Wetland vegetation automatic classification method based on time sequence remote sensing data segmentation

Abstract

The invention discloses a wetland vegetation automatic classification method based on time sequence remote sensing data segmentation. Comprising the following steps: step 1, performing Savitzky-Golay filtering preprocessing on NDVI data of a time sequence, and reconstructing NDVI data of a dense time sequence; step 2, based on a vegetation habitat similarity principle, using an X-means algorithm to segment NDVI dense time sequence data; and step 3, based on sample plot survey data and literature data, determining a mapping relation between the classification of the segmented plaques and the classification of the actual vegetation clusters, and realizing classification and precision evaluation of the vegetation clusters. According to the method, automatic classification on the vegetation cluster level can be achieved, the precision of vegetation classification based on medium-resolution optical data is improved, the defect that priori knowledge is insufficient in wetland vegetation remote sensing classification is overcome, and a new technical means is provided for large-scale vegetation mapping.

Description

Technical Field

[0001] The invention relates to the field of remote sensing and geographic information, and is an automatic classification method for wetland vegetation based on time series remote sensing data segmentation. Background Art

[0002] Wetland vegetation is the producer of wetland ecosystems and the main source of energy for the growth and metabolism of other biological groups in wetlands. The type and spatial distribution of wetland vegetation is an intuitive basis for judging the hydrological conditions and soil salinity of wetlands, and is also an important indicator of wetland biodiversity. The high intra-class variability and low inter-class variability make remote sensing classification of wetland vegetation challenging. The results of remote sensing classification of wetland vegetation cannot better serve ecological research, and it is difficult to provide more detailed information for refined wetland management and protection needs.

[0003] The spectral difference of vegetation is the theoretical basis for vegetation classification and mapping, but the phenomenon of "same object, different spectrum" and "same spectrum, different objects" of ground objects and the limited number of bands of multispectral images lead to low classification accuracy between different vegetation species. With the ability to reflect the subtle differences in the spectrum of ground objects, hyperspectral images have shown significant advantages in vegetation classification and mapping. For example, researchers use hyperspectral data such as AVIRIS and HyMap, combined with OBIA, SVM or RF classifiers to carry out salt marsh vegetation mapping. The results show that hyperspectral data with fine spatial resolution has great potential in vegetation mapping of heterogeneous wetland ecosystems. The combination of hyperspectral data and unsupervised classification has been widely used in target detection and land cover classification research, and can fully explore the hidden patterns in spectral information. The study of Belluco et al. shows that when there is less ground reference data, the use of K-means clustering method can provide more robust and reliable classification results than machine learning methods because it can utilize all available spectral information and reduce the Hughes effect. However, the high acquisition cost and complex processing technology of hyperspectral data limit its application in wetland vegetation remote sensing classification research.

[0004] In the temporal dimension, vegetation will show obvious dynamics with seasonal changes throughout the year, that is, the phenology of vegetation. Based on the phenological characteristics of vegetation (germination, maturity, aging, etc.), the remote sensing classification of wetland vegetation using time series optical remote sensing data has achieved good results, such as distinguishing reeds, Spartina alterniflora and Suaeda salsa. With the development of remote sensing cloud platforms such as Google EarthEngine, researchers can more conveniently access long-term series of remote sensing images to carry out wetland mapping research, synthesize a year's high-resolution remote sensing images into time series according to months or quarters, extract features such as indexes and quantiles, and use machine learning methods to achieve classification.

[0005] However, there are some problems with the existing time series classification methods: first, a large number of studies synthesize remote sensing images at a certain time interval to ensure that effective observation pixels can cover the entire study area, but this method will cover the observation information at a specific time point; second, due to the influence of surface emergencies (such as floods) and sensor noise, outliers often appear in time series data; third, due to the complexity of wetland environmental conditions and the volatility of hydrological conditions, the diversity and complex spatial composition and evolution characteristics of wetland vegetation are caused. The supervised classification based on prior knowledge is difficult to well reflect the complex spatial structure and composition of wetland vegetation. For example, in the area where Tamarix, Phragmites australis and Suaeda salsa grow alternately, due to different dominant species, different structures such as Tamarix & Phragmites australis associations, Phragmites australis & Suaeda salsa associations, and Tamarix & Phragmites australis & Suaeda salsa associations are formed. Summary of the invention

[0006] In order to address the deficiencies of the above-mentioned prior art and the need for scientific research and management of rapid acquisition of wetland vegetation cover category information, the purpose of the present invention is to provide an automated remote sensing classification method that can finely map wetland vegetation cover types on a small scale.

