Wetland landscape pattern classification method based on optimized feature parameters

Abstract

This invention discloses a wetland landscape pattern classification method based on optimized feature parameters. First, multispectral and topographic data of the target wetland are preprocessed to extract feature parameters from the preprocessed data. Wetland landscape categories are defined, identifying 13 land cover types. Training samples are determined by referencing field survey data, original multispectral image data, principal component analysis (PCA1) of original remote sensing data, and the tasseled cap transformation brightness component of original remote sensing data. For the preprocessed multispectral data, wetland information extraction is performed using both maximum likelihood classification and decision tree classification. The classification results from the two extractions are combined with the extracted feature parameters to form fused data. Information extraction based on maximum likelihood classification is then performed on the fused data to obtain dynamic change analysis results of wetland landscape classification. Combined with optimized feature parameters, multiple feature parameters are fused to achieve information extraction from remote sensing images and refined classification of wetland landscapes.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology, specifically relating to a wetland landscape classification method based on data fusion of multiple feature parameters. Background Technology

[0002] Land use and cover classification is one of the main methods for studying wetland evolution. With the increase in global satellite data, remote sensing big data is applied in many fields, and wetland observation can be achieved more effectively based on remote sensing imagery and field survey data. Using long-term remote sensing imagery data to classify wetland landscapes and conduct research on the evolution of wetland landscape patterns is of great research significance for accurately understanding and assessing wetland ecological health.

[0003] Based on existing research, deep learning methods (U-NET) can effectively identify vegetation and non-vegetation classifications, but cannot achieve three-level (13 categories) landscape classification. Currently, methods for land cover classification based on remote sensing imagery mainly include supervised classification (including object-oriented and pixel-based methods), decision tree methods, random forests, and neural network methods. Analysis revealed that the maximum likelihood method is the most commonly used supervised classification method. Through field verification points obtained from on-site surveys, the overall classification accuracy reached 75.29%, with a Kappa coefficient of 0.71. However, reeds were found in higher-altitude dry areas, indicating that some classification error still exists. Furthermore, the use of decision trees for classification requires the selection of appropriate parameters. Based on previous research results, six feature parameters were selected: Brightness, NDVI, MNDWI, NIR, SI, and DEM. The decision tree method was used to perform preliminary classification of wetland landscapes. The results showed that there were still problems in distinguishing between forest land and dry land, rice paddies and reeds, and bare land and residential areas. Moreover, the classification effect for 13 landscapes was not achieved. For example, the decision tree identified water bodies, but it could not classify them into rivers, reservoirs, and shallow water areas. Therefore, its classification accuracy was relatively low. The overall classification accuracy reached 56.82%, and the Kappa coefficient was 0.57.

[0004] To achieve a three-level classification of wetland landscapes with high accuracy, it is urgent to propose a wetland landscape classification method based on optimized feature parameters. Summary of the Invention

[0005] This invention addresses the shortcomings of existing wetland landscape classification methods, such as certain errors and the inability to achieve three-level landscape classification. It proposes a wetland landscape classification method based on data fusion of multiple feature parameters, which has high classification accuracy and can achieve three-level landscape classification.

[0006] This invention is achieved using the following technical solution: a wetland landscape pattern classification method based on optimized feature parameters, comprising the following steps:

[0007] Step A: Acquire the original image and preprocess the multispectral data and topographic data of the target wetland; the original image includes remote sensing images and DEM data of the target wetland from multiple years, and the preprocessing process includes, but is not limited to, radiometric correction, atmospheric correction and mosaicking and cropping operations to obtain the image data of the target wetland.

[0008] Step B: Extract the data feature parameters from the preprocessed data. The feature parameters include the first component of principal component analysis (PCA1), the tasseled cap transformation brightness component, normalized difference vegetation index (NDVI), normalized difference water index (MNDWI), red band spectral value (RED), near-infrared band spectral value (NIR), bare soil index (SI), urban building index (IBI), and digital elevation model (DEM).

[0009] Step C: Remote sensing image information extraction and classification:

[0010] (1) Sample selection: The wetland landscape category is defined by a three-level landscape classification, and the training samples are determined by referring to the field survey data, the first component of principal component analysis (PCA1) of the original multispectral image data, the tasseled cap transformation brightness component of the original remote sensing data;

[0011] The three-tiered landscape classification includes:

[0012] Primary classification: Natural wetlands, constructed wetlands, and non-wetlands;

[0013] Secondary classification: Natural wetlands are further subdivided into water bodies, marsh wetlands, mudflats, and bare land; Non-wetlands are further subdivided into woodlands, farmland, industrial and mining buildings, and residential areas;

[0014] The three-level classification system further subdivides water bodies into river waters and shallow sea waters, marshes into reed beds and Suaeda salsa areas, and artificial wetlands into salt pans, aquaculture areas, and reservoirs; and farmland into paddy fields and dry land.

