Method for identifying ground object types by integrating time-series remote sensing information and KL-divergence

Abstract

The application provides a ground object type recognition method combining time sequence remote sensing information and KL-divergence, sample data of various ground objects in a research area is extracted, and time sequence remote sensing parameters of the ground object types are selected; the sample points are used to extract corresponding parameter values from the time sequence remote sensing parameters to form standard sequence curves of each type of ground object; the time sequence remote sensing parameter values corresponding to the pixels to be classified are used as another distribution data, which is combined with the distributions corresponding to the standard sequence curves of various ground objects, and the KL values of the pixels of the ground objects to be classified are calculated based on a KL-divergence formula to form a KL layer; the n KL values corresponding to each pixel are compared, and the pixel is classified into a ground type category corresponding to the minimum KL value. The application fully utilizes the change characteristics of the ground objects on the sequence, and combines the KL-divergence which has obvious advantages in measuring the similarity of probability distribution, so that the ground object types can be better classified and recognized, and the recognition precision is improved.

Description

Technical Field

[0001] This invention relates to the technical field of remote sensing monitoring, and in particular to a method for identifying land cover types by integrating temporal remote sensing information and KL-divergence. Background Technology

[0002] Currently, the main methods for identifying land cover types using remote sensing technology are methods based on spectral information and methods based on time-varying characteristics.

[0003] Spectral-based identification methods are primarily used for medium- to high-resolution optical imagery. The principle is to utilize the statistical characteristics of pixel values ​​to determine the similarity between the pixel to be identified and the sample. If medium- to high-resolution remote sensing data covering the study area can be acquired within an ideal timeframe, high-precision classification results can be obtained. However, the acquisition of optical imagery is severely affected by weather conditions and revisit cycles. Under conditions such as dense cloud cover, even with satellite transit, remote sensing data cannot be acquired. Therefore, even with a remote sensing satellite constellation, it is difficult to guarantee the acquisition of remote sensing images covering the entire study area at a specific time.

[0004] Recognition methods based on temporal variation features are mainly used for time-series remote sensing images and products. Their main principle is to identify different land cover types by utilizing the differences in their temporal variation patterns. However, this type of remote sensing data typically has low spatial resolution, resulting in low final recognition accuracy. Therefore, many researchers have begun to combine spectral and temporal features in hybrid pixel decomposition studies. However, these methods usually yield abundance maps, which, while improving quantitative accuracy, cannot specifically describe the spatial distribution of land cover within a pixel, thus remaining a significant inconvenience in application.

[0005] Kullback–Leibler (KL) divergence, a metric in probability and information theory used to measure the difference between two probability distributions, has proven highly effective in remote sensing image classification. For example, combining KL divergence with temporal variation information in winter wheat identification studies has yielded higher accuracy compared to conventional methods due to the significant advantage of KL divergence in measuring the differences between different distributions. However, a drawback is that it only supports identification of a single crop type and does not consider the simultaneous identification of multiple crop types or land cover types.

[0006] In existing technologies, using remote sensing imagery to monitor land cover types is a relatively quick and intuitive method. However, each technique has its own inherent advantages and disadvantages. For example, conventional classification methods, especially parametric methods, typically assume that land cover features exhibit a normal distribution, which is unrealistic for high-dimensional data such as time-series data. Furthermore, common similarity matching classification methods place too much emphasis on the morphological characteristics of curves, making them relatively more complex. KL divergence, on the other hand, does not focus on whether there is a normal distribution or morphological characteristics; it uses information theory and probability theory to determine the "distance" between the pixel to be classified and the true distribution, exhibiting strong sensitivity. Therefore, introducing new data methods can obtain accurate monitoring data. However, in current technologies, there is a lack of research on the identification of multiple land cover types by integrating time-series remote sensing information and KL divergence. Summary of the Invention

