growth period is extracted, which are recorded as NDVI min1 ,NDVI max and NDVI min2 , and then construct the winter crop index WCI based on these indicators; Step 2-4: Divide the study area into 1°×1° grids, perform adaptive threshold segmentation on WCI grid by grid based on the Otsu adaptive threshold segmentation algorithm, and automatically extract winter crop pixels; Step 2-5: Combine the cultivated land layer of the land cover product FROM_GLC10 to mask the extracted winter crop pixels and eliminate the errors in the extracted winter crop pixels.

3. The large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 2 is characterized in that: In step 2-3, the specific formula of winter crop index WCI is: WCI=ΔNDVI EGS ×ΔNDVI LGS =(NDVI max -NDVI min1 )×(NDVI max -NDVI min2 ) Among them, ΔNDVI EGS and ΔNDVI LGS are the increase and decrease of NDVI in the early and late growth stages of winter crops, respectively. min1 ,NDVI max and NDVI min2 Represent the NDVI values ​​of winter crops at the sowing, heading and harvesting stages respectively.

4. The large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 1 is characterized in that: In step 2-4, the study area is divided into 1°×1° grids, and the optimal segmentation threshold OT is found based on the Otsu adaptive threshold segmentation algorithm. The WCI is segmented grid by grid based on OT, so as to automatically extract winter crop pixels. The specific search formula for OT is: s 2 (t)=P0(μ0-μ T ) 2 +P1(μ1-μ T ) 2 Among them, σ 2 (t) is the inter-class variance generated by dividing winter crops and other objects based on threshold t, P is the probability that the undefined pixel belongs to winter crops or other objects, μ0, μ1 and μ T It is the average of winter crops, other features and all features.

5. The large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 1 is characterized in that: Step 3 specifically includes: Step 3-1: Extract the spectral index and backscattering coefficient of winter wheat and winter rapeseed based on the Sentinel-1 and Sentinel-2 time series according to the ground survey data; Step 3-2: Analyze the differences in spectral index and polarization bands between winter wheat and winter rapeseed, and determine the sensitive band and phenological period, recorded as SP t ; Step 3-3: According to SP t threshold to distinguish winter wheat and winter rape. If SP t < T, it is considered as the pixel of winter wheat in winter crops, otherwise it is the pixel of winter rape.

6. The large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 1 is characterized in that: In step 3, the Sentinel-1VH backscatter coefficient threshold is used to distinguish winter rapeseed from winter wheat. The specific steps are as follows: Firstly, the monthly NDVI, NDYI, VV and VH characteristic time series were extracted based on the ground survey data of winter rapeseed and winter wheat; Then, based on the t-test and time series characteristics, the most suitable period and characteristics for distinguishing winter rapeseed from winter wheat were analyzed, and finally the VH polarization in the late flowering period of rapeseed was selected for distinguishing winter wheat from winter rapeseed; Finally, the optimal threshold for distinguishing winter rapeseed and winter wheat was determined in combination with the pixel distribution frequency map to achieve accurate distinction between winter wheat and winter rapeseed pixels.

7. The large-scale winter wheat early identification method based on automatic extraction of training samples according to claim 1 is characterized in that: In step 4, a stratified random sampling method is used to generate training samples grid by grid, and then the spatial neighborhood analysis technique is further used to reduce the errors in the training samples and enhance the quality and representativeness of the training samples.

Images