[0015] (2) First maximum likelihood classification: For the multispectral data after preprocessing in step A, wetland information extraction based on maximum likelihood classification is performed for the first time, combined with the determined training samples.

[0016] (3) Decision tree classification: Based on the multispectral data preprocessed in step A, wetland information is extracted using decision tree classification in combination with the determined training samples.

[0017] (4) Band combination: The first maximum likelihood classification result is used as 1 data band, the decision tree classification result is used as 1 band, and combined with the 9 feature parameters extracted in step B to form 11 bands of fused data.

[0018] (5) Second maximum likelihood classification: Information extraction based on maximum likelihood classification is performed on the new fused data to obtain the analysis results of the dynamic changes in the target wetland landscape classification.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0020] This solution extracts feature parameters from multispectral data and improves the accuracy of wetland landscape classification by fusing and optimizing parameters. It also establishes a three-level wetland landscape classification system, breaking through the application of refined wetland landscape classification. By fusing optimized feature parameters and combining classifiers, it solves the problem of three-level landscape classification (13 classes) that deep learning and decision trees cannot achieve, and improves the accuracy of maximum likelihood classification.

[0021] Moreover, this solution leverages the advantages of machine learning technology in data processing to construct training samples for wetland landscape classification, find threshold settings for decision tree classification, and clarify optimal feature parameters. This enables the application of remote sensing data and machine learning technology in wetland landscape classification. Optimizing feature parameters can highlight and enhance wetland landscape information, and the combination of decision tree threshold setting and maximum likelihood method can compensate for each other's shortcomings.

[0022] This invention is pioneering and innovative, providing a scientific basis for effectively estimating the blue carbon reserves, ecological and economic value of wetlands, and for developing scientific, effective and accurate schemes for identifying and monitoring the spatial and ecological characteristics of wetlands. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the classification method described in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the Liaohe River Delta region image obtained by cropping an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram showing the distribution of verification points for remote sensing interpretation results in the Liaohe River Delta according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of decision tree parameter settings in an embodiment of the present invention for classification based on decision trees;

[0027] Figure 5 This is a schematic diagram comparing the classification results for 2007 using decision tree (a), maximum likelihood (b), and a combination of both (c) in an embodiment of the present invention.

[0028] Figure 6 The images show the landscape distribution of the Liaohe River Delta in 1987, 1997, 2007, and 2017, as per embodiments of the present invention. Detailed Implementation

[0029] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may be practiced in other ways than those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0030] This embodiment proposes a wetland landscape pattern classification method based on optimized feature parameters, such as... Figure 1 As shown, it includes the following steps:

[0031] Step A: Acquire the original image and preprocess the multispectral and topographic data of the target wetland;

[0032] Step B: Extract the data feature parameters from the preprocessed data. The feature parameters include the first component of principal component analysis (PCA1), the tasseled cap transformation brightness component, normalized difference vegetation index (NDVI), normalized difference water index (MNDWI), red band spectral value (RED), near-infrared band spectral value (NIR), bare soil index (SI), urban building index (IBI), and digital elevation model (DEM).

[0033] Step C: Remote sensing image information extraction and classification:

[0034] (1) Sample selection: The wetland landscape category is defined by a three-level landscape classification, and the training samples are determined by referring to the field survey data, the first component of principal component analysis (PCA1) of the original multispectral image data, the tasseled cap transformation brightness component of the original remote sensing data;

[0035] (2) First maximum likelihood classification: For the multispectral data after preprocessing in step A, wetland information extraction based on maximum likelihood classification is performed for the first time, combined with the determined training samples.

[0036] (3) Decision tree classification: Based on the multispectral data preprocessed in step A, wetland information is extracted using decision tree classification in combination with the determined training samples.

[0037] (4) Band combination: The first maximum likelihood classification result is used as 1 data band, the decision tree classification result is used as 1 band, and combined with the 9 feature parameters extracted in step B to form 11 bands of fused data.

[0038] (5) Second maximum likelihood classification: Information extraction based on maximum likelihood classification is performed on the new fused data to obtain the analysis results of the dynamic changes in the target wetland landscape classification.