[0007] To address the technical problems of existing conventional spectral-based remote sensing identification methods, which require specific assumptions and time-series matching methods that overemphasize morphological features, resulting in inconvenient applications, this invention proposes a land cover type identification method that integrates time-series remote sensing information and KL-divergence. This method uses time-series remote sensing information combined with KL divergence for land cover type classification and identification, without any mandatory assumptions. It uses probability theory and information theory to identify land cover types by judging the "distance" between two distributions. Any data that can form a parameter sequence reflecting land cover characteristics (such as time-series parameters reflecting land cover growth characteristics) can be applied, making it more convenient and flexible than conventional methods.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0009] This invention utilizes time-series remote sensing data to generate time-series remote sensing information, such as time-series vegetation indices; selects samples for each land cover category to be identified, requiring them to be pure pixel samples at that scale; constructs a standard time-series curve for each land cover category using the mean of the sample pixels; calculates the KL value for each land cover category based on the standard time-series curve using the KL divergence formula; finally, compares the KL value of each pixel with respect to the land cover to be identified, and assigns the pixel to the category with the smallest KL value, thereby obtaining the classification and identification result.

[0010] A method for identifying land cover types by integrating temporal remote sensing information and KL-divergence, the steps of which are as follows:

[0011] Step 1: Identify samples of various land cover types in the study area and select sequential remote sensing parameters that have identification characteristics for n types of land cover.

[0012] Step 2: Use sample points to extract corresponding parameter values ​​from the sequence remote sensing parameters to form standard sequence curves for each type of land cover;

[0013] Step 3: Using the remote sensing parameter values ​​of the sequence corresponding to the pixel to be classified as another distribution data, and combining it with the distribution corresponding to the standard sequence curves of various land cover types, calculate the KL value of each land cover type corresponding to the pixel to be classified based on the KL-divergence formula, and extend it to the entire study area to form n KL layers.

[0014] Step 4: Compare the n KL values ​​corresponding to each pixel and assign the pixel to the land use category corresponding to the smallest KL value; apply this to all pixels to obtain the classification results for the study area.

[0015] Preferably, the sequence remote sensing parameters are time-series remote sensing parameters characterizing the temporal differences in vegetation type growth and development or sequence data formed from hyperspectral data reflecting the differences in spectral curves of ground objects.

[0016] Preferably, the time series remote sensing parameters are formed using MODIS NDVI data, and the sequence data is formed by adding blue band, red band and near-infrared band reflectance to MODIS NDVI.

[0017] Preferably, the method for implementing the standard sequence curve in step two is as follows:

[0018] A. Based on field survey data or research experience, select pure pixel samples for the land cover types to be classified;

[0019] B. Based on the selected samples, extract the corresponding parameter values ​​from the sequence remote sensing parameters;

[0020] C. The extracted parameter values ​​are averaged according to the land cover category, thus forming a standard sequence curve for each land cover category.

[0021] Preferably, the types of land features include evergreen forests, deciduous forests, closed grasslands, open grasslands, arbor savanna, savanna, grasslands, farmland, water bodies, and bare land; the distinct growth and change characteristics of different vegetation types are manifested in the variation patterns of peaks and valleys of the standard sequence curves and the different magnitudes of NDVI values.

[0022] Preferably, the method for calculating the KL value is as follows:

[0023] The formula for calculating KL-divergence is:

[0024]

[0025] In the formula, P represents the true distribution sequence, i.e., the standard sequence curve of each class; Q is the data sequence corresponding to a pixel to be classified; i represents the sequence number in the two distributions; and n represents the n land categories in the classification system.

[0026] KL values ​​are represented by the mean of the positive and negative KL-divergence:

[0027] KL = (D KL (P||Q)+D KL (Q||P)) / 2;

[0028] In the formula, KL is the final KL value, and D KL (P||Q) and D KL (Q||P) represent positive KL-divergence and negative KL-divergence, respectively.

[0029] Preferably, the KL value for each class is a layer or band with the same number of rows and columns as the original remote sensing data. The KL values ​​of all pixels of the land cover to be classified are calculated to obtain a KL data layer corresponding to the land cover of that class, i.e., the KL layer.

[0030] Preferably, for each pixel, the size of its corresponding n KL values ​​is compared, and the pixel is assigned to the land cover category corresponding to the smallest KL value.

