Extraction method of soybean planting area based on domestic GF-6 WFV data

Abstract

The present invention relates to a soybean planting area extraction method based on domestic GF-6WFV data, including: obtaining GF-6WFV data of a to-be-extracted area, and preprocessing the data; using a decision tree hierarchical step-by-step extraction strategy to eliminate non-crop pixels in the preprocessed GF-6WFV data, and obtaining overall distribution data of farmland vegetation; generating a candidate remote sensing feature set; obtaining a preferred feature set, and then screening a classifier to obtain an optimal classifier; combining the preferred feature set with the optimal classifier to obtain an optimal extraction model, and evaluating the performance of the optimal extraction model, and examining the mapping effect of the optimal extraction model on the soybean planting area. The present invention utilizes a decision tree hierarchical step-by-step extraction strategy to significantly reduce the interference of other landform types such as water bodies, built-up areas, bare soil, and trees on soybean extraction results, which helps to obtain more accurate and reliable results; while reducing the storage space and amount of computation occupied by data, thereby improving execution efficiency.

Description

Technical Field

[0001] The invention relates to the technical field of crop remote sensing identification, in particular to a soybean planting area extraction method based on domestic GF-6 WFV data. Background Art

[0002] Remote sensing technology provides a feasible method for large-area soybean mapping. Different remote sensing data have been widely used to monitor global and regional soybean growth. Although great progress has been made in domestic and foreign research, large-area soybean mapping is still challenging in fragmented landscapes with complex crop planting conditions, such as the summer soybean planting area in the Huanghuaihai region of China. Many crops exist during the soybean production period, resulting in a complex planting structure in the Huanghuaihai region. Worldview-2, Worldview-3, Rapideye, Sentinel-2 and GF-6 satellites all contain repositioning band sensors. The spatial resolutions of Worldview-2, Worldview-3 and Rapideye are 1.8 meters, 1.24 meters and 5 meters, respectively, which are suitable for small-area soybean mapping. Sentinel-2MSI images have a spatial resolution of 10 meters to 60 meters, a width of 290 kilometers, and a revisit period of 5 days. It has been widely used in regional and large-area soybean mapping. In addition, MODIS and Landsat satellites are the most widely used satellite datasets for large-area soybean mapping. MODIS data have a high revisit frequency, however, their coarse spatial resolution (250m–1km) limits their applicability to monitoring small, scattered patches of farmland. Landsat satellite imagery is a suitable data source with a spatial resolution of 30 m, but due to its low temporal resolution (16 days), it is difficult to obtain sufficient cloud-free images for time series analysis. The Gaofen-6 satellite, combined with the Gaofen-1 satellite, has a temporal resolution of 2 days. The GF-6WFV data has a spatial resolution of 16 m, a width of 800 km, and two red edge bands, combining the advantages of most satellites. Therefore, under complex planting structures, the GF-6 WFV data has great potential to effectively map the distribution of soybean planting areas. Summary of the invention

[0003] The purpose of the present invention is to provide a soybean planting area extraction method based on domestic GF-6 WFV data, which helps to obtain more accurate and reliable extraction results, while ensuring accuracy and reducing the storage space and computational complexity occupied by data, thereby improving execution efficiency.

[0004] To achieve the above object, the present invention adopts the following technical solution: a soybean planting area extraction method based on domestic GF-6 WFV data, the method comprising the following steps in order:

[0005] (1) Obtaining GF-6 WFV data of the area to be extracted and preprocessing the data to obtain preprocessed GF-6 WFV data;

[0006] (2) A decision tree hierarchical extraction strategy was used to remove non-crop pixels from the preprocessed GF-6 WFV data to obtain the overall distribution data of farmland vegetation;

[0007] (3) Generate a set of candidate remote sensing features based on the overall distribution data of farmland vegetation;

[0008] (4) Perform feature optimization on the candidate remote sensing feature set to obtain the optimal feature set, and then screen the classifier to obtain the best classifier;

[0009] (5) The optimal feature set and the best classifier are combined to obtain the optimal extraction model, and the performance of the optimal extraction model is evaluated to examine the effect of the optimal extraction model on mapping soybean planting areas.

