Long-time-sequence remote sensing crop classification method

Abstract

The invention discloses a long-time-sequence remote sensing crop classification method. The method comprises the following steps: extracting spectral features of a to-be-detected region in different periods; constructing an NDVI phenological characteristic curve of the typical crops; judging the crop category based on the similarity between the curves, and generating a crop classification map of 10 meters of the to-be-detected region; performing upscaling and stratified sampling processing on the crop classification result of the sentinel No.2 satellite, and constructing training and verification samples of Landsat satellite crop classification; carrying out space-time fusion on the Landsat multiband spectral image, and constructing crop supervision classification input based on a characteristic spectrum of the synthesized image and a derived vegetation index; and constructing a crop classifier of a Landsat satellite based on the random forest model, and generating a crop classification map of the to-be-detected region. According to the method, crop classification is carried out by fully utilizing the advantages of high resolution of the sentinel No.2 satellite and long sequence of the Landsat satellite, and long-sequence crop classification drawing of the to-be-detected area can be realized.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing crop classification, and in particular relates to a long-time series remote sensing crop classification method. Background Art

[0002] Crop classification is of great significance to agricultural production, research and management. It is not only related to the planting, management, harvesting and processing of crops, but also an important basis for relevant departments to carry out agricultural precision management, yield forecasting, policy formulation and international trade. Traditional manual crop classification methods (such as field surveys and questionnaire statistics) played an important role in a certain historical period, but their inefficiencies, limited accuracy, insufficient standardization, difficult data management and high costs have limited their application in modern agriculture. With the development of remote sensing and artificial intelligence technology, intelligent classification methods (such as remote sensing technology, machine learning, computer vision, etc.) are gradually replacing traditional manual classification methods, greatly improving the efficiency and accuracy of crop classification.

[0003] Unlike the intensive management of large farms in Europe and the United States, my country's farmland is mainly operated by small-scale family farms and small farmers, and the crop planting structure is complex and diverse. Crop classification based on remote sensing satellites requires satellite images with high spatial resolution. The Sentinel-2 satellite image data launched by the European Space Agency in 2015 has the advantages of a 5-day re-entry period, a spatial resolution of 10m, and multi-spectral, and is widely used in crop classification in various parts of my country. However, the image data of the Sentinel-2 satellite has a short time span (2016 to present), and practical applications often require long-sequence crop type distribution information. Landsat satellite images spanning decades (1982 to present) can make up for the shortcomings of the short sequence of Sentinel-2 satellite data, but its resolution is coarse (30m), and the classification accuracy is not high in areas with complex and changing crop planting structures. Summary of the invention

[0004] The purpose of the present invention is to provide a long-time remote sensing crop classification method to solve the technical problems in the prior art that the remote sensing crop classification time span is short and the accuracy is not high.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for constructing a typical crop NDVI phenological characteristic curve, which specifically comprises the following steps:

[0007] Step 1: Based on statistical yearbooks, literature and field surveys, obtain the types and growth period characteristics of the main crops in the test area, including sowing period, emergence period, vigorous period and harvest period;

[0008] Step 2: Obtain remote sensing images from Sentinel-2 and Landsat satellites during the crop growing season of the test area, and screen and preprocess the remote sensing image data, including radiometric calibration, atmospheric correction, orthorectification, and minimum cloud composition (cloud removal processing);

[0009] Step 3: Using the existing cultivated land distribution data as a mask, extract the Sentinel-2 remote sensing image within the cultivated land area of ​​the test area;

[0010] Step 4: Calculate the Normalized Difference Vegetation Index (NDVI) based on the B4 red visible light (Red) band and the B8 near infrared (NIR) band in the Sentinel-2 remote sensing image: NDVI = (B8 B4) / (B8 + B4);

[0011] Step 5: Based on field surveys, a classification label library (true value sample points) of the main crops in the test area is constructed; based on the classification label library, the NDVI time series during the growth period of each type of crop is extracted to construct the NDVI phenological characteristic curve of each type of crop.