[0031] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention fully utilizes the sequential variation characteristics of ground features, such as time series or spectral series, and closely combines them with KL-divergence, which has a significant advantage in measuring the similarity of probability distributions. This allows for better classification and identification of ground feature types, improving identification accuracy. This invention introduces KL-divergence into remote sensing classification research, providing a new research perspective for remote sensing classification work. It is applicable to remote sensing data that reflects the characteristics of ground feature types and forms sequences at different scales, and is easy to apply in practical monitoring at certain regional scales, thus possessing good potential for widespread application. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the process of the present invention.

[0034] Figure 2 This is a map showing sample data for the study area.

[0035] Figure 3 The standard time series curves for ten types of land features in the study area are shown.

[0036] Figure 4 This is a schematic diagram of the classification method of the present invention.

[0037] Figure 5 This is a schematic diagram of the classification results of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] like Figure 1 As shown, a land cover type identification method integrating time-series remote sensing information and KL-divergence is proposed. First, based on field surveys or research experience, samples are selected for each land cover type using the available remote sensing imagery. Second, the corresponding parameters required for classification, such as NDVI, are calculated from the time-series remote sensing data, and these parameters also form a time series. Third, using samples from each land cover type, corresponding parameters are extracted from the time-series remote sensing parameter dataset. By averaging the sample parameters for each land cover type, a standard time-series curve for each land cover type is obtained. Fourth, the standard time-series curve is used as one distribution, and the time-series data corresponding to each pixel in the time-series parameters is used as another distribution. The KL value is calculated for each pixel using the KL-divergence formula, thus generating a band of data for each land cover type. Fifth, the generated sequence KL values ​​are determined pixel by pixel. The KL values ​​for each pixel corresponding to each land cover type are compared, and the pixel is assigned to the land cover type with the smallest KL value. After determining all pixels, the classification and identification results are obtained. The specific steps of this invention are as follows:

[0040] Step 1: Using field surveys and experience, select samples of various land features in the study area, and select sequential remote sensing parameters that have identification characteristics for n types of land features.

[0041] Based on the characteristics that can distinguish land cover types, a sequence of remote sensing parameters is prepared to be formed. For example, time series vegetation index data, which can identify land cover types based on the temporal differences in vegetation growth and development; or sequence data formed from hyperspectral data, which can reflect the differences in the spectral curves of land cover types, thereby identifying land cover types.

[0042] Samples are used to identify land cover types. Field surveys determine the land cover type corresponding to each sample point. Data from these sample points is then extracted from the sequence remote sensing parameters as the basis for classification. The classification system consists of the 10 land cover types mentioned below. Specific processing methods include synthesizing these parameter bands and then extracting parameters using the sample points, or direct extraction.

[0043] For example, MODIS NDVI data can be used to form time-series remote sensing parameters that reflect the temporal variation patterns of different land cover types. Specifically, in this embodiment, the present invention uses MODIS NDVI data plus blue, red, and near-infrared reflectance to form a sequence of data for classification experiments. This sequence of data integrates temporal variation characteristics and spectral characteristics, enabling better identification of land covers.

[0044] Step 2: Extract remote sensing parameter data from the sequence remote sensing parameters using samples to form standard sequence curves for each type of land cover.

[0045] The method for generating standard sequence curves is as follows:

[0046] A. Based on field survey data or research experience, select pure pixel samples for the land cover types to be classified.

[0047] A pure pixel refers to a pixel that corresponds to a single type of land cover. If a pixel contains more than one land cover type, it is considered a mixed pixel. Selecting a pure pixel ensures that the sample corresponds to only one type of land cover.

[0048] B. Based on the selected samples, extract the corresponding parameter values ​​from the sequence remote sensing parameter dataset.

[0049] C. Calculate the mean of the extracted data according to the land cover category, thereby forming a standard sequence curve for each land cover category.

[0050] Calculating the mean of remote sensing parameters for each land cover type can reflect the representativeness of that type of land cover at the parameter classification center, which is more representative than a single sample data. Each land cover type corresponds to a standard curve, which can be considered as the true distribution of that land cover type. Strictly speaking, it can be called a "quasi-true distribution".