[0010] The step (1) specifically comprises the following steps:

[0011] (1a) Download the GF-6 WFV data of the area to be extracted from the website of China Resources Satellite Application Center;

[0012] (1b) Data preprocessing, including radiation calibration, atmospheric correction and orthorectification, is performed based on the one-key atmospheric correction and orthorectification module of RSD software. Then, the data is clipped using vector administrative boundary data to obtain preprocessed GF-6 WFV data.

[0013] The step (2) specifically comprises the following steps:

[0014] (2a) The decision tree hierarchical extraction method includes a first-level extraction, a second-level extraction and a third-level extraction, wherein the non-crop pixels refer to non-agricultural landforms other than field vegetation; the first-level extraction is to separate water bodies and buildings in GF-6 WFV data by using the normalized water index NDWI and the normalized vegetation index NDVI, and obtain the first-level extraction result, i.e., GF-6 WFV data with water bodies and building pixels removed;

[0015] (2b) The second layer of extraction is to use the near-infrared band reflectance of the GF-6 WFV data to remove non-agricultural objects other than field vegetation in the first layer of extraction results. Non-agricultural objects other than field vegetation include bare soil and trees, and the second layer of extraction results are obtained, that is, the GF-6 WFV data without bare soil and trees.

[0016] (2c) The third layer of extraction uses the cultivated land distribution data provided by the FROM-GLC10 global land use product in 2017. The spatial resolution of the cultivated land distribution data is resampled to 16 meters to obtain the resampled cultivated land distribution data. The resampled cultivated land distribution data is used to generate a mask file. The mask file is used to mask the second layer of extraction results to obtain the overall distribution data of farmland vegetation.

[0017] The step (3) specifically comprises the following steps:

[0018] (3a) Adding vegetation index and texture features to the preprocessed GF-6 WFV data;

[0019] (3b) Calculate the OIF index of different band combinations of the preprocessed GF-6 WFV data to obtain the optimal band combination, and perform HSV transformation on the preprocessed GF-6 WFV data to obtain three color space features: hue, saturation, and brightness;

[0020] (3c) All vegetation indices, texture features, color space features and preprocessed GF-6WFV data are fused together, and the fused data are masked using the overall distribution data of farmland vegetation to form a complete feature set based on the overall distribution of farmland vegetation, namely, a candidate remote sensing feature set.

[0021] The step (4) specifically comprises the following steps:

[0022] (4a) Based on the candidate remote sensing feature set, sample points covering the main crops in the area to be extracted are selected to generate an area of ​​interest, the value of each candidate feature is extracted, and the importance of the candidate feature is calculated using the ReliefF algorithm;

[0023] (4b) Based on the importance ranking of candidate features, the sequential forward selection method is used to obtain the optimal feature sets of three classifiers: random forest RF, BP neural network BPNN, and support vector machine SVM;

[0024] (4c) Finely classify the 3-meter resolution Planet image to obtain the distribution of soybeans in each sample plot;

[0025] (4d) Using the optimal feature set as input, the soybean classification result diagram of the three classifiers, RF, BPNN, and SVM, was obtained. Based on the Kappa coefficient, the soybean distribution obtained from the Planet image was used to verify the accuracy of the classification results. The higher the Kappa value, the better the classification result. The classifier with the highest Kappa value is the best classifier for extracting soybean planting areas.

[0026] The step (5) specifically comprises the following steps:

[0027] (5a) Input the preferred feature set corresponding to the best classifier into the best classifier to form the best extraction model, and establish three schemes to evaluate the performance of the best extraction model: input the original farmland vegetation overall distribution data into the best classifier, which is scheme one; use the candidate remote sensing feature set as the input of the best classifier, which is scheme two; the best extraction model is scheme three, and the three schemes are used to perform classification using the classifier algorithm in ENVI software to obtain three corresponding soybean planting area extraction results;

[0028] (5b) Using the distribution of soybeans in each sample plot of the Planet image as a reference, a confusion matrix was generated for the soybean planting area extraction results of the three schemes to verify the accuracy. The advantages of the optimal extraction model compared with other schemes were checked based on the Kappa coefficient value, and the mapping effect of the optimal extraction model on the soybean planting area was examined.