[0012] In a second aspect, the present invention provides a method for supervised crop classification based on Landsat satellite images. In this method, the crop classification results of the Sentinel-2 satellite are used as training data for supervised crop classification of Landsat satellites, and specifically include the following steps:

[0013] S1: Based on the GEE platform, the Sentinel-2 satellite images and Landsat-5, Landsat-7, and Landsat-8 remote sensing images of the crop growing season in the test area are obtained, and the remote sensing image data are screened and preprocessed, including radiation calibration, atmospheric correction, orthorectification, and minimum cloud synthesis;

[0014] S2: Based on field research, a classification label library of typical crops in the test area is established, and the NDVI phenological characteristic curve of typical crops is constructed;

[0015] S3: Based on TWDTW, the similarity between the NDVI curve of the pixel to be classified and the NDVI phenological characteristic curve of typical crops is calculated, and the threshold of the similarity of each type of crop is determined based on the statistical yearbook data to generate a crop classification map of the Sentinel-2 satellite with a resolution of 10 meters in the area to be tested;

[0016] S4: The 10-meter resolution crop classification map generated by the Sentinel-2 satellite was upscaled and “purified”, that is, a 3×3 spatial filter was applied to retain 9 (3×3) pixel matrices with the same crop category; the “purified” classification map was stratified sampled and further screened to obtain the training samples for supervised classification of Landsat images, that is, 70% of the pixels were randomly selected for each crop category to construct “purified” and representative training samples, and the remaining 30% were used as validation samples;

[0017] S5: Perform spatiotemporal fusion of multi-band spectral images of Landsat-5, Landsat-7 and Landsat-8 to generate a 16-day interval, 30-meter resolution synthetic image as the first type of characteristic image for crop classification; generate 6 vegetation index data based on the 6-band spectral images of Landsat satellites as the second type of characteristic image for crop classification;

[0018] S6: Extract two types of feature images at the location of the training samples and input them into the random forest classifier for model training; verify the accuracy of the crop classification results based on the verification samples generated in step S4, use cross-validation between different years to improve the interannual migration performance of the classifier, and finally obtain the random forest classifier with the highest classification accuracy; based on the multispectral synthetic images of Landsat-5, Landsat-7 and Landsat-8 and the derived vegetation index to drive the random forest classifier with the highest classification accuracy, draw a long-term 30-meter resolution crop classification map of the test area.

[0019] Furthermore, the six spectral bands used are blue, green, red, near infrared (NIR), shortwave infrared 1 (SWIR1) and shortwave infrared 2 (SWIR2).

[0020] Furthermore, the derived indicator indices are normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), green chlorophyll vegetation index (GCVI), plant senescence reflectance index (PSRI), normalized difference tillage index (NDTI), and normalized difference senescent vegetation index (NDSVI).

[0021] The crop classification method proposed in the present invention is applicable to the classification of crops of different regions and types, including but not limited to common crops such as wheat, corn, rice and sunflower. In practical applications, the model parameters and classification strategies can be adjusted according to specific needs to adapt to the requirements of different regions and tasks.

[0022] Based on the above technical solution, the embodiments of the present invention can at least produce the following technical effects:

[0023] The crop classification method based on Sentinel-2 satellite and Landsat satellite combined with phenology and machine learning provided by the present invention can efficiently draw a long-time, high-resolution crop classification map for any agricultural area. Compared with the traditional remote sensing crop classification method based on phenology or machine learning, this method makes full use of the advantages of the high resolution of Sentinel-2 satellite and the long sequence of Landsat satellite, has higher classification accuracy and stronger generalization ability, and is particularly suitable for long-term crop type identification in planting areas with complex planting structures, and can provide strong support for the implementation and management of regional precision agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0025] Figure 1 is a flow chart of the remote sensing crop classification method provided by the present invention;

[0026] Figure 2 Taking corn as an example, the comparison between the original and shape-matched NDVI curves and the standard curve is shown in the form of a schematic diagram;

[0027] Figure 3 yes Figure 2 The cumulative distance matrix between the NDVI curve of the pixels to be classified and the standard NDVI curve and its optimal distortion path in the time dimension. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0029] The present invention provides a long-term remote sensing crop classification method, referring to Figure 1 As shown, the specific steps are as follows:

[0030] Step S1: Based on the GEE (Google Earth Engine) platform, the Sentinel-2 satellite images and Landsat-5, Landsat-7, and Landsat-8 remote sensing images of the crop growing season in the test area are obtained, the existing cultivated land distribution data are obtained as masks, and the remote sensing image data are screened and preprocessed, including radiation calibration, atmospheric correction, orthorectification, and minimum cloud synthesis (declouding processing).

[0031] Specifically, in step S1, based on the remote sensing image preprocessing, a remote sensing image of the target area containing only cultivated land is extracted based on the existing cultivated land mask data. The cultivated land mask can be produced by extracting cultivated land grids from public, high-resolution land use data.

[0032] Step S2: Based on field surveys, a classification label library (including crop type, longitude and latitude, and growth period information) of the main crops in the test area is constructed. 20% of the pixels are randomly selected from the ground sample points of each type of crop through the classification label library to average the NDVI time series of its growth period to generate the NDVI phenological characteristic curve of each type of crop.

