A classification method based on weighted vegetation time series attribute characteristic curve

Abstract

The present invention provides a classification method based on weighted vegetation time series attribute characteristic curves, selects multi-view dual-polarization Sentinel-1 data within one year, uses an object-oriented multi-scale segmentation method to segment the image and takes the segmented units as objects, performs weighted analysis on its time series characteristics according to the vegetation growing season conditions, uses vegetation growing season weighted DTW to obtain the correlation between object time series curves as a new time series attribute characteristic curve, and then uses the nearest neighbor classification KNN algorithm for classification. The present invention increases the weight between vegetation growing seasons in a targeted manner and extracts the similarity between vegetation time series curves of each object as an attribute value to form a new time series attribute curve, and improves the classification accuracy of vegetation categories by mining the correlation between deep-level attribute curves.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and mainly relates to polarimetric SAR remote sensing data temporal feature extraction and classification, and specifically is a classification method based on weighted vegetation temporal attribute characteristic curves. Background Art

[0002] In recent years, synthetic aperture radar (SAR) technology has received extensive attention in the field of remote sensing research because of its multiple characteristics such as all-day, all-weather, high resolution, wide bandwidth, strong penetration, etc., which can make up for the defects of missing optical data and cloud cover. Polarimetric SAR has multiple polarization channels, and can obtain vegetation information that cannot be monitored by optical remote sensing through polarization scattering and polarization vegetation index. Therefore, it has been widely used in dense vegetation areas with high water vapor content in recent years, and is also used to monitor the entire growth cycle of vegetation.

[0003] Most vegetation has certain seasonal and phenological characteristics, so remote sensing image vegetation time series analysis is crucial for the extraction of vegetation information. At present, the research on vegetation time series curves is mainly based on two directions: phenological information extraction and similarity. Phenological information extraction is to obtain representative special points that can reflect the growth state of vegetation over time from the vegetation time series characteristic curve, which is used to distinguish different vegetation; similarity is based on the curve itself, and the distance between curves is calculated as the similarity measure of the curves, and the curves are classified according to this similarity. Dynamic Time Warping (DTW) calculates the similarity between two time series by stretching the time series. However, the similarity measure of DTW curves is calculated and analyzed based on the entire curve, and there is little focus on local time periods.

[0004] In addition, most of the previous remote sensing research was based on pixel analysis, while ignoring the relevant characteristic attributes such as the shape, texture and spatial topological relationship of the target object. The object-oriented remote sensing image analysis method can effectively avoid the "pepper and salt effect" caused by the large spectral variation of similar objects in high-resolution images while taking into account the above characteristic attributes. In addition, according to the optimal segmentation scale of different regional features, object-oriented multi-scale segmentation can divide the image into several non-intersecting areas, ensure the consistency of object information, greatly reduce the time required for classification and further improve the classification accuracy.

[0005] Aiming at multi-time series polarimetric SAR data, the present invention proposes a classification method based on the combination of object-oriented segmentation and weighted vegetation time series attribute features. The segmentation unit is taken as the object and the vegetation growing season is taken as the key monitoring period. The classification is performed after obtaining the similarity of the vegetation time series attribute (DTW) curves of each object. Summary of the invention

[0006] The purpose of the present invention is to propose a classification method based on weighted vegetation time series attribute characteristic curves, select multi-view dual-polarization Sentinel-1 data within one year, use object-oriented multi-scale segmentation method to segment the image and take the segmented units as objects, perform weighted analysis on its time series characteristics according to the vegetation growing season conditions, use vegetation growing season weighted DTW to obtain the correlation between object time series curves as new time series attribute characteristic curves, and then use the nearest neighbor classification (K-Nearest Neighbor, KNN) algorithm for classification. By increasing the weight of the growing season, the calculation and analysis of the similarity measurement of the time series curve is more targeted and the classification accuracy is higher.

[0007] Therefore, a classification method based on weighted vegetation time series attribute characteristic curve, specifically, the steps are as follows;

[0008] Step 1: Download multi-view dual-polarization Sentinel-1SAR images within one year and perform radiometric calibration, multi-view filtering and registration operations to eliminate or reduce coherent speckle noise and improve the visual interpretation of the image;

[0009] Step 2, selecting one scene from the preprocessed multi-scene images for object-oriented multi-scale segmentation and selecting objects of different categories as training samples;

[0010] Step 3, extracting the polarization vegetation index of each scene image;

[0011] Step 4, averaging each segmented object in each scene polarization vegetation index feature image and sequentially constructing an object-oriented polarization vegetation index time series polyline;

[0012] Step 5, smoothing and filtering the vegetation time series polyline of each object;

[0013] Step 6, using the weighted DTW method to extract the transformation attributes of the vegetation time series curve of each object and form an attribute curve;

[0014] Step 7, using the KNN algorithm to classify the attribute curve of the object to be classified according to the attribute curve of the training sample;

[0015] Furthermore, in step 2, an object-oriented multi-scale segmentation method is used to merge pixels from bottom to top according to the shape, texture and spatial position relationship characteristics of the target object, so that the above characteristics show consistency or similarity within the same object, and the optimal segmentation scale is obtained through multiple experiments combined with visual interpretation of the segmentation results.