[0051] Specifically, in this embodiment, the present invention utilizes MODIS NDVI time-series data, based on field surveys and remote sensing imagery, to extract 1775 samples from 10 land cover categories (1. evergreen forest, 2. deciduous forest, 3. closed grassland, 4. open grassland, 5. arborized savanna, 6. savanna, 7. grassland, 8. farmland, 9. water body, 10. bare land). Figure 2As shown. Approximately half of the samples (882) were used for classification, and the rest were used for validation. Time-series MODIS NDVI data were extracted from the land cover classification samples, and the mean was calculated to obtain the standard NDVI curve for each land cover type, as shown below. Figure 3 As shown.

[0052] Depend on Figure 3 It can be seen that different vegetation types exhibit distinct growth and change characteristics, which forms the basis for distinguishing land cover types. This is reflected in the variation patterns of the peaks and troughs of the curves and the differences in NDVI values.

[0053] Step 3: Using the remote sensing parameter value of the sequence corresponding to a pixel to be classified as another distribution data, and combining it with the distribution corresponding to the standard sequence curve of various land cover types, calculate the KL value of the pixel to be classified based on the KL-divergence formula, and extend it to the entire study area to form n KL data layers.

[0054] The specific formula for calculating KL-divergence is as follows:

[0055]

[0056] In the formula, P represents the true distribution sequence, i.e., the standard sequence curve for each class; Q is the data sequence corresponding to a pixel to be classified; i represents the sequence number in the two distributions. || is the symbol for conditional probability. D KL (P||Q) is a representation of conditional probability, where P(i) and Q(i) are the i-th parameter values ​​in the sequence parameters, and n represents the number of land types.

[0057] Because KL-divergence is asymmetric, i.e., D KL (P||Q)≠D KL (Q||P), typically represented by the mean of the positive and negative KL-divergence values. The calculation formula is as follows:

[0058] KL = (D KL (P||Q)+D KL (Q||P)) / 2 (2)

[0059] In the formula, KL is the final KL value, which is the positive KL-divergence D. KL (P||Q) and reverse KL-divergenceD KL The mean of (Q||P).

[0060] The KL value is calculated by using the standard curve for each class as one distribution and the sequence of each pixel in the sequenced remote sensing parameters as another distribution. The calculation result for each class is a layer or band with the same number of rows and columns as the original remote sensing data. Assuming there are n land classes to be classified, n layers will be formed.

[0061] Each standard sequence curve can be seen as a real discrete distribution of that type of land cover, while a set of time series parameters corresponding to a pixel to be classified can be seen as another distribution. The KL value between the pixel and the standard curve can be calculated using formula (2). By calculating the KL value of all pixels of the land cover to be classified in this way, a KL data layer corresponding to that type of land cover can be obtained.

[0062] Since each cell needs to be compared with the standard curves of all land cover types, 10 KL data layers can be obtained in this embodiment, each corresponding to a land cover type.

[0063] Step 4: Compare the n KL values ​​corresponding to each pixel, assign the pixel to the land use category corresponding to the smallest KL value, and apply this method to all pixels to obtain the classification result.

[0064] Based on the calculated KL values, for each pixel, its corresponding n KL values ​​are compared, i.e., compared with the KL values ​​formed by the standard sequence curves of each land cover category. The pixel is then assigned to the land cover category corresponding to the smallest KL value. Since KL-divergence can measure the distance or proximity between two discrete distributions, by comparing the KL values ​​of each land cover category corresponding to each pixel to be classified, the minimum value represents the pixel being closest to that land cover category, thus assigning the pixel to that category. This process is repeated for each pixel to obtain the classification results for the entire study area. The classification process is as follows: Figure 4 As shown.

[0065] Specifically, in this embodiment, the classification results are as follows: Figure 5 As shown, the distribution of different land cover types within the study area can be observed. To verify the accuracy of the land cover remote sensing classification method provided by this invention, the classification results were analyzed using another half of the sample points based on the confusion matrix method. Overall, the overall classification accuracy reached 85.12%, and the KAPPA coefficient was 0.84.

[0066] The verification results show that the method provided by this invention achieves a high level of accuracy in land cover classification, accurately reflecting the true distribution pattern of land cover. Especially in large-scale studies conducted across the entire country, the accuracy is significantly improved compared to less than 80% under the conditions of conventional methods.

