Plant identification method, device and equipment based on improved multidimensional vector similarity model

Abstract

This invention discloses a plant identification method, apparatus, and equipment based on an improved multidimensional vector similarity model, comprising: selecting a plant planting study area and sample points; downloading Sentinel-2 images for atmospheric correction and stitching / cropping processing, and batch calculating NDVI after processing; reconstructing the temporal NDVI using a fitting analysis algorithm to eliminate the influence of clouds and fog, and extracting vegetation growth curves. Using sample points and growth curves, the plant growth characteristic time is extracted, and the corresponding temporal NDVI and red-edge bands are extracted using the growth characteristic time. Band synthesis is used to reconstruct new multi-band image data. Multidimensional vectors of the synthesized band image are extracted based on the sample points, and the average vector and vector similarity threshold of the sample points are calculated. Then, an improved multidimensional vector similarity model with outlier weights is constructed using the pixel vectors of the synthesized image to extract the plant planting range pixel by pixel. The beneficial effects of this invention are: effectively reducing the complexity of the plant extraction algorithm and ensuring the accuracy of plant extraction.

Description

Technical Field

[0001] The invention belongs to the field of vegetation extraction, and in particular relates to a plant identification method, device and equipment based on an improved multi-dimensional vector similarity model. Background Art

[0002] In recent years, with the extensive application of remote sensing technology, the technology of using spectral remote sensing images to identify plant species in a large area has become more and more mature. However, due to the large differences between the north and the south of my country, the complex climate and geography, and the large number of plant species, the accuracy of using spectral remote sensing images to identify plant species is relatively low, and it is easy to have problems such as misclassification of similar species, different spectra for the same object, and different spectra for the same object. Although deep learning classification algorithms, integrating multi-spectral, hyperspectral, INSAR and other multi-source remote sensing image data, have improved the accuracy of plant identification to a certain extent, deep learning algorithms are relatively complex, and the data processing workload based on multi-source remote sensing images is large, which can easily lead to a significant increase in production costs and time costs, and the phenomenon of different objects with the same spectrum and the same object with different spectra is still relatively serious.

[0003] There are three main reasons for the above problems: one is that the plant growth areas in the central and southern parts of the country are more complex, with lush vegetation; in the northern region, plants grow lushly in summer and autumn, which makes it easy to misclassify autumn crops. Therefore, the use of spectral images in a single time phase can easily cause the same spectrum for different objects and different spectra for the same object; another reason is that the southern region has heavy water vapor, and the process of spectral image identification of plant species is easily affected by water vapor; the third reason is that in areas with more clouds and fog, clouds and fog will cause certain deviations when satellite sensors receive ground reflectivity, thereby affecting the accuracy of plant identification. Summary of the invention

[0004] In order to solve the shortcomings of existing plant identification and classification technologies, the present invention provides a plant identification method based on an improved multidimensional vector similarity model, which can accurately identify plant species.

[0005] The technical solution of the present invention provides a plant identification method based on an improved multidimensional vector similarity model, the method comprising the following steps:

[0006] S1. Obtain the first-level surface reflectance product images of the Sentinel-2 satellite sensor for 12 consecutive months a year, and perform atmospheric correction to obtain the corrected remote sensing images;

[0007] S2. Based on the actual location of the study area or the scope of the administrative area, the corrected remote sensing images are stitched and cut using vectors to obtain the remote sensing images to be processed;

[0008] S3, calculating the normalized vegetation index NDVI of the remote sensing image to be processed, and forming a time series NDVI according to the time sequence;

[0009] S4, using the visual interpretation method, collect samples of the target plants in the remote sensing image to be processed to obtain sample points;

[0010] S5. Perform growth analysis on the time series NDVI and obtain the dates T of all peaks and troughs of the growth curve within the growth cycle of the target plant according to the sample points. i ;

[0011] S6. According to date T i Extract the red edge band of the remote sensing image to be processed in the corresponding time phase, and compare the red edge band with T i The time series NDVI of the time phase is processed by band synthesis to obtain a band synthetic image;

[0012] Extract pixel vectors of band synthetic images based on sample points, purify sample points using clustering methods, remove erroneous sample points, and divide sample points into training samples and verification samples;

[0013] S7, calculating the similarity between the average vector of all sample points of the training sample and the vector of a single sample point of the training sample, and averaging all similarities to obtain an average similarity; the average similarity is used as the similarity threshold of the multidimensional vector similarity model;