[0039] To better understand the present invention, this embodiment performs a three-level classification of the Liaohe River Delta wetland landscape. The classification effect is verified by analyzing long-term remote sensing data to validate the dynamic changes of the Liaohe River Delta wetland, as detailed below:

[0040] Step 1: Preprocess the multispectral and topographic data of the target wetland;

[0041] The downloaded remote sensing images and DEM data from 2017, 2007, 1997, and 1987 underwent preprocessing including radiometric correction, atmospheric correction, and mosaicking / cropping, as detailed below:

[0042] (1) The data sources are 6 Landsat 5 TM images, 2 Landsat 8 OLI images, and 4 ASTER GDEM V2 images, all with a spatial resolution of 30 meters. Satellite images from September to October have virtually no cloud cover in the study area and are of good quality. Therefore, this invention selects 2 images from September 21, 1987, 2 images from October 18, 1997, 2 images from September 28, 2007, and 2 images from September 7, 2017. ASTER GDEM V2 global DEM data was officially released as global topographic elevation data on January 6, 2015. This digital elevation model of the Earth's land surface contains 1.3 million stereo images collected by the Advanced Spaceborne Thermal and Reflective Radiometer (ASTER) in Japan, with a horizontal accuracy of 30 meters.

[0043] Table 1 Remote sensing image information of the study area

[0044]

[0045] (2) Using the Radiometric Calibration module in ENVI software, the original image DN values ​​were converted into apparent radiance.

[0046] (3) Then, the apparent radiance data is converted into surface reflectance data using the FLAASH atmospheric correction module;

[0047] (4) Based on the stitched image map, and according to the existing vector region bounding map, the image data of the Liaohe River Delta region is obtained by cropping, such as... Figure 2 As shown.

[0048] Step 2: Extract the feature parameters of the preprocessed data (9 feature parameters);

[0049] The feature parameters extracted from the above 8 remote sensing images include Principal Component Analysis (PCA1) component 1, Tasselled Cap Transformation (TCT) brightness component, NDVI, MNDWI, Red Band Spectral Value (RED), Near Infrared Band Spectral Value (NIR), Soil Index (SI), Urban Building Index (IBI), and Digital Elevation Model (DEM).

[0050] Table 2. Statistics of Feature Parameters

[0051]

[0052]

[0053] Principal component analysis (PCA) is widely used in research related to vegetation cover, remote sensing monitoring of ecological changes, and land cover classification. It involves removing redundant (closely related) variables from all previously proposed variables and creating as few new variables as possible that are pairwise uncorrelated. These new variables should retain as much of the original information as possible in reflecting the topic. The contribution rates of the first principal component (PCA1) for 1987, 1997, 2007, and 2017 were 89.18%, 92.47%, 85.93%, and 83.53%, respectively, representing the vast majority of information for each band. The synthesized first principal component is used to replace the indices of each component. The PCA data in this example was converted from the PCA module of ENVI software.

[0054] Tasseled Cap Transform, also known as KT Transform or Kauth-Thomas Transform, is an empirical linear orthogonal transformation of images based on the information distribution structure of soil, vegetation, and other data in multidimensional spectral space from multispectral remote sensing. The first three components after the transformation are named "Brightness," "Greenness," and "Moisture," reflecting the soil and rock, vegetation, and moisture information within the soil and vegetation, respectively. In this embodiment, the tasseled cap analysis data was converted using the Tasseled Cap module of ENVI software.

[0055] Normalized Difference Vegetation Index (NDVI) is one of the most commonly used indices for indicating vegetation growth status. It is widely used in monitoring vegetation growth and ecological environment quality. The formula is:

[0056] NDVI=(long(nir)-ρ(red)) / (I(nir)+i(red)) (1)

[0057] The Modified Normalized Difference Water Index (MNDWI) is used to extract water body information from images. MNDWI can effectively distinguish between shadows and water bodies, solving the problem of shadow removal in water body extraction. Its formula is as follows:

[0058] MNDWI=(yin(green)-ρ(swir1)) / (ρ(green)+r(swir1)) (2)

[0059] Bare soil is a major surface landscape feature in areas prone to soil erosion. This data plays a crucial role in the monitoring and control of soil erosion. The formula for calculating the Bare Soil Index (SI) is as follows:

[0060] SI=[ρ(swir1)+ρ(red)-(ρ(bule)+ρ(nir))] / [ρ(swir1)+ρ(red)+(ρ(bule)+ρ(nir))] (3)

[0061] The formula for calculating the Urban Building Index (IBI) is as follows:

[0062]

[0063] In formulas (1) to (4), ρ(blue) represents the surface reflectance value of the blue band, ρ(nir) represents the surface reflectance value of the near-infrared band, ρ(green) represents the surface reflectance value of the green band, ρ(red) represents the surface reflectance value of the red band, ρ(swir1) represents the surface reflectance value of the shortwave infrared 1 band, and ρ(swir2) represents the surface reflectance value of the shortwave infrared 2 band.