[0016] Furthermore, in step 3, the formula of the Radar Vegetation Index (RVI) is as follows:

[0017]

[0018] In the above formula and are the backscattering coefficients of polarization image data, respectively.

[0019] Furthermore, in step 5, the extracted polarization vegetation index time series curve is smoothed by using the difference method and the Savitzky-Golay filter (usually referred to as the SG filter) to obtain a time series fitting curve. The biggest feature of this filter is that it can ensure that the shape and width of the time series curve remain unchanged while filtering out noise. The two parameters contained in the SG filter include the sliding window value and the polynomial order. In order to prevent details from being filtered out or the smoothing effect from being unclear, the sliding window value is generally between 3 and 7, so that more values ​​are involved in the fitting and the effect is better; the polynomial is generally between 2 and 4, and the lower the order, the better the smoothing effect.

[0020] Furthermore, in step 6, since the values ​​at each time point in each time series can be used as the attribute values ​​of this time series, they are important factors affecting the similarity measurement results, and the time series attributes can more effectively measure the correlation between them and the analysis accuracy. Generally, time series analysis is based on the time values ​​in the time series as attributes for analysis. The present invention uses the similarity measurement algorithm of vegetation growing season weighted DTW to calculate the similarity measurement value of each two segmented object time series curves to form a new attribute time series curve, and then analyzes the similarity of the object time series curves after the attribute transformation for the next step of analysis.

[0021] Weighted DTW is based on DTW and performs distance weighting within the range of vegetation growing season, and obtains the similarity measurement value between curves according to vegetation growing season in a targeted manner. The weighting is performed using the modified logistic weight function (MLWF) which is flexible in setting weight limits.

[0022]

[0023] Where i = 2, 3, ..., m represents the row index of the matrix, and j = 2, 3, ..., n represents the column index of the matrix. D(i, j) is the minimum cumulative value of the path. 1 <i<i 2 Refers to the vegetation growing season, d ij is the Euclidean distance between two values ​​in the matrix formed by two time series curves. i-j is the weight between the matching points of the two time series curves. The weight function is as follows:

[0024]

[0025] Where i = 1, 2, ... m, m is the length of the sequence, m cis the midpoint of the sequence. max is the expected upper limit of the weight parameter, g is the empirical constant of the control function curvature (slope), the optimal g ranges from 0.01 to 0.6, and the larger the g, the greater the weight;

[0026] Furthermore, in step 7, the Euclidean distance between the object to be classified and the training sample is calculated based on the transformed attribute time series obtained in the previous step, and then the nearest neighbor classification (KNN) algorithm is used for classification. The Euclidean distance is as follows:

[0027]

[0028] Where X and Y are the training sample object curve and the sample object curve to be classified respectively; n refers to the corresponding point randomly selected on the two curves; x i is the value of the i-th point on the X curve; y i is the value of the i-th point on the Y curve. Find the K training samples that are the nearest neighbors to the object to be classified. The object to be classified is assigned to the category with the largest number of these K training samples.

[0029] The present invention provides a classification method based on weighted vegetation time series attribute characteristic curves, which has the technical effect of specifically increasing the weights between vegetation growing seasons and extracting the similarities between vegetation time series curves of each object as attribute values ​​to form new time series attribute curves, and improving the classification accuracy of vegetation categories by mining the correlation between deep-level attribute curves. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A schematic diagram of a flow chart of a classification method based on a weighted vegetation time series attribute characteristic curve provided by the present invention;

[0031] Figure 2 A schematic diagram of the time series attribute transformation provided by the present invention;

[0032] Figure 3 A schematic diagram of the DTW (Dynamic Time Warping) dynamic time warping algorithm provided by the present invention;

[0033] Figure 4 This is a schematic diagram of the weighted area of ​​vegetation growing season provided by the present invention. DETAILED DESCRIPTION

[0034] The following examples are only used to more clearly illustrate the technical solution of the present invention. Figure 1 :

[0035] Step 1: Download multiple Sentinel-1 time-series images within a year and perform preprocessing such as radiometric calibration, multi-view filtering, and registration to eliminate or reduce coherent speckle noise and improve the visual interpretation of images. The selection of Sentinel-1 time-series images is mainly based on the growing season image data of the research area and object, supplemented by the off-season image data.