[0014] Calculate the similarity between the sample vector of the sample point and the average vector of all sample points to obtain the minimum similarity;

[0015] S8. Calculate the similarity pixel by pixel using the pixel vector of the band synthetic image and the average vector of all sample points, and compare and analyze with the minimum similarity and similarity threshold of the sample points, identify those greater than the similarity threshold as target plants, those less than the minimum similarity as other plants or ground objects, and those between the minimum similarity and the threshold as abnormal similarities, and calculate the abnormal value weight;

[0016] S9. For pixel vectors with abnormal similarity, use the similarity threshold, multidimensional vector similarity, and outlier weight to construct a multidimensional vector similarity model with weighted factors to identify plant species within the outlier range, and verify the accuracy of the classified and identified plants.

[0017] Furthermore, in step S3, the normalized vegetation index NDVI is calculated in batches using python and the gdal library.

[0018] Furthermore, the growth analysis in step S5 uses an NDVI time series analysis model to extract the growth characteristics of vegetation; the NDVI time series analysis model formula is as follows:

[0019]

[0020] In the above formula (1), t represents time, A0 is the harmonic remainder, L represents the length of the time series f(t), ω n is the frequency of each harmonic, θ n is the initial phase of the nth harmonic, A n is the amplitude of the harmonic.

[0021] Furthermore, the similarity r of the multidimensional vector similarity model i The calculation formula is as follows:

[0022]

[0023] In the above formula (2), r i is the similarity between the i-th sample vector and the average vector, x i is the vector of variable x at point i, is the mean vector of variable x, y i is the vector of variable y at point i, is the average vector of variable y; use the training sample set N to extract the pixel vector α on the band synthetic image i , and calculate the training sample n i The average vector Use formula (2) to calculate the similarity r i , and calculate r i The average As the similarity threshold of the multidimensional vector similarity model.

[0024] Further, step S8 is specifically as follows: traverse the pixel vector β of the band synthetic image i The average vector of the plant training samples Use formula (2) to calculate the similarity r i , filter out the minimum similarity value r min , judge r i and and r min The size relationship will be greater than The identification of target plants will be less than r min Identification of other types of plants and features, if r i exist An improved similarity model with outlier weights is used to identify plant species, where the outlier weights are calculated as follows:

[0025]

[0026] In step S9, The r j, using similarity threshold, multidimensional vector similarity, and outlier weight, a multidimensional vector similarity model with outlier weight is constructed. The model formula is as follows:

[0027]

[0028] According to the calculation result in formula (4), r is greater than Identification as target plant, less than are other plants or landforms; after the entire plant identification is completed, the validation sample set is used in the ENVI software to calculate the kappa coefficient and plant classification accuracy to verify the accuracy of the target plant extraction results.

[0029] A plant identification device based on an improved multidimensional vector similarity model, the device comprising:

[0030] Data acquisition unit: obtains the first-level surface reflectance product images of the Sentinel-2 satellite sensor for 12 consecutive months a year, and performs atmospheric correction to obtain the corrected remote sensing images;

[0031] Data processing unit: according to the actual location of the study area or the scope of the administrative area, the vector is used to complete the stitching and cutting of the corrected remote sensing image to obtain the remote sensing image to be processed;

[0032] Time series NDVI calculation unit: calculates the normalized vegetation index NDVI of the remote sensing image to be processed, and composes the time series NDVI according to the time sequence;

[0033] Sample acquisition unit: Use visual interpretation method to collect samples of target plants in the remote sensing image to be processed and obtain sample points;

[0034] Time series NDVI analysis unit: Perform growth analysis on the time series NDVI and obtain the dates T of all peaks and troughs of the growth curve within the growth cycle of the target plant based on the sample points i ;

[0035] Band synthetic image acquisition unit: according to date T i Extract the red edge band of the remote sensing image to be processed in the corresponding time phase, and compare the red edge band with T i The time series NDVI of the time phase is processed by band synthesis to obtain a band synthetic image;

[0036] Pixel vector acquisition unit: extracts pixel vectors of band synthetic images according to sample points, purifies sample points using clustering method, removes erroneous sample points, and divides sample points into training samples and verification samples;