[0064] Step 3: Define the wetland landscape categories of the Liaohe River Delta and clarify 13 land cover types;

[0065] The Liaohe River Delta region boasts diverse topography, encompassing large areas of natural wetlands such as reed beds and mudflats, as well as extensive paddy fields and dry land. As a coastal area, coastal development has created large areas of artificial wetlands. The primary classification, based on hydrological and surface soil conditions, categorizes wetlands into three types: natural wetlands, artificial wetlands, and non-wetlands. Building upon this primary classification, the secondary classification further subdivides natural wetlands into water bodies, marshes, and mudflats, and non-wetlands into forest land, farmland, industrial and mining buildings, residential land, and unused land. To study the evolution of wetlands in the Liaohe River Delta and to classify them in more detail, this paper divides the wetlands in the study area into 13 tertiary types, as shown in Table 3. Based on the secondary classification, water bodies are further subdivided into rivers and shallow sea areas, and marshes are subdivided into reed beds and Suaeda salsa areas. Due to local economic development needs, artificial wetlands have been continuously developed and expanded, further subdividing them into salt pans, aquaculture areas, and reservoirs. Since rice is produced in the Liaohe River Delta region, farmland is further divided into paddy fields and dry land.

[0066] Table 3. Landscape Classification System of the Liaohe River Delta

[0067]

[0068]

[0069] Step 4: Select training samples by referring to the field survey data, the original multispectral image data, the first component of principal component analysis (PCA1) of the original remote sensing data, and the tasseled cap transformation brightness component of the original remote sensing data.

[0070] The training samples were drawn using original imagery, PCA feature images, TCT feature images, and field survey data. The number of samples is shown in Table 4. There were 479 samples in 1987, 495 in 1997, 723 in 2007, and 621 in 2017. To verify the accuracy of the remote sensing interpretation results for the Liaohe River Delta, validation samples were obtained using field survey data, totaling 232 validation points, distributed as follows: Figure 3 As shown, a vegetation index (NDVI) of 0.38 is used as the threshold to distinguish whether the information is vegetation; if it is greater than 0.38, it is considered vegetation. A water body index (MNDWI) greater than -0.1 indicates that the area is a water body. Near-infrared band (threshold 0.16) and tasseled cap transform brightness component values ​​(threshold 460) are used to distinguish mudflats from residential areas and industrial / mining areas. A bare soil index (SI) of 0.4 is the critical value; a value greater than 0.4 indicates that the area is dry land. Paddy fields, woodlands, and reed beds are distinguished using water body index (MNDWI), vegetation index (NDVI), and topographic elevation (DEM). Generally, paddy fields have higher water body information values, while woodlands have higher topographic elevation values.

[0071] Table 4. Sample size statistics for each year

[0072]

[0073] Step 5: For the preprocessed multispectral data, perform the first wetland information extraction based on maximum likelihood classification;

[0074] Supervised classification (using the maximum likelihood method) selects suitable samples and automatically extracts land cover classification information. Its advantages are high accuracy, small error, and fast speed, and it does not require GPU configuration. However, practical operation has found that the maximum likelihood method classification results for a single data source still have some errors (through field surveys and verification points, the overall classification accuracy reached 75.29%, and the Kappa coefficient was 0.71), and reeds appeared in the higher-altitude dry land.

[0075] Step 6: Extract wetland information based on decision tree classification from the preprocessed multispectral data;

[0076] Six feature parameters were selected: Brightness, NDVI, MNDWI, NIR, SI, and DEM. A decision tree method was used to perform preliminary classification of wetland landscapes. The parameter settings are as follows: Figure 4 As shown (where the parameter thresholds are determined based on the critical values ​​of spectral histogram transitions for different objects), this decision tree method still suffers from problems such as failing to distinguish between forest and dry land, rice and reeds, and bare land and residential areas (e.g., Figure 5 As shown in the figure, the decision tree does not achieve 13 classification results. For example, it identifies water bodies but does not categorize them into rivers, reservoirs, or shallow water areas. Therefore, its classification accuracy is relatively low, with an overall classification accuracy of 56.82% and a Kappa coefficient of 0.57. Thus, decision tree classification is not suitable for three-level classification, i.e., it is not applicable to situations with higher landscape categories.