[0029] It can be seen from the above technical scheme that the beneficial effects of the present invention are as follows: First, the present invention utilizes a decision tree hierarchical extraction strategy, which can significantly reduce the interference of other landform types such as water bodies, built-up areas, bare soil, and trees on the soybean extraction results, and helps to obtain more accurate and reliable results; Second, the optimal soybean extraction model based on GF-6 WFV images of the present invention uses the preferred feature set as input, which can reduce the storage space and computational complexity of the data while ensuring accuracy, thereby improving execution efficiency. Third, the present invention explores the effect of domestic GF-6 WFV data in achieving planting area extraction in soybean producing areas with complex planting structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2 Schematic diagram of decision tree screening rules;

[0032] Figure 3 This is a schematic diagram of descending order of importance of candidate features;

[0033] Figure 4 , 5 , 6 are schematic diagrams showing the relationship between the number of features and classification accuracy. DETAILED DESCRIPTION

[0034] like Figure 1 As shown, a soybean planting area extraction method based on domestic GF-6 WFV data includes the following steps in order:

[0035] (1) Obtaining GF-6 WFV data of the area to be extracted and preprocessing the data to obtain preprocessed GF-6 WFV data;

[0036] (2) A decision tree hierarchical extraction strategy was used to remove non-crop pixels from the preprocessed GF-6 WFV data to obtain the overall distribution data of farmland vegetation;

[0037] (3) Generate a set of candidate remote sensing features based on the overall distribution data of farmland vegetation;

[0038] (4) Perform feature optimization on the candidate remote sensing feature set to obtain the optimal feature set, and then screen the classifier to obtain the best classifier;

[0039] (5) The optimal feature set and the best classifier are combined to obtain the optimal extraction model, and the performance of the optimal extraction model is evaluated to examine the effect of the optimal extraction model on mapping soybean planting areas.

[0040] The step (1) specifically comprises the following steps:

[0041] (1a) Download the GF-6 WFV data of the area to be extracted from the website of China Resources Satellite Application Center;

[0042] (1b) Data preprocessing, including radiation calibration, atmospheric correction and orthorectification processing, is implemented based on the one-key atmospheric correction and orthorectification module of the domestic RSD software, and then the vector administrative boundary data is used for clipping to obtain the preprocessed GF-6 WFV data.

[0043] like Figure 2 As shown, the step (2) specifically includes the following steps:

[0044] (2a) The decision tree hierarchical extraction method includes a first-level extraction, a second-level extraction and a third-level extraction, wherein the non-crop pixels refer to non-agricultural landforms other than field vegetation; the first-level extraction is to separate water bodies and buildings in GF-6 WFV data by using the normalized water index NDWI and the normalized vegetation index NDVI, and obtain the first-level extraction result, i.e., GF-6 WFV data with water bodies and building pixels removed;

[0045] (2b) The second layer of extraction is to use the near-infrared band reflectance of the GF-6 WFV data to remove non-agricultural objects other than field vegetation in the first layer of extraction results. Non-agricultural objects other than field vegetation include bare soil and trees, and the second layer of extraction results are obtained, that is, the GF-6 WFV data without bare soil and trees.

[0046] (2c) The third layer of extraction uses the cultivated land distribution data provided by the FROM-GLC10 global land use product in 2017, type number 10, and resamples the spatial resolution of the cultivated land distribution data to 16 meters to obtain the resampled cultivated land distribution data. The resampled cultivated land distribution data is used to generate a mask file, and the mask file is used to mask the second layer of extraction results to obtain the overall distribution data of farmland vegetation.

[0047] The step (3) specifically comprises the following steps:

[0048] (3a) Adding vegetation index and texture features to the preprocessed GF-6 WFV data;

[0049] Traditionally, band reflectance is usually used as the remote sensing discriminant feature of a designated object. However, in reality, not all working bands may be effective enough for soybean identification. Therefore, some extended features are considered, such as multiple vegetation indices (calculated based on band reflectance) and texture features. First, PCA transformation is performed on the GF-6 data. Based on the first principal component with the largest amount of information, four texture features that can show the distribution of soybeans are obtained through gray-level co-occurrence matrix calculation: Mean, Entroy, Second Moment, and Homogeneity.