[0037] Multidimensional vector similarity model threshold acquisition unit: calculates the similarity between the average vector of all sample points of the training sample and the vector of a single sample point of the training sample, and averages all similarities to obtain an average similarity; the average similarity is used as the similarity threshold of the multidimensional vector similarity model;

[0038] Calculate the similarity between the sample vector of the sample point and the average vector of all sample points to obtain the minimum similarity;

[0039] Comparative analysis unit: Use the pixel vector of the band synthetic image and the average vector of all sample points to calculate the similarity pixel by pixel, and compare and analyze with the minimum similarity and similarity threshold of the sample points. Those with a similarity greater than the threshold are identified as plants, those with a similarity less than the minimum similarity are identified as other plants or landforms, and those between the minimum similarity and the threshold are considered abnormal similarities, and the weight of the abnormal value is calculated;

[0040] Plant identification unit: For pixel vectors with abnormal similarity, a multidimensional vector similarity model with weighted factors is constructed using similarity threshold, multidimensional vector similarity, and outlier weight to identify plant species within the outlier range, and the accuracy of the classified and identified plants is verified.

[0041] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the plant identification step based on a multi-dimensional vector similarity model is implemented.

[0042] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of plant identification based on a multidimensional vector similarity model are implemented.

[0043] Compared with the prior art, the beneficial effects of the present invention include: utilizing the growth characteristics of different plants to solve the problem of the same object but different spectra and the same spectrum but different objects in spectral images of many plants; using the red edge band of the time-series Sentinel-2 image to effectively reduce the interference caused by water vapor and water bodies, and the combination of the two basically solves the problem of the same object but different spectra, the same object with different spectra, and the misclassification of water bodies; adopting the construction of multidimensional vector similarity as the plant recognition method, effectively reducing the complexity of the extraction algorithm, and utilizing the characteristics of the similarity between plant multidimensional vectors and the outlier weight to ensure the recognition accuracy of the target plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a plant identification method based on an improved multi-dimensional vector similarity model provided by the present invention;

[0045] Figure 2 This is the satellite image for this example;

[0046] Figure 3 This is the plant growth curve of this example;

[0047] Figure 4 This is the plant planting distribution map for this example. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] The present invention provides a plant identification method based on an improved multidimensional vector similarity model. Figure 1 , Figure 1 It is a flow chart of the method of the present invention; the method comprises the following steps:

[0050] S1. Obtain the first-level surface reflectance product images of the Sentinel-2 satellite sensor for 12 consecutive months a year, and perform atmospheric correction to obtain the corrected remote sensing images;

[0051] It should be noted that in step 1, the Sentinel-2 primary and secondary product images within one year are obtained through the GEE platform. The primary product images are atmospherically corrected using the Sen2Cor-02.08.00 software, while the secondary products have been atmospherically corrected by ESA and do not need to be processed, so only the primary product images are processed;

[0052] S2. Based on the actual location of the study area or the scope of the administrative area, the corrected remote sensing images are stitched and cut using vectors to obtain the remote sensing images to be processed;

[0053] It should be noted that the vector range is manually drawn according to the location of the study area, or the administrative division vector is used to stitch and crop the Sentinel-2 satellite remote sensing images that have completed atmospheric correction; specifically, the administrative division map and GEE platform are used to stitch and crop the images.

[0054] S3, calculating the normalized vegetation index NDVI of the remote sensing image to be processed, and forming a time series NDVI according to the time sequence;

[0055] It should be noted that in step S3, Python and the gdal library are used to batch calculate the normalized vegetation index NDVI, and its calculation formula is as follows:

[0056]

[0057] In the above formula, ρ INRis the near infrared band of the image, ρ RED is the red light band of the image;

[0058] S4, using the visual interpretation method, collect samples of the target plants in the remote sensing image to be processed to obtain sample points;

[0059] It should be noted that in step S4, high-resolution images are used to collect plant sample points, and arcgis is used to draw a vector range honeycomb map according to the size of the study area to ensure that the sampling points are randomly located inside the hexagon of the honeycomb map.