[0036] Step 2: Select one scene from the preprocessed multi-scene images for object-oriented multi-scale segmentation, obtain objects of different shapes and sizes, and select objects of different categories as samples for subsequent classification; object-oriented multi-scale segmentation can merge pixels from bottom to top according to the shape, texture, and spatial position relationship of the target object. The optimal segmentation scale is obtained through multiple experiments combined with visual interpretation of the segmentation results.

[0037] Step 3: Extract the polarization vegetation index of each scene image; the formula of the polarization vegetation index RVI (Radar Vegetation Index) is as follows:

[0038]

[0039] In the above formula and are the backscatter coefficients of polarization image data respectively. The polarization vegetation index formula is only for dual-polarization data, and the polarization vegetation index formula for full polarization is different from that shown in the present invention.

[0040] Step 4: average the segmented objects in each scene's polarimetric vegetation index feature image and sequentially construct an object-oriented polarimetric vegetation index time series polyline. All objects have a time series polyline representing their internal features and proceed to the next step of analysis.

[0041] Step 5, smoothing and filtering the vegetation time series curves of each object: the extracted polarization vegetation index time series curves are smoothed and filtered using the difference method and the Savitzky-Golay filter (usually referred to as the SG filter) to obtain the time series fitting curve. The biggest feature of this filter is that it can ensure that the shape and width of the time series curve remain unchanged while filtering out noise. The two parameters contained in the SG filter include the sliding window value and the polynomial order. In order to prevent details from being filtered out or the smoothing effect is not obvious, the sliding window value is generally between 3 and 7, so that more values ​​are involved in the fitting and the effect is better; the polynomial is generally between 2 and 4, and the lower the order, the better the smoothing effect.

[0042] Step 6, using the weighted DTW method to extract the transformation attributes of each object vegetation time series curve; since the values ​​at each time point in each time series can be used as the attribute value of this time series, they are important factors affecting the similarity measurement results, and time series attributes can more effectively measure their mutual correlation and analysis accuracy. Generally, time series analysis is based on the time value in the time series as an attribute for analysis. The present invention uses the similarity measurement algorithm of the weighted DTW of the vegetation growing season to calculate the similarity measurement value of each two segmented object time series curves to form a new attribute time series curve, and then analyzes the new attribute curve for the next step of analysis, such as Figure 2 .

[0043] In addition, the basic dynamic time warping algorithm (DTW) is a dynamic programming algorithm that calculates the similarity between two time series, especially sequences of different lengths. Given two time series curves A and B, their lengths are m and n respectively (m=n when processing remote sensing data).

[0044] A=a 1 , a 2 , a 3 ...a m

[0045] B=b 1 , b 2 , b 3 ...b n

[0046] First construct an n×m matrix D for alignment operation. The matrix element d ij Indicates a i and b j The distance between two points d(a i , b j ), this distance calculation uses the Euclidean distance, that is,

[0047]

[0048] The DTW algorithm is to find a path from the origin to (A m ,B n ) is the shortest path. The path is defined as W, W = {w 1 ,w 2, w 3… w k}, the kth element w of W k =(c ij ) k ,like Figure 3 The curved path should meet the following three conditions:

[0049] (1) Boundedness, i.e. w 1=(1,1),w 2 =(m,n) and max{m,n} <K≤m+n-1;

[0050] (2) Continuity means that if w k-1 =(i',j'), then w k =(i,j) must satisfy (i-i')<<1 and (j-j')<<1;

[0051] (3) Monotonicity means that if w k-1 =(i',j'), then w k =(i,j) must satisfy (i-i')>>0 and (j-j')>>0;

[0052] Find the shortest curved path, i.e. the minimum cumulative distance DTW(A,B), under the following conditions:

[0053]

[0054]

[0055] Where i = 2, 3, ..., m represents the row index of the matrix, j = 2, 3, ..., n represents the column index of the matrix. D(i, j) is the minimum cumulative value of the path.

[0056] The weighted DTW performs distance weighting within the vegetation growing season based on DTW, and obtains the similarity measurement value between curves according to the vegetation growing season. The improved logistic weight function (MLWF) is used for weighting, such as Figure 4 .

[0057]

[0058] Where i 1 <i<i 2 Refers to the vegetation growing season period, v i-j is the weight between the matching points of the two time series curves. The weight function is as follows:

[0059]

[0060] Where i = 1, 2, ... m, m is the length of the sequence, m c is the midpoint of the sequence. max is the expected upper limit of the weight parameter, g is the empirical constant that controls the curvature (slope) of the function, and the optimal g ranges from 0.01 to 0.6. The larger g is, the greater its weight is.