[0077] Step 7: The first maximum likelihood classification result is used as one data band, the decision tree classification result is used as one band, and combined with the 9 feature parameters to form 11 bands of fused data.

[0078] The 11 bands (1+1+9) are combined using the LayerStack function module of ENVI software to form new image data with 11 bands.

[0079] Step 8: Extract information from the new fused data based on maximum likelihood classification to obtain the dynamic change analysis results of the Liaohe River Delta wetland landscape classification.

[0080] This embodiment uses field validation data to evaluate the accuracy of the classification results. Using the maximum likelihood classification method alone, the overall classification accuracy is 75.29%, and the Kappa coefficient is 0.71. Using the decision tree classification method alone, the overall classification accuracy reaches 66.82%, and the Kappa coefficient is 0.57, but it cannot classify 13 types. The method of this invention achieves an overall classification accuracy of 87.71%, and a Kappa coefficient of 0.85, which is 16.50% higher than the accuracy of using supervised classification alone, and 20.72% higher in Kappa coefficient accuracy.

[0081] This invention, through the above scheme, achieves the formation of new fused data based on the optimized combination of feature parameters, extracts wetland landscape information of the Liaohe River Delta in 1987, 1997, and 2017, and compares the dynamic changes of the Liaohe River wetland landscape over 30 years, such as... Figure 6 As shown.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A wetland landscape pattern classification method based on optimized feature parameters, characterized in that, Includes the following steps: Step A: Acquire the original image and preprocess the multispectral and topographic data of the target wetland; Step B: Extract the data feature parameters from the preprocessed data. The characteristic parameters include the first component of principal component analysis (PCA1), the tasseled cap transformation brightness component, normalized difference vegetation index (NDVI), normalized difference water index (MNDWI), red band spectral value (RED), near-infrared band spectral value (NIR), bare soil index (SI), urban building index (IBI), and digital elevation model (DEM). Step C: Remote sensing image information extraction and classification: (1) Sample selection: The wetland landscape category is defined by a three-level landscape classification, and the training samples are determined by referring to the field survey data, the first component of principal component analysis (PCA1) of the original multispectral image data, the brightness component of the tassel transformation of the original remote sensing data; (2) First maximum likelihood classification: For the multispectral data after preprocessing in step A, wetland information is extracted by first maximum likelihood classification in combination with the determined training samples; (3) Decision tree classification: Based on the preprocessed multispectral data in step A, wetland information is extracted using decision tree classification in combination with the determined training samples; (4) Band combination: The first maximum likelihood classification result is used as 1 data band, the decision tree classification result is used as 1 band, and combined with the 9 feature parameters extracted in step B to form 11 bands of fused data; (5) Second maximum likelihood classification: Information extraction based on maximum likelihood classification is performed on the new fused data to obtain the analysis results of the dynamic changes in the target wetland landscape classification.

2. The wetland landscape pattern classification method based on optimized feature parameters according to claim 1, characterized in that, In step A, the original image includes remote sensing images and DEM data of the target wetland from multiple years. The preprocessing process includes, but is not limited to, radiometric correction, atmospheric correction, and stitching and cropping operations to obtain image data of the target wetland.

3. The wetland landscape pattern classification method based on optimized feature parameters according to claim 1, characterized in that, In step C, the three-level landscape classification includes: Primary classification: Natural wetlands, constructed wetlands, and non-wetlands; Secondary classification: Natural wetlands are further subdivided into water bodies, wetlands, mudflats, and bare land; Non-wetlands are further subdivided into woodlands, farmland, industrial and mining buildings, and residential areas; The three-level classification system further subdivides water bodies into river waters and shallow sea waters, marshes into reed beds and Suaeda salsa areas, artificial wetlands into salt pans, aquaculture areas, and reservoirs, and farmland into paddy fields and dry land.

4. The wetland landscape pattern classification method based on optimized feature parameters according to claim 3, characterized in that, In step C, when performing decision tree classification, six feature parameters are selected as threshold judgment parameters: tasseled cap transform brightness component, normalized difference vegetation index (NDVI), normalized difference water index (MNDWI), near-infrared spectral value (NIR), bare soil index (SI), and digital elevation model (DEM).

Images