[0060] S5. Perform growth analysis on the time series NDVI and obtain the dates T of all peaks and troughs of the growth curve within the growth cycle of the target plant according to the sample points. i ;

[0061] It should be noted that the analysis of the time series NDVI in step S5 can extract the growth cycle of various vegetation and the NDVI index characteristics of each growth stage, which can fully reflect the differences between different vegetation. In the time series NDVI analysis method, Matlab is used to write the NDVI time series fitting analysis model to extract the growth characteristics of vegetation. The core algorithm of this model is to decompose the time spectrum data into many sine curves and cosine curves of different frequencies, and select several curves that can reflect the characteristics of the time series for superposition to achieve the reconstruction of the time series data. This method can quantitatively monitor the dynamic changes of vegetation and eliminate the impact of clouds and fog on NDVI. The model formula is as follows:

[0062]

[0063] In the above formula (1), t represents time, A0 is the harmonic remainder, L represents the length of the time series f(t), ω n is the frequency of each harmonic, θ n is the initial phase of the nth harmonic, A n is the amplitude of the harmonic;

[0064] According to the plant sample points, the dates T of all the peaks and troughs of the growth curve in the plant growth cycle are obtained. i , and use the ENVI band splitting function to split the synthetic time series NDVI into bands.

[0065] S6. According to date T i Extract the red edge band of the remote sensing image to be processed in the corresponding time phase, and compare the red edge band with T i The time series NDVI of the time phase is processed by band synthesis to obtain a band synthetic image;

[0066] Extract pixel vectors of band synthetic images based on sample points, purify sample points using clustering methods, remove erroneous sample points, and divide sample points into training samples and verification samples;

[0067] It should be noted that the red edge band of the remote sensing image in step S6 is more sensitive to the growth conditions of water and vegetation. In the Sentinel-2 image, bands 5, 6, and 7 are red edge bands with wavelengths of 0.705 μm, 0.740 μm, and 0.783 μm, respectively. i Extract the NDVI and red edge bands of the corresponding phase, use ENVI to perform band synthesis processing, and merge them into a multi-band raster image;

[0068] After the plant feature NDVI and red edge bands were synthesized, a clustering algorithm was written based on Matlab. The vector values ​​of the band synthetic images were first extracted using sample points, and a clustering screening was performed to remove the wrong samples. The screened sample points were randomly divided into a training sample set (n i ∈N) and the validation sample set (m i ∈M).

[0069] S7, calculating the similarity between the average vector of all sample points of the training sample and the vector of a single sample point of the training sample, and averaging all similarities to obtain an average similarity; the average similarity is used as the threshold of the multidimensional vector similarity model;

[0070] It should be noted that in step S7, python and third-party libraries such as GDAL and Scikit-learn are used to write a multidimensional vector similarity model as a plant extraction model. The model is optimized by cosine similarity and covariance, eliminating the influence of different standards of each component (score expansion), and has translation invariance and scale invariance. The numerical range of similarity r is [-1, 1]. The closer to 1, the higher the similarity. The similarity r of the multidimensional vector similarity model is i The calculation formula is as follows:

[0071]

[0072] In the above formula (2), r i is the similarity between the i-th sample vector and the average vector, x i is the vector of variable x at point i, is the mean vector of variable x, y i is the vector of variable y at point i, is the average vector of variable y; use the training sample set N to extract the pixel vector α on the band synthetic image i , and calculate the training sample n i The average vector Use formula (2) to calculate the similarity r i , and calculate r i The average As the similarity threshold of the multidimensional vector similarity model.

[0073] S8. Use the pixel vector of the band synthetic image and the average vector of all sample points to calculate the similarity pixel by pixel, and compare and analyze with the minimum similarity and similarity threshold of the sample points. Those greater than the similarity threshold are identified as target plants, those less than the minimum similarity are judged as other plants or ground objects, and those between the minimum similarity and the threshold are considered abnormal similarities, and the outlier weight is calculated.

[0074] It should be noted that step S8 specifically includes: traversing the pixel vector β of the band synthetic image i The average vector of the plant training samples Use formula (2) to calculate the similarity r i , filter out the minimum similarity value r min , judge r i and and r min The size relationship will be greater than The identification of target plants will be less than r min Identification of other types of plants and features, if r i exist An improved similarity model with outlier weights is used to identify plant species, where the outlier weights are calculated as follows:

[0075]

[0076] S9. For pixel vectors with abnormal similarity, use the similarity threshold, multidimensional vector similarity, and outlier weight to construct a multidimensional vector similarity model with weighted factors to identify plant species within the outlier range, and verify the accuracy of the classified and identified plants.