[0033] Step S3: Calculate the crop NDVI time series based on the red visible light and near infrared reflectance values ​​of the Sentinel-2 satellite, and count the similarity (cumulative distance) between the NDVI curve of the pixel to be classified and the NDVI phenological characteristic curve of the typical crop. Pixels with similarity less than a certain threshold are identified as crop types consistent with the standard curve. Based on this method, a crop classification map with a resolution of 10 meters can be generated for the area to be tested.

[0034] Step S4: The crop classification map generated by the Sentinel-2 satellite is upscaled and “purified”, that is, a 3×3 spatial filter is applied to retain 9 (3×3) pixel matrices with the same crop category. On this basis, stratified sampling and further screening are performed on the “purified” classification map, that is, 70% of the pixels are randomly selected for each type of crop to construct a “purified” and representative supervised classification training sample, and the remaining 30% are used as verification samples.

[0035] Step S5: The multi-band spectral images of Landsat-5, Landsat-7 and Landsat-8 are spatially and temporally fused to generate a 16-day, 30-meter resolution synthetic image and its derived vegetation index, and the original spectrum and derived vegetation index of the synthetic image are extracted as feature inputs of the random forest classifier.

[0036] Step S6: Based on the training samples and feature inputs generated by S4 and S5, the random forest classifier is trained. The crop classification results are evaluated for accuracy using the validation samples generated by S4. The validation samples from different years are cross-validated to improve the inter-annual migration performance of the classifier, thereby obtaining the random forest classifier with the highest classification accuracy. The classifier with the highest classification accuracy is driven by the long sequence of Landsat satellite synthetic images and their derived vegetation indices, and a long-series, 30-meter resolution annual crop classification map is drawn.

[0037] In step S2, the NDVI time series is calculated based on the B4 red visible light (Red) and B8 near infrared (NIR) band reflectance values ​​of the Sentinel-2 satellite: NDVI = (B8-B4) / (B8+B4). The NDVI time series is smoothed using the Savitzky-Golay (SG) filter algorithm to eliminate the time series noise caused by differences in atmospheric conditions, with the filter window size and order being 5 and 3, respectively.

[0038] In step S3, the dynamic time planning method based on time weight (TWDTW) stretches or shortens the NDVI curve of the pixel to be tested in the time dimension. The purpose is to deal with the inconsistency of NDVI curves caused by differences in sowing time for the same type of crops, and find the best alignment with the NDVI phenological characteristic curve of typical crops for shape matching. Figure 2 The standard NDVI curve of corn, the original NDVI curve, and the NDVI curve distorted by the time dimension are shown in schematic form. Figure 3 The cumulative distance matrix and the optimal distortion path of the standard NDVI curve and the original NDVI curve are shown. In addition, in the specific implementation, the similarity threshold of each type of crop can be dynamically optimized and determined based on the sown area of ​​each type of crop provided by the regional statistical yearbook.

[0039] In step S5, the six spectral bands used are blue, green, red, near infrared (NIR), shortwave infrared 1 (SWIR1) and shortwave infrared 2 (SWIR2). The derived vegetation indices (Table 1) are normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), green chlorophyll vegetation index (GCVI), plant senescence reflectance index (PSRI), normalized difference tillage index (NDTI), and normalized difference senescent vegetation index (NDSVI) (see Table 1).

[0040] Table 1 Vegetation index derived from Landsat synthetic images

[0041]

[0042] In step S6, the classification accuracy of crops is evaluated based on the Kappa coefficient and the overall accuracy (OA). The specific calculation formulas of these two indicators are as follows:

[0043]

[0044] In the formula, N represents the total number of pixels to be classified, k is the number of crop categories, and x ii represents the corresponding value of the ith and ith columns, x ij represents the number of pixels correctly classified for the j-th crop, x i+ represents the marginal total of the i-th row, x +i represents the marginal total of the ith column.

[0045] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the attached claims and their equivalents.