[0007] The technical solution of the present invention is: a method for automatic classification of wetland vegetation based on time series remote sensing data segmentation, comprising the following steps:

[0008] Step 1: Select appropriate parameters, perform Savitzky-Golay filter preprocessing on the time series data, and establish a complete dense time series NDVI data;

[0009] Step 2: Based on the similarity of vegetation habitats, K-means and X-means algorithms are used to segment the dense time series data NDVI;

[0010] Step 3: Establish a mapping relationship between the ecological site survey data and the segmentation patches to complete the classification and accuracy assessment.

[0011] The present invention has the following beneficial effects:

[0012] Accurate spatial distribution information of wetland vegetation is not only the basis for assessing the health and diversity of wetland ecosystems, but also a prerequisite for monitoring wetland environmental changes and predicting future ecological trends. In this invention, a new working framework is developed. Based on the principle of temporal similarity of similar vegetation coverage, the NDVI dense time series data is segmented using the X-means algorithm to obtain homogeneous vegetation patches, and the survey data is used to determine the correspondence between the patch category and the actual vegetation cluster category. The segmentation results are mapped to the classification system, and finally the mapping of wetland vegetation cluster categories is completed. Within this framework, the complexity of the vegetation community structure in the study area is taken into account, and the separability of vegetation categories based on NDVI dense time series data is improved through unsupervised segmentation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The figure is a flow chart of the method of the present invention.

[0014] Figure 2 This is an image example of the classification system of the present invention.

[0015] Figure 3 This is an example of the classification result of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0017] The present invention discloses an automatic classification method for wetland vegetation based on time series remote sensing data segmentation. This classification method is based on the Chinese vegetation classification method based on "plant community science-ecology", makes full use of the time dimension information of dense time series remote sensing data, and can realize the fine mapping of wetland vegetation categories at the cluster level.

[0018] like Figure 1 FIG. 1 is a flow chart of a wetland vegetation automatic classification method based on time series remote sensing data segmentation of the present invention, comprising the following steps:

[0019] Step 1: Preprocess the Sentinel-2 time series data, fill in missing values, and perform Savitzky-Golay filtering to process outliers; specifically:

[0020] Step 1.1, obtain Sentinel-2 images, remove clouds and crop them, and calculate the NDVI of each image; NDVI is the normalized vegetation index, NDVI=(B8-B4) / (B8+B4), where B4 and B8 are the red and near-infrared bands of Sentinel-2;

[0021] Step 1.2: Use linear time interpolation to fill in the missing values ​​in NDVI and construct a 5-day NDVI time series. The formula for linear time interpolation is:

[0022] ;

[0023] Among them, y is the missing value to be interpolated, y1 is the data before the missing value, y2 is the data after the missing value, and t, t1, and t2 are the corresponding times. After interpolation, the spatiotemporally continuous NDVI time series data will be obtained;

[0024] Step 1.3: Use Savitzky-Golay filtering to smooth the NDVI time series and process outliers. The key to applying Savitzky-Golay filtering is to solve the matrix operator. Suppose the filter window width is 2d=2m+1, and each measurement point is x= (-m, -m+1,...-1, 0, 1,...m-1, m). Use a k-order polynomial to fit the data points in the window:

[0025] ;

[0026] So there are n such equations, a0, a1... are the unknowns that need to be solved, forming a k+1 linear equation system, namely:

[0027] ;

[0028] Among them, Y is the measurement value matrix, X is the measurement point matrix, A is the matrix of unknown number a, E is the unit matrix, and the subscripts of each parameter represent their respective dimensions. If the system of equations is solvable, 2m+1 should be greater than k+1. By using the least squares method, the fitting parameters can be obtained. The solution is:

[0029] ;

[0030] Then the filtering value P of the model is:

[0031] ;

[0032] Among them, B is the relationship matrix between the filtered value and the observed value, then:

[0033] ;

[0034] The B matrix can be used to convert the observed values ​​into filtered values, and the NDVI dense time series data can be obtained;

[0035] The Savitzky-Golay filter parameters in step 1 include:

[0036] The sliding window half-width d represents 1 / 2 of the time range for applying the Savitzky-Golay filter. A larger d value means that the polynomial fitting is performed over a longer time range, which will produce a smoother effect.

[0037] When fitting the polynomial order k, a smaller k value will produce smoother results, but will mask some change information. A larger k value will retain more original information, but the smoothing effect on noise information will also become worse.

[0038] It is recommended to set k in the range of 2-4 and d between 4-7 time points, so there are 12 combinations of k and d. The root mean square error (RMSE) is a common indicator for evaluating the fitting effect of Savitzky-Golay filter, also known as SG filter. The formula is:

[0039] ;

[0040] Where n is the number of fitting points, y i is the ith observation, is the i-th fitting value. It is necessary to use the field measurement points to calculate the RMSE values ​​before and after filtering with different k and d combinations to obtain the optimal filtering parameters.

[0041] Step 2: Two-step segmentation based on NDVI dense time series, including:

[0042] Step 2.1: Use the K-means algorithm to segment the NDVI dense time series. The K-means algorithm divides the n objects in the data set into K clusters, so that the sum of the distances from each object to the center (or mean point, centroid) of the cluster to which it belongs is minimized. Set the number of categories to 100, and manually identify non-vegetation categories based on the survey data and the NDVI time curve of each category, exclude these categories, and retain the vegetation coverage area;

[0043] Step 2.2: In the vegetation coverage area, use the X-means algorithm to further automatically segment the NDVI dense time series. The X-means algorithm is an improvement on the K-means algorithm, which solves the problems of high computational complexity, the need to preset the K value, and the tendency to fall into local optimality. The X-means algorithm can judge the quality of the model under different numbers of segmentation based on the Bayesian Information Criterion (BIC), thereby automatically determining the optimal number of segmentation;

[0044]

[0045] r is the number of model parameters, n is the number of samples, and L is the likelihood function. The larger L is, the smaller the BIC value is, so the larger L is, the better.

[0046] Step 2.3: Determine the classification system based on the survey data and references. The construction of the classification system requires comprehensive consideration of the actual distribution and community structure of wetland vegetation in the target area, as well as the spatial scale of remote sensing observations. Therefore, at the spatial resolution of 10m in Sentinel-2 data, trees, shrubs, and tall grass vegetation can be included in the classification system. Map the segmented categories to the classification system to complete the classification mapping.

[0047] Step 3: Evaluate the accuracy of the classification results, including: independently acquiring samples of different wetland types, and evaluating the accuracy of the classification results, including overall accuracy, mapping accuracy, user accuracy, Kappa coefficient, and Macro-F1.

[0048] According to one embodiment of the present invention, the classification system is designed as follows:

[0049] In a certain area, based on field surveys and literature research, the main vegetation species and their distribution in the area were obtained. The image examples and sample numbers are shown in Table 1.

[0050] Table 1. Classification system of wetland vegetation in the Yellow River Delta region

[0051]

[0052] According to an embodiment of the present invention, the classification result is as follows: Figure 3 shown.

[0053] According to an embodiment of the present invention, the classification accuracy evaluation is specifically as follows:

[0054] The final classification results were generated based on the survey data through category mapping classification. The classification results included 9 vegetation cover types: reed community, reed community, sedge community, locust community, elm community, reed-sedge community, tamarisk-reed community, willow-reed community, and tamarisk-reed-sedge community. Based on the classification map, the confusion matrix was constructed using the validation samples to evaluate the classification results.