[0077] It should be noted that in step S9, The r j , using similarity threshold, multidimensional vector similarity, and outlier weight, a multidimensional vector similarity model with outlier weight is constructed. The model formula is as follows:

[0078]

[0079] According to the calculation result in formula (4), r is greater than Identification as target plant, less than are other plants or landforms; after the entire plant identification is completed, the validation sample set is used in the ENVI software to calculate the kappa coefficient and plant classification accuracy to verify the accuracy of the target plant extraction results.

[0080] The present invention provides a specific embodiment as follows (in this embodiment, the target plant is rice):

[0081] Step 1, the plant extraction area is selected as Jianli County, Jingzhou City, Hubei Province, the time is 2020, and the GEE platform is used to download the Sentinel-2 remote sensing images of the time series within one year, and the Sen2Cor-02.08.0 software is used to perform atmospheric correction on the L1C-level remote sensing images to generate L2A image products. The temporal resolution of Sentinel-2 images is 5 days, but there are many clouds in the area and the clouds and fog are seriously blocked. Therefore, the images with more than 60% clouds and fog in Jianli County are removed, and finally, remote sensing images of no less than 1 period are selected for analysis every month. This embodiment has a total of 23 images. Among them, Jianli County harvests and plants double-season rice from April to November.

[0082] Step 2, Figure 2 For high-resolution image maps of Jianli County, the 2020 Google Image and Gaofen-2 images were used to collect rice sample points in Jianli County.

[0083] Step 3: Use Python and GDAL 3.3 to batch calculate the NDVI of Sentinel-2 images, and then arrange them in sequence to form time series data.

[0084] Step 4: Use Matlab to write a time series analysis algorithm, synthesize the time series NDVI data into bands, and then analyze the growth of rice. According to the growth curve and rice sample points, extract the rice growth curve, and use the peaks and troughs of the curve as the growth characteristics of rice. Finally, use the band splitting function of ENVI software to split the band-synthesized time series NDVI image to generate a single-scene NDVI. Figure 3 This is the average growth curve of rice in Jianli County. It can be seen from the figure that the rice is at trough when the horizontal axis is 8, 11, 15, and 22, corresponding to the early April, early June, mid-July, and mid-November; it is at peak when the horizontal axis is 10, 12, and 16, corresponding to the early May, the end of June, and the end of August.

[0085] Step 5, extract the red edge band and NDVI of the remote sensing images in early April, early May, early June, late June, mid-July, and late August, and re-synthesize the bands in chronological order, then use the sample points to extract the vectors of the synthetic images, and finally use the clustering algorithm to screen the sampling points, and randomly divide the sample points into training samples and verification samples in a ratio of 7:3.

[0086] Step 6, extract the vector of the synthetic image using the training sample in ArcGIS, and then use Python to calculate the average vector, so that each extracted sample vector is similar to the average vector, and finally calculate the average similarity. In this embodiment, the minimum similarity of the sample points is 0.9701, the average similarity is 0.9762, and this data is used as the threshold.

[0087] Step 7, use Python to read the pixel vector of the synthetic image pixel by pixel, and use the multidimensional vector similarity model to calculate the multidimensional vector similarity of each pixel based on the sample average vector and compare it with 0.9762. Those greater than 0.9762 are judged as rice, and those less than 0.9701 are judged as other objects. The outlier weight ∈ is calculated between [0.9701, 0.9762]. In this embodiment, ∈=2(0.9762-r i ) / r i +0.9762.

[0088] Step 8, use 0.9762, r i ,∈,r=r i +∈The rice in [0.9701,0.9762] is identified and extracted again, and the results r less than 0.9762 are identified as rice, and those less than For other plants or land features.

[0089] Step 9: After the rice extraction is completed, the verification sample and the classification accuracy analysis module of ENVI software are used to analyze the extraction accuracy of the rice. The total accuracy of rice extraction by this method is 96.8%. Figure 4 This is the rice extraction result of this embodiment. It can be seen from the figure that the rice planting area in Jianli County is relatively concentrated and large in scale.