[0050] (3b) Calculate the OIF index of different band combinations of the preprocessed GF-6 WFV data to obtain the optimal band combination, and perform HSV transformation on the preprocessed GF-6 WFV data to obtain three color space features: hue, saturation, and brightness;

[0051] According to the principle of color synthesis of remote sensing images, color images can better identify target objects. Generally, the OIF index is used as the standard. The larger the value, the lower the correlation of the bands and the better the combination effect. The present invention combines the original eight bands to obtain the optimal band combination of band3, 4, and 6, and performs HSV transformation on them to obtain three color space features: Hue, Sat, and Val.

[0052] (3c) All vegetation indices, texture features, color space features and preprocessed GF-6WFV data are fused together, and the fused data are masked using the overall distribution data of farmland vegetation to form a complete feature set based on the overall distribution of farmland vegetation, namely, a candidate remote sensing feature set.

[0053] Table 1 24 candidate remote sensing feature variables selected by the present invention

[0054]

[0055] The step (4) specifically comprises the following steps:

[0056] (4a) Based on the candidate remote sensing feature set, sample points covering the main crops in the area to be extracted are selected to generate an area of ​​interest, the value of each candidate feature is extracted, and the importance of the candidate feature is calculated using the ReliefF algorithm;

[0057] About 1500 sample points covering four main types of land features (soybean, corn, sorghum, and others) were selected to generate regions of interest, and the values ​​of each candidate feature were extracted. The ReliefF algorithm was used to evaluate the importance of 24 candidate feature factors in soybean identification. Due to the inherent randomness of the algorithm, the weight evaluation results may have certain uncertainties. In this paper, the ReliefF algorithm was executed 20 times and the average of the calculation results was taken as the weight value of each feature, as shown in Figure 1. Figure 3 shown.

[0058] (4b) Based on the ranking of candidate features, the sequential forward selection method is used to obtain the optimal feature sets of three classifiers: random forest RF, BP neural network BPNN, and support vector machine SVM, as shown in Table 2 below:

[0059] Table 2 Preferred feature sets for different classifiers

[0060]

[0061] In view of the strong subjectivity of the traditional threshold determination method, the present invention uses an optimal dimension determination method coupled with a classifier, namely the Sequential Forward Selection (SFS). Based on the ranking of candidate feature importance, 24 features are added to the RF, BPNN, and SVM classifiers in order from high to low, one feature at a time. First, the feature with the highest score is input into the classifier as the basic data and the classification accuracy is calculated; then the adjacent features with lower scores are added in turn, and together with the previously added features, new input data is formed and the new extraction accuracy is calculated until all features are input. When the classification accuracy reaches the local optimum, with the continuous addition of subsequent features, the accuracy does not change much and the number of features fluctuates within a small range; in addition, the features that cause the accuracy to decrease after addition are discarded, and the remaining features are used as the preferred feature set. The present invention implements feature variable optimization based on MATLAB 2018a.

[0062] like Figure 4As shown in the figure, with the continuous addition of features, the classification accuracy continues to improve; for the RF classifier, the results show that when the feature dimension reaches 9, marked with a triangle, the classification accuracy reaches the local optimum, and with the addition of subsequent features, the accuracy decreases slightly and fluctuates slightly; when all 24 feature factors are involved in the classification, the classification accuracy reaches the maximum value, but it is only 0.12% higher than the accuracy obtained by the first 9 features; features that fail to improve the classification accuracy after being added are discarded in this study. In the end, the RF classifier selected the top 8 feature factors ranked in the top 9 in feature weight as the preferred feature subset of the model. Similarly, the optimal feature dimension of the BPNN classifier is 9, and two of the features that did not improve the classification accuracy are discarded, such as Figure 5 As shown; the number of optimal feature factors of the SVM classifier is 7, as shown Figure 6 shown.

[0063] (4c) Finely classify the 3-meter resolution Planet image to obtain the distribution of soybeans in each sample plot;

[0064] (4d) Using the optimal feature set as input, the soybean classification result diagram of the three classifiers, RF, BPNN, and SVM, was obtained. Based on the Kappa coefficient, the soybean distribution obtained from the Planet image was used to verify the accuracy of the classification results. The higher the Kappa value, the better the classification result. The classifier with the highest Kappa value is the best classifier for extracting soybean planting areas.