[0055] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments, for those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A wetland vegetation automatic classification method based on time series remote sensing data segmentation, characterized in that: The steps include: Step 1: Select appropriate parameters, perform Savitzky-Golay filter preprocessing on the time series data, and establish a complete dense time series NDVI data; Step 2: Based on the similarity of vegetation habitats, the K-means and X-means algorithms are used to segment the dense time series data NDVI to obtain segmentation patches of different categories; Step 3: Establish a mapping relationship between the ecological site survey data and the segmentation patches of different categories to complete the classification and accuracy assessment.

2. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1 is characterized in that: The step 1 specifically includes: Step 1.1: Acquire Sentinel-2 images. The time range is from December 1st of the year before the target year to January 31st of the year after the target year, i.e. 14 months. Perform cloud removal and cropping preprocessing operations on each original image. The original NDVI time series is calculated using the formula NDVI = (B8-B4) / (B8+B4), where B4 and B8 are the red and near-infrared bands of Sentinel-2. Step 1.2: Perform linear time interpolation on the original NDVI time series to fill in the image gaps caused by cloud removal and obtain complete NDVI time series data that is continuous in time and space; Step 1.3: Apply Savitzky-Golay filtering to the interpolated NDVI time series data.

3. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 2 is characterized in that: The formula for linear time interpolation is: ； Among them, y is the missing value that needs to be interpolated, y1 is the data before the missing value, y2 is the data after the missing value, and t, t1, and t2 are the corresponding times.

4. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 2 is characterized in that: Specifically, the Savitzky-Golay filter is as follows: assuming that the filter window width is 2d=2m+1, each measurement point is x = (-m, -m+1, ... -1, 0, 1, ... m-1, m), and a k-order polynomial is used to fit the data points in the window: ； We get n such equations, where a0, a1, etc. are the unknowns to be solved, forming a k+1 linear equation system, namely: ； Among them, Y is the measurement value matrix, X is the measurement point matrix, A is the matrix of unknown number a, E is the unit matrix, and the subscripts of each parameter represent their respective dimensions; the value 2m+1 is greater than k+1, and the fitting parameters are obtained by the least squares method. The solution is: ； Then the filtering value P is: ； Among them, B is the relationship matrix between the filtered value and the observed value, then: 。 5. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: Use the K-means algorithm to segment the NDVI time series of the target area. Set the number of segments to a larger value. Determine whether it is vegetation based on the time series curve of each segmentation category. Keep the vegetation area and remove the non-vegetation area. Step 2.2: In the vegetation area, use the X-means algorithm to automatically segment the NDVI time series and automatically determine the optimal number of segmentation categories.

6. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1 is characterized in that: The step 3 specifically includes: Step 3: Determine the vegetation classification system of the target area based on the field survey data and literature data, and map the different categories of patches generated by segmentation to the classification system to complete the classification.

7. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1 is characterized in that: The Savitzky-Golay filter parameters in step 1 include: Sliding window half width d; Fitting polynomial order k; Set k in the range of 2-4 and d between 4-7 time points. Therefore, there are 12 combinations of k and d. The root mean square error (RMSE) is used to evaluate the Savitzky-Golay filter fitting effect. The formula is: ； Where n is the number of fitting points, y i is the ith observation, is the i-th fitting value; using the field measurement points, the RMSE values ​​before and after filtering with different k and d combinations are calculated to obtain the optimal filtering parameters.

8. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1 is characterized in that: Based on the NDVI time curves of the survey sampling points, the NDVI thresholds for vegetation and non-vegetation were determined. Areas with the highest NDVI throughout the year below the threshold were classified as non-vegetation categories, and these categories were excluded to retain vegetation-covered areas.

9. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1, characterized in that: The wetland classification method based on time series remote sensing data segmentation is based on the change of NDVI of wetland vegetation in a year. The X-means algorithm aggregates pixels with similar change characteristics into one category based on the principle of temporal similarity.

10. The method for automatic classification of wetland vegetation based on time series remote sensing data segmentation according to claim 1, characterized in that: The step 3 specifically includes: The above-mentioned different wetland vegetation type samples were obtained independently, and the accuracy of the classification results was evaluated, including overall accuracy, mapping accuracy, user accuracy, Kappa coefficient and Macro-F1.

Images