[0090] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0091] The beneficial effects of the present invention are as follows: the growth characteristics of different plants are utilized to solve the problem of the same object with different spectra and the same spectrum with different objects in spectral images of many plants; the red edge band of the time-series Sentinel-2 image is used to effectively reduce the interference caused by water vapor and water bodies, and the combination of the two basically solves the problem of the same object with different spectra, the same object with different spectra, and the misclassification of water bodies; the construction of multidimensional vector similarity is adopted as the plant recognition method, which effectively reduces the complexity of the extraction algorithm, and utilizes the characteristics of the similarity between plant multidimensional vectors and the outlier weight, thereby ensuring the recognition accuracy of the target plant.

[0092] The specific implementation of the present invention described above does not constitute a limitation on the protection scope of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A plant identification method based on an improved multidimensional vector similarity model, characterized in that: The following steps are involved: S1. Obtain the first-level surface reflectance product images of the Sentinel-2 satellite sensor for 12 consecutive months a year, and perform atmospheric correction to obtain the corrected remote sensing images; S2. Based on the actual location of the study area or the scope of the administrative area, the corrected remote sensing images are stitched and cut using vectors to obtain the remote sensing images to be processed; S3, calculating the normalized vegetation index NDVI of the remote sensing image to be processed, and forming a time series NDVI according to the time sequence; S4, using the visual interpretation method, collect samples of the target plants in the remote sensing image to be processed to obtain sample points; S5. Perform growth analysis on the time series NDVI and obtain the dates T of all peaks and troughs of the growth curve within the growth cycle of the target plant according to the sample points. i ; S6. According to date T i Extract the red edge band of the remote sensing image to be processed in the corresponding time phase, and compare the red edge band with T i The time series NDVI of the time phase is processed by band synthesis to obtain a band synthetic image; Extract pixel vectors of band synthetic images based on sample points, purify sample points using clustering methods, remove erroneous sample points, and divide sample points into training samples and verification samples; S7, calculating the similarity between the average vector of all sample points of the training sample and the vector of a single sample point of the training sample, and averaging all similarities to obtain an average similarity; using the average similarity as a similarity threshold of the multidimensional vector similarity model; Calculate the similarity between the sample vector of the sample point and the average vector of all sample points to obtain the minimum similarity; S8. Calculate the similarity pixel by pixel using the pixel vector of the band synthetic image and the average vector of all sample points, and compare and analyze with the minimum similarity and similarity threshold of the sample points, identify those greater than the similarity threshold as target plants, those less than the minimum similarity as other plants or ground objects, and those between the minimum similarity and the threshold as abnormal similarities, and calculate the abnormal value weight; S9, for pixel vectors with abnormal similarity, using similarity threshold, multidimensional vector similarity, and outlier weight, construct a multidimensional vector similarity model with weighted factors to identify plant species within the outlier range, and verify the accuracy of the classified and identified plants; The similarity r of the multidimensional vector similarity model in step S7 i The calculation formula is as follows: In the above formula (2), r i is the similarity between the i-th sample vector and the average vector, x i is the vector of variable x at point i, is the mean vector of variable x, y i is the vector of variable y at point i, is the average vector of variable y; use training sample set N to extract pixel vector α1 on the band synthetic image and calculate training sample n i The average vector Use formula (2) to calculate the similarity r i , and calculate r i The average As a similarity threshold for the multi-dimensional vector similarity model; Step S8 is specifically as follows: traverse the pixel vector β of the band synthetic image i The average vector of the plant training samples Use formula (2) to calculate the similarity r i , filter out the minimum similarity value r min , judge r i and and r min The size relationship will be greater than The identification of target plants will be less than r min Identification of other types of plants and features, if r i exist An improved similarity model with outlier weights is used to identify plant species, where the outlier weights are calculated as follows: In step S9, The r between j , using similarity threshold, multidimensional vector similarity, and outlier weight, a multidimensional vector similarity model with outlier weight is constructed. The model formula is as follows: According to the calculation result in formula (4), r is greater than Identification as target plant, less than are other plants or landforms; after the entire plant identification is completed, the validation sample set is used in the ENVI software to calculate the kappa coefficient and plant classification accuracy to verify the accuracy of the target plant extraction results.

2. A plant identification method based on an improved multidimensional vector similarity model as claimed in claim 1, characterized in that: In step S3, Python and the gdal library are used to batch calculate the normalized vegetation index NDVI.