Claims

1. A long-term remote sensing crop classification method, characterized in that: The following steps are involved: S1: Extract the spectral characteristics of the area to be measured at different times based on the remote sensing image of the area to be measured, and screen and preprocess the remote sensing image data; S2: Based on field research, a classification label library of typical crops in the test area is established, and the NDVI phenological characteristic curve of typical crops is constructed; S3: Determine the crop category based on the similarity between the NDVI curve of the pixel to be classified and the NDVI phenological characteristic curve of typical crops, and generate a 10-meter crop classification map during the Sentinel-2 satellite observation period; S4: Upscaling and stratified sampling of crop classification results from Sentinel-2 satellites to construct training and validation samples for supervised crop classification of Landsat satellite long-series images; S5: The Landsat multi-band spectral images are spatially and temporally fused, and the input for crop supervised classification is constructed based on the characteristic spectra of the synthetic images and their derived vegetation indices; S6: Construct a crop classifier for Landsat satellite based on the random forest model to generate a long sequence of crop classification maps with a resolution of 30 meters for the area to be tested.

2. The long-term remote sensing crop classification method according to claim 1, characterized in that: In step S1, the remote sensing image data is screened and preprocessed including radiation calibration, atmospheric correction, orthorectification and minimum cloud cover synthesis.

3. The long-term remote sensing crop classification method according to claim 1, characterized in that: In step S2, the following steps are specifically included: S21: Obtain the types and growth period characteristics of the main crops in the tested area through various methods; S22: Obtain remote sensing images from Sentinel-2 and Landsat satellites during the crop growing season in the area to be tested, and screen and preprocess the remote sensing image data; S23: extracting the Sentinel-2 remote sensing image within the scope of the cultivated land in the area to be measured based on the existing cultivated land distribution data as a mask; S24: Calculates the Normalized Difference Vegetation Index (NDVI) based on the B4 red visible light band and B8 near infrared band in the Sentinel-2 remote sensing image; S5: Based on field surveys, a classification label library of major crops in the test area is constructed; based on the classification label library, the NDVI time series during the growth period of each type of crop is extracted to construct the NDVI phenological characteristic curve of each type of crop.

4. The long-term remote sensing crop classification method according to claim 3 is characterized in that: In step S24, the NDVI time series is calculated based on the reflectance values ​​of the B4 red visible light and B8 near-infrared bands of the Sentinel-2 satellite: NDVI = (B8-B4) / (B8+B4). The SG filtering algorithm is used to smooth the NDVI time series to eliminate the time series noise caused by differences in atmospheric conditions. The window size and order of the filter are 5 and 3, respectively.

5. The long-term remote sensing crop classification method according to claim 1, characterized in that: In step S3, a dynamic time planning method based on time weight is used to calculate the similarity between the NDVI curve of the pixel to be classified and the NDVI phenological characteristic curve of typical crops. The threshold of the similarity for each type of crop is determined based on the statistical yearbook data, and a crop classification map with a 10-meter resolution of the Sentinel-2 satellite of the area to be measured is generated.

6. The long-term remote sensing crop classification method according to claim 1, characterized in that: The specific steps of step S4 are as follows: upscaling and purifying the crop classification map generated by the Sentinel-2 satellite, applying a 3×3 spatial filter to retain 9 pixel matrices with the same crop category, stratified sampling and further screening the purified classification map, randomly selecting 70% of the pixels for each type of crop to construct a purified and representative supervised classification training sample, and the remaining 30% as a verification sample.

7. The long-term remote sensing crop classification method according to claim 1, characterized in that: In step S5, the six spectral bands used are blue band, green band, red band, near infrared band, short wave infrared 1 and short wave infrared 2 respectively.

8. The long-term remote sensing crop classification method according to claim 1, characterized in that: In step S5, the derived vegetation indices are respectively normalized difference vegetation index, enhanced vegetation index, chlorophyll vegetation index, plant senescence reflectance index, normalized difference tillage index, and normalized difference senescence vegetation index.

9. The long-term remote sensing crop classification method according to claim 1, characterized in that: The specific steps of step S6 are: training a random forest classifier based on the training samples and feature inputs generated in steps S4 and S5, evaluating the accuracy of crop classification results using the validation samples generated in S4, cross-validating using validation samples from different years to improve the interannual migration performance of the classifier, obtaining the random forest classifier with the highest classification accuracy, driving the classifier with the highest classification accuracy based on a long sequence of Landsat satellite synthetic images and their derived vegetation indices, and drawing an annual crop classification map with a long time series and a resolution of 30 meters.

10. The long-term remote sensing crop classification method according to claim 9, characterized in that: In step S6, the classification accuracy of crops is evaluated based on the Kappa coefficient and the overall accuracy OA. The specific calculation formula is as follows: In the formula, N represents the total number of pixels to be classified, k is the number of crop categories, and x ii represents the corresponding value of the ith and ith columns, x ij represents the number of pixels correctly classified for the j-th crop, x i+ represents the marginal total of the i-th row, x +i represents the marginal total of the ith column.

Images