A fast water body extraction method from remote sensing images based on multi-scale GLCM feature fusion

Abstract

The present invention discloses a method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion, which relates to the technical field of water body remote sensing monitoring. The method comprises the following steps: S1, data collection; S2, calculating a water body index matrix W based on the collected data; S3, extracting multi-scale GLCM texture features; S4, weightedly fusing the water body index matrix W with multi-scale texture features to generate a comprehensive feature matrix; S5, defining a parameter space; S6, automatic threshold segmentation using the Otsu method; S7, generating a binary water body matrix; and S8, iteration. The present invention has the beneficial effect of introducing a Monte Carlo sampling and iterative optimization mechanism, performing a global search in a broad parameter space, finding an optimal parameter set, achieving fine weighted fusion of features at different scales, and automatically optimizing the weighted parameters of texture features, thereby significantly improving the accuracy and efficiency of water body recognition in complex environments, while reducing the need for manual parameter adjustment and having higher applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of water body remote sensing monitoring, and in particular to a method for rapidly extracting water bodies from remote sensing images by fusion of multi-scale GLCM features. Background Art

[0002] With advances in remote sensing technology and big data processing capabilities, remote sensing imagery is increasingly being used in water resource monitoring, environmental management, and disaster response. However, water area extraction techniques face multiple challenges, including accuracy and adaptability to complex environments. Existing water extraction techniques are primarily categorized as spectral feature-based methods and machine learning-based methods, each with its own advantages and disadvantages in multi-scale environments. Traditional spectral feature-based methods primarily rely on constructing water indices, such as the Normalized Difference Water Index (NDWI) and the Modified Normalized Difference Water Index (MNDWI). These methods calculate the reflectance of specific spectral combinations, such as near-infrared (NIR), green, and short-wave infrared (SWIR), to segment water bodies from non-water bodies. These methods are effective under specific conditions and are particularly suitable for single water environments or large water areas. However, NDWI and MNDWI have significant limitations in complex terrain and multi-scale environments, such as low recognition accuracy for small water bodies, poor extraction effect at complex boundaries, and susceptibility to interference from highly reflective areas such as vegetation and buildings.

[0003] Machine learning-based water extraction methods have developed rapidly in recent years. Using algorithms such as convolutional neural networks (CNNs) and support vector machines (SVMs) to identify water areas, these methods have demonstrated enhanced learning and generalization capabilities. These methods typically rely on large amounts of labeled data for training, which can improve the ability to identify detailed information in water extraction to a certain extent. However, machine learning methods have limited adaptability in multi-scale applications, particularly in cross-regional and spatiotemporal water extraction, where the effectiveness of identifying different water types can be significantly reduced. Furthermore, deep learning models typically require high computational resources, resulting in high computational costs when applied to large-scale remote sensing data.

[0004] To overcome the shortcomings of spectral and machine learning methods, some researchers have proposed multi-scale water extraction methods based on texture features. These methods use the Gray-Level Co-occurrence Matrix (GLCM) to extract image texture features, including inertia, angular second moment (ASM), and entropy, thereby providing water texture information to assist in identification. GLCM features, due to their ability to capture the spatial structure of images, have shown certain advantages in complex terrain and multi-scale environments. However, single texture features lack adaptability at different scales and still have limitations in processing detailed edges. Summary of the Invention

[0005] To solve the above problems, the present invention provides a method for rapid water body extraction from remote sensing images by fusion of multi-scale GLCM features, which is achieved through the following technical solutions.

[0006] A method for rapidly extracting water bodies from remote sensing images by fusion of multi-scale GLCM features, the method comprising the following steps:

[0007] S1, data collection, obtains remote sensing image data including green light, near infrared and short wave infrared bands from the radar image dataset of Sentinel-2 satellite. The original data matrices of remote sensing images of the three bands are I Green , I NIR and I SWIR , preprocess the original data matrix to obtain I' Green 、I' NIR and I' SWIR ;

[0008] S2, through I' Green 、I' NIR and I' SWIR The water index matrix W is calculated;

[0009] S3, multi-scale GLCM texture feature extraction, GLCM is the gray level co-occurrence matrix, and the elements of GLCM are defined as:

[0010] P(i,j|d,θ)

[0011] Where i, j represents the gray level, d is the pixel distance, and θ is the direction;

[0012] The following three texture features are extracted from GLCM:

[0013] Moment of inertia:

[0014]

[0015] Angular second moment:

[0016]

[0017] entropy:

[0018]

[0019] The above three texture features are calculated by kernels of different scales to obtain the multi-scale texture feature matrix T Inertia 、T ASM and T Entropy ;