3. A plant identification method based on an improved multidimensional vector similarity model as claimed in claim 1, characterized in that: The growth analysis in step S5 uses the NDVI time series analysis model to extract the growth characteristics of vegetation; the NDVI time series analysis model formula is as follows: In the above formula (1), t represents time, A0 is the harmonic remainder, L represents the length of the time series f(t), ω n is the frequency of each harmonic, θ n is the initial phase of the nth harmonic, A n is the amplitude of the harmonic.

4. A plant identification device based on an improved multidimensional vector similarity model, characterized in that: The device comprises: Data acquisition unit: obtains the first-level surface reflectance product images of the Sentinel-2 satellite sensor for 12 consecutive months a year, and performs atmospheric correction to obtain the corrected remote sensing images; Data processing unit: according to the actual location of the study area or the scope of the administrative area, the vector is used to complete the stitching and cutting of the corrected remote sensing image to obtain the remote sensing image to be processed; Time series NDVI calculation unit: calculates the normalized vegetation index NDVI of the remote sensing image to be processed, and composes the time series NDVI according to the time sequence; Sample acquisition unit: Use visual interpretation method to collect samples of target plants in the remote sensing image to be processed and obtain sample points; Time series NDVI analysis unit: Perform growth analysis on the time series NDVI and obtain the dates T of all peaks and troughs of the growth curve within the growth cycle of the target plant based on the sample points i ; Band synthetic image acquisition unit: according to date T i Extract the red edge band of the remote sensing image to be processed in the corresponding time phase, and compare the red edge band with T i The time series NDVI of the time phase is processed by band synthesis to obtain a band synthetic image; Pixel vector acquisition unit: extracts pixel vectors of band synthetic images according to sample points, purifies sample points using clustering method, removes erroneous sample points, and divides sample points into training samples and verification samples; Multidimensional vector similarity model threshold acquisition unit: calculates the similarity between the average vector of all sample points of the training sample and the vector of a single sample point of the training sample, and averages all similarities to obtain an average similarity; the average similarity is used as the similarity threshold of the multidimensional vector similarity model; Calculate the similarity between the sample vector of the sample point and the average vector of all sample points to obtain the minimum similarity; Comparative analysis unit: Use the pixel vector of the band synthetic image and the average vector of all sample points to calculate the similarity pixel by pixel, and compare and analyze with the minimum similarity and similarity threshold of the sample points. Those with a similarity greater than the threshold are identified as plants, those with a similarity less than the minimum similarity are identified as other plants or landforms, and those between the minimum similarity and the threshold are considered abnormal similarities, and the weight of the abnormal value is calculated; Plant identification unit: For pixel vectors with abnormal similarity, a multidimensional vector similarity model with weighted factors is constructed to identify plant species within the abnormal value range using similarity threshold, multidimensional vector similarity, and outlier weight, and the accuracy of the classified and identified plants is verified; In the multidimensional vector similarity model threshold acquisition unit, the similarity r of the multidimensional vector similarity model i The calculation formula is as follows: In the above formula (2), r i is the similarity between the i-th sample vector and the average vector, x i is the vector of variable x at point i, is the mean vector of variable x, y i is the vector of variable y at point i, is the average vector of variable y; use the training sample set N to extract the pixel vector α on the band synthetic image i , and calculate the training sample n i The average vector Use formula (2) to calculate the similarity r i , and calculate r i The average As a similarity threshold for the multi-dimensional vector similarity model; The comparison and analysis unit is specifically: traversing the pixel vector β of the band synthetic image i The average vector of the plant training samples Use formula (2) to calculate the similarity r i , filter out the minimum similarity value r min , judge r i and and r min The size relationship will be greater than The identification of target plants will be less than r min Identification of other types of plants and features, if r i exist An improved similarity model with outlier weights is used to identify plant species, where the outlier weights are calculated as follows: Comparative analysis unit The r between j , using similarity threshold, multidimensional vector similarity, and outlier weight, a multidimensional vector similarity model with outlier weight is constructed. The model formula is as follows: According to the calculation result in formula (4), r is greater than Identification as target plant, less than are other plants or landforms; after the entire plant identification is completed, the validation sample set is used in the ENVI software to calculate the kappa coefficient and plant classification accuracy to verify the accuracy of the target plant extraction results.

5. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the plant identification method based on the improved multidimensional vector similarity model as claimed in any one of claims 1 to 3 are implemented.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the plant identification method based on the improved multidimensional vector similarity model as claimed in any one of claims 1 to 3 are implemented.

Images