[0065] The present invention uses 3-meter resolution Planet images covering Woyang County to verify the classification results. Eight sample plots of 5 km × 5 km are arranged in the area to be extracted. The sample plots are arranged to be as evenly distributed as possible in space and the proportion of artificial objects contained therein is as small as possible. The present invention uses the maximum likelihood method to classify the Planet images, and the fine soybean distribution results obtained are resampled to the same 16-meter resolution as the GF-6 WFV data as a reference for accuracy verification.

[0066] The present invention uses the preferred feature set corresponding to each classifier as input to perform classification to obtain soybean planting areas. Table 3 shows the extraction accuracy of the three classifiers. The results show that the overall accuracy and Kappa of the eight sample plots obtained based on the support vector machine SVM are higher than those of the other two classifiers. Except for sample plots 6 and 7, the soybean overestimation of the BP neural network BPNN is more serious than that of the support vector machine SVM. The classifier misclassifies more other land types as soybeans, resulting in low user accuracy. The extraction result of the support vector machine SVM is closer to the actual distribution of soybeans, and the extraction effect is better than the other two classifiers. Therefore, the present invention selects the support vector machine SVM as the best classifier for soybean planting area extraction.

[0067] Table 3 Soybean extraction accuracy of three classifiers

[0068]

[0069] The step (5) specifically comprises the following steps:

[0070] (5a) As shown in Table 4, the preferred feature set corresponding to the best classifier is input into the best classifier to form the optimal extraction model, and three schemes are established to evaluate the performance of the optimal extraction model: the original farmland vegetation overall distribution data is input into the best classifier, which is scheme 1; the candidate remote sensing feature set is used as the input of the best classifier, which is scheme 2; the optimal extraction model is scheme 3, and the three schemes are used to perform classification using the classifier algorithm in the ENVI software to obtain three corresponding soybean planting area extraction results;

[0071] Table 4 Different extraction schemes in soybean growing areas

[0072]

[0073] (5b) Using the distribution of soybeans in each sample plot of the Planet image as a reference, a confusion matrix was generated for the soybean planting area extraction results of the three schemes to verify the accuracy. The advantages of the optimal extraction model compared with other schemes were checked based on the Kappa coefficient value, and the mapping effect of the optimal extraction model on the soybean planting area was examined.

[0074] Table 5 shows the soybean extraction accuracy of the eight ground verification plots involved in the three schemes. The Kappa coefficients of the scheme C established based on the preferred features are all over 0.70 except for the 1st and 8th plots. The kappa of scheme B is 0.02 higher than that of scheme C in plot 8, but not higher than that of scheme C in the other plots. In addition, the data volume of the preferred feature set is reduced by 70.83% compared with all features. Therefore, the soybean planting area mapping based on the preferred features not only has greater advantages in data volume and running time, but also the extraction accuracy can meet the actual needs, and the prospects are very broad.

[0075] Table 5 Comparison of extraction accuracy of three schemes

[0076]

[0077] In summary, the present invention utilizes a decision tree hierarchical extraction strategy, which can significantly reduce the interference of other landform types such as water bodies, built-up areas, bare soil, and trees on soybean extraction results, and is helpful to obtain more accurate and reliable results; the optimal soybean extraction model based on GF-6 WFV images of the present invention uses the preferred feature set as input, which can reduce the storage space and computational complexity occupied by data while ensuring accuracy, thereby improving execution efficiency; the present invention explores the effect of domestic GF-6 WFV data in achieving planting area extraction in soybean producing areas with complex planting structures.