[0061] Step 7: Use the KNN algorithm to classify the attribute curves of the objects to be classified according to the attribute curves of the training samples; use the Euclidean distance to calculate the distance between the attribute curves of the samples to be classified and the attribute curves of the training samples and sort them. The distance formula is as follows:

[0062]

[0063] Where X and Y are the training sample object curve and the sample object curve to be classified respectively; n refers to the corresponding point randomly selected on the two curves; x i is the value of the i-th point on the X curve; y i is the value of the i-th point on the Y curve. By determining the K value, that is, the number of objects to be classified that are closest to the training sample, find the K training samples that are the nearest neighbors to the sample to be classified, and count the category with the most occurrences of these K neighboring training samples, and then assign the object to be classified to this category.

Claims

1. A classification method based on weighted vegetation time series attribute characteristic curve, characterized in that: Multi-view dual-polarization Sentinel-1 data within one year are selected, and the image is segmented by object-oriented multi-scale segmentation method. The segmented units are taken as objects, and the time series characteristics are weighted analyzed according to the vegetation growing season. The correlation between the object time series curves is obtained by using the vegetation growing season weighted DTW as the new time series attribute characteristic curve, and then the nearest neighbor classification KNN algorithm is used for classification. The steps include: Step 1: Download multi-view dual-polarization Sentinel-1SAR images within one year and perform radiometric calibration, multi-view filtering and registration operations to eliminate or reduce coherent speckle noise and improve the visual interpretation of the image; Step 2, selecting one scene from the preprocessed multi-scene images for object-oriented multi-scale segmentation and selecting objects of different categories as training samples; Step 3, extracting the polarization vegetation index of each scene image; Step 4, averaging each segmented object in each scene polarization vegetation index feature image and sequentially constructing an object-oriented polarization vegetation index time series polyline; Step 5, smoothing and filtering the vegetation time series polyline of each object; Step 6, using the weighted DTW method to extract the transformation attributes of the vegetation time series curve of each object and form an attribute curve; Using the similarity measurement algorithm of vegetation growing season weighted DTW, the similarity measurement value between each two segmented object time series curves is calculated to form a new attribute time series curve, and then the next step of analysis is carried out by analyzing the similarity of the object time series curves after the attribute transformation; Weighted DTW performs distance weighting within the vegetation growing season based on DTW, and obtains similarity measurement values ​​between curves according to the vegetation growing season in a targeted manner; it uses the improved logistic weight function MLWF, which is flexible in setting weight limits, for weighting: （2）； Where i=2,3,…,m represents the row index of the matrix, j=2,3,…,n represents the column index of the matrix; D(i,j) is the minimum cumulative value of the path; Refers to the vegetation growing season. is the Euclidean distance between two values ​​in the matrix composed of two time series curves; is the weight between the matching points of the two time series curves; The weight function is as follows: （3）； Where i=1,2,…m, m is the length of the sequence, is the midpoint of the sequence; is the expected upper limit of the weight parameter, g is the empirical constant that controls the curvature of the function, and the range of g is 0.01~0.

6. The larger the g, the greater the weight. Step 7: Use the KNN algorithm to classify the attribute curves of the objects to be classified according to the attribute curves of the training samples.

2. A classification method based on weighted vegetation time series attribute characteristic curve according to claim 1, characterized in that: In step 2, an object-oriented multi-scale segmentation method is used to merge pixels from bottom to top according to the shape, texture and spatial position relationship characteristics of the target object, so that the above characteristics show consistency or similarity within the same object. The optimal segmentation scale is obtained through multiple experiments combined with visual interpretation of the segmentation results.

3. The classification method based on weighted vegetation time series attribute characteristic curve according to claim 1 is characterized in that: In step 3, the polarization vegetation index RVI formula is as follows: （1）； In the above formula and are the backscattering coefficients of polarization image data, respectively.

4. The classification method based on weighted vegetation time series attribute characteristic curve according to claim 1 is characterized in that: In the step 5, the extracted polarization vegetation index time series broken line is smoothed by using the difference method and the SG filter to obtain a time series fitting curve.

5. The classification method based on weighted vegetation time series attribute characteristic curve according to claim 1 is characterized in that: In step 7, the Euclidean distance between the object to be classified and the training sample is calculated according to the transformed attribute time series obtained in the previous step, and then the nearest neighbor classification KNN algorithm is used for classification; wherein the Euclidean distance is as follows: （4）； Where X and Y are the training sample object curve and the sample object curve to be classified respectively; n refers to the corresponding points randomly selected on the two curves; is the value of the i-th point on the X curve; is the value of the i-th point on the Y curve; Find the K training samples that are the nearest neighbors to the object to be classified. The object to be classified is assigned to the category with the largest number of these K training samples.

Images