[0067] In summary, this invention introduces the KL-divergence method, a metric for measuring the distance between two distributions from probability theory and information theory, into remote sensing classification. By analyzing the "distance" between the pixel to be classified and the true distribution, it identifies land cover types. Simultaneously, it overcomes the assumptions of traditional remote sensing classification methods, such as the requirement for land cover to have a normal distribution, and the overemphasis on curve morphology, achieving higher accuracy. Therefore, this invention has significant application value in land cover remote sensing classification, crop type identification, and hyperspectral remote sensing classification, while also providing a new perspective and reference for remote sensing classification work. The data used in this invention can be time-series remote sensing data, hyperspectral bands, various indices, and combinations thereof, offering flexibility and ease of practical monitoring applications at the regional scale, thus possessing significant potential for widespread application.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying land cover types by integrating temporal remote sensing information and KL-divergence, characterized in that, The steps are as follows: Step 1: Identify samples of various land cover types in the study area and select sequential remote sensing parameters that have identification characteristics for n types of land cover. Step 2: Use sample points to extract corresponding parameter values ​​from the sequence remote sensing parameters to form standard sequence curves for each type of land cover; Step 3: Using the remote sensing parameter values ​​of the sequence corresponding to the pixel to be classified as another distribution data, and combining it with the distribution corresponding to the standard sequence curves of various land cover types, calculate the KL value of each land cover type corresponding to the pixel to be classified based on the KL-divergence formula, and extend it to the entire study area to form n KL layers. Step 4: Compare the n KL values ​​corresponding to each pixel and assign the pixel to the land use category corresponding to the smallest KL value; The classification results for the study area are obtained by applying the method to all pixels. KL values ​​are represented by the mean of the positive and negative KL-divergence: ； In the formula, KL For the final KL value, and These are positive KL-divergence and negative KL-divergence, respectively. The KL value for each class is a layer or band with the same number of rows and columns as the original remote sensing data. The KL values ​​of all pixels of the land cover to be classified are calculated to obtain a KL data layer corresponding to the land cover of that class, i.e., the KL layer. For each cell, compare its corresponding n KL values ​​and assign the cell to the land cover category corresponding to the smallest KL value.

2. The method for identifying land cover types by integrating temporal remote sensing information and KL-divergence according to claim 1, characterized in that, The sequence remote sensing parameters are time-series remote sensing parameters that characterize the temporal differences in the growth and development of vegetation types, or sequence data formed from hyperspectral data that reflect the differences in the spectral curves of ground objects.

3. The method for identifying land cover types by integrating temporal remote sensing information and KL-divergence according to claim 2, characterized in that, The time series remote sensing parameters are formed using MODIS NDVI data, and the sequence data is formed by adding blue band, red band and near-infrared band reflectance to MODIS NDVI.

4. The method for identifying land cover types by integrating temporal remote sensing information and KL-divergence according to any one of claims 1-3, characterized in that, The method for implementing the standard sequence curve in step two is as follows: A. Based on field survey data or research experience, select pure pixel samples for the land cover types to be classified; B. Based on the selected samples, extract the corresponding parameter values ​​from the sequence remote sensing parameters; C. The extracted parameter values ​​are averaged according to the land cover category, thus forming a standard sequence curve for each land cover category.

5. The method for identifying land cover types by integrating temporal remote sensing information and KL-divergence according to claim 4, characterized in that, The types of land features include evergreen forests, deciduous forests, closed grasslands, open grasslands, arbor savanna, savanna, grasslands, farmland, water bodies, and bare land; the distinct growth and change characteristics of different vegetation types are reflected in the peak and valley variation patterns of the standard sequence curves and the different NDVI values.

6. The method for identifying land cover types by integrating temporal remote sensing information and KL-divergence according to claim 5, characterized in that, The method for calculating the KL value is as follows: The formula for calculating KL-divergence is: ； In the formula, P The curve representing the true distribution sequence, i.e., the standard sequence curve for each class; Q A data sequence corresponding to a pixel to be classified; i Represents the sequence number in two distributions; n Representing the classification system n Land categories.

Images