Claims

1. A soybean planting area extraction method based on domestic GF-6WFV data, characterized in that: The method comprises the following steps in order: (1) Obtaining GF-6WFV data of the area to be extracted, and preprocessing the data to obtain preprocessed GF-6WFV data; (2) A decision tree hierarchical extraction strategy was used to remove non-crop pixels from the preprocessed GF-6WFV data to obtain the overall distribution data of farmland vegetation; (3) Generate a set of candidate remote sensing features based on the overall distribution data of farmland vegetation; (4) Perform feature optimization on the candidate remote sensing feature set to obtain the optimal feature set, and then screen the classifier to obtain the best classifier; (5) Combine the optimal feature set and the best classifier to obtain the optimal extraction model, evaluate the performance of the optimal extraction model, and examine the effect of the optimal extraction model on mapping soybean planting areas; The step (2) specifically comprises the following steps: (2a) The decision tree hierarchical extraction method includes a first-level extraction, a second-level extraction and a third-level extraction, wherein the non-crop pixels refer to non-agricultural landforms other than field vegetation; the first-level extraction is to separate water bodies and buildings in GF-6WFV data by using the normalized water index NDWI and the normalized vegetation index NDVI, and obtain the first-level extraction result, i.e., GF-6WFV data with water bodies and building pixels removed; (2b) The second layer of extraction is to use the near-infrared band reflectance of the GF-6WFV data to remove non-agricultural objects other than field vegetation in the first layer of extraction results. Non-agricultural objects other than field vegetation include bare soil and trees, and the second layer of extraction results are obtained, that is, the GF-6WFV data without bare soil and trees. (2c) The third layer of extraction uses the cultivated land distribution data provided by the FROM-GLC10 global land use product in 2017. The spatial resolution of the cultivated land distribution data is resampled to 16 meters to obtain the resampled cultivated land distribution data. The resampled cultivated land distribution data is used to generate a mask file. The mask file is used to mask the second layer of extraction results to obtain the overall distribution data of farmland vegetation.

2. The soybean planting area extraction method based on domestic GF-6WFV data according to claim 1, characterized in that: The step (1) specifically comprises the following steps: (1a) Download the GF-6WFV data of the area to be extracted from the website of China Resources Satellite Application Center; (1b) Data preprocessing, including radiation calibration, atmospheric correction and orthorectification processing, is implemented based on the one-key atmospheric correction and orthorectification module of RSD software, and then the vector administrative boundary data is used for clipping to obtain the preprocessed GF-6WFV data.

3. The soybean planting area extraction method based on domestic GF-6WFV data according to claim 1 is characterized in that: The step (3) specifically comprises the following steps: (3a) Adding vegetation index and texture features to the preprocessed GF-6WFV data; (3b) Calculate the OIF index of different band combinations of the preprocessed GF-6WFV data to obtain the optimal band combination, and perform HSV transformation on the preprocessed GF-6WFV data to obtain three color space features: hue, saturation and brightness; (3c) All vegetation indices, texture features, color space features and preprocessed GF-6WFV data are fused together, and the fused data are masked using the overall distribution data of farmland vegetation to form a complete feature set based on the overall distribution of farmland vegetation, namely, a candidate remote sensing feature set.

4. The soybean planting area extraction method based on domestic GF-6WFV data according to claim 1 is characterized in that: The step (4) specifically comprises the following steps: (4a) Based on the candidate remote sensing feature set, sample points covering the main crops in the area to be extracted are selected to generate an area of ​​interest, the value of each candidate feature is extracted, and the importance of the candidate feature is calculated using the ReliefF algorithm; (4b) Based on the importance ranking of candidate features, the sequential forward selection method is used to obtain the optimal feature sets of three classifiers: random forest RF, BP neural network BPNN, and support vector machine SVM; (4c) Finely classify the 3-meter resolution Planet image to obtain the distribution of soybeans in each sample plot; (4d) Using the optimal feature set as input, the soybean classification result diagram of the three classifiers, RF, BPNN, and SVM, was obtained. Based on the Kappa coefficient, the soybean distribution obtained from the Planet image was used to verify the accuracy of the classification results. The higher the Kappa value, the better the classification result. The classifier with the highest Kappa value is the best classifier for extracting soybean planting areas.

5. The soybean planting area extraction method based on domestic GF-6WFV data according to claim 1 is characterized in that: The step (5) specifically comprises the following steps: (5a) Input the preferred feature set corresponding to the best classifier into the best classifier to form the best extraction model, and establish three schemes to evaluate the performance of the best extraction model: input the original farmland vegetation overall distribution data into the best classifier, which is scheme one; use the candidate remote sensing feature set as the input of the best classifier, which is scheme two; the best extraction model is scheme three, and the three schemes are used to perform classification using the classifier algorithm in ENVI software to obtain three corresponding soybean planting area extraction results; (5b) Using the distribution of soybeans in each sample plot of the Planet image as a reference, a confusion matrix was generated for the soybean planting area extraction results of the three schemes to verify the accuracy. The advantages of the optimal extraction model compared with other schemes were checked based on the Kappa coefficient value, and the mapping effect of the optimal extraction model on the soybean planting area was examined.

Images