[0020] S4, the water index matrix W and the multi-scale texture feature T Inertia 、T ASM 、T Entropy Weighted fusion to generate a comprehensive feature matrix S(x,y):

[0021]

[0022] Where α is the water index weight coefficient; β, γ, and δ are the weight coefficients of each texture feature respectively; w k is the comprehensive weight coefficient of multi-scale texture features;

[0023] S5, parameter space definition, define α, β, γ, δ and w k The parameter space is constructed based on the value range of , and the Monte Carlo method is used to randomly sample the parameter combination η in the parameter space. The comprehensive feature matrix S(x,y) is calculated based on the parameter combination η;

[0024] S6, Otsu method automatic threshold segmentation, applies Otsu method to the comprehensive feature matrix S of the determined parameter combination η, and automatically determines the optimal segmentation threshold of binarization. The formula is as follows:

[0025]

[0026] in represents the between-class variance;

[0027] S7, generate a binary water matrix, binarize the comprehensive feature matrix S according to the threshold τ, and generate a water mask matrix M(x,y):

[0028]

[0029] S8, iteration, set the number of iterations N, repeat steps S5 to S7 until the number of iterations is reached. In each iteration, the error E is calculated based on the generated water mask matrix. η :

[0030]

[0031] Among them, A true is the reference water area, A η is the water area calculated using the parameter combination η based on the water mask matrix M(x,y).

[0032] Preferably, in step S1, the pre-processing operation of the original data matrix is: processing the original data matrix I by median filtering or Gaussian filtering Green , I NIR and I SWIR , remove noise and outliers, then perform atmospheric and radiation corrections, and obtain I' Green 、I' NIR and I′ SWIR .

[0033] Preferably, in step S2, the calculation formula of the water index matrix W is as follows:

[0034] W(x,y) represents the water index value of the pixel.

[0035] Preferably, in step S2, the calculation formula of the water index matrix W is as follows:

[0036] W(x,y) represents the water index value of the pixel.

[0037] Preferably, in step S3, the kernel sizes used in the GLCM calculation are 3×3, 5×5 and 7×7 respectively.

[0038] Preferably, in step S5, α, β, γ, δ and w k The value range is [0.5,3.5], and the step size is 0.5.

[0039] Preferably, in step S5, α, β, γ, δ and w k The value range is [1,5], and the step size is 1.

[0040] Preferably, in step S5, α, β, γ, δ and w k The value range is [0.1,1], and the step size is 0.1.

[0041] Preferably, in step S8, in each iteration, A η is recorded and statistically analyzed in all iterations. Based on the maximum likelihood estimation principle, the calculated value A of the parameter combination η with the highest frequency is η It best reflects the area of ​​the water body.

[0042] The beneficial effects of the present invention are as follows:

[0043] 1. The present invention enhances the recognition ability of small-scale water bodies and complex boundary areas by fusing multi-scale GLCM texture features and automatically optimizing weights, avoiding the ambiguity and information loss problems of traditional NDWI and MNDWI methods on boundary details, thereby achieving high-precision water body shape extraction.

[0044] 2. The present invention introduces multi-scale texture features, extracts the spatial structure information of the image through the GLCM method, and combines it with Monte Carlo iterative optimization to effectively distinguish high-reflection areas such as vegetation and buildings, reducing the false detection rate in the water body extraction process. It is particularly suitable for water body identification in diverse and complex backgrounds.

[0045] 3. By iteratively optimizing the weights of multi-scale texture features through the Monte Carlo algorithm, the present invention can automatically adjust the water segmentation threshold in different environments, enhance the adaptability to complex terrain and multi-scale environments, and overcome the limitation of the traditional Otsu method that performs poorly in complex environments.

[0046] 4. The present invention realizes automatic parameter selection and weight adjustment through Monte Carlo iterative optimization, avoiding the inefficient process of manual repeated parameter adjustment, making the water extraction process more efficient, and improving the application convenience and stability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the description of the specific implementation methods. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 :Flowchart of a method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to the present invention;

[0049] Figure 2 : Schematic diagram of the present invention in extracting large lakes and filtering small water bodies;

[0050] Figure 3 : Comparison results of extracting small water bodies using this method and the traditional method based on water body index;

[0051] Figure 4 : Results of 100 Monte Carlo iterations. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0053] like Figure 1-4 As shown in FIG, a method for rapidly extracting water bodies from remote sensing images by fusion of multi-scale GLCM features is provided. The method comprises the following steps:

[0054] S1, data collection, obtains remote sensing image data including green light, near infrared and short wave infrared bands from the radar image dataset of Sentinel-2 satellite. The original data matrices of remote sensing images of the three bands are I Green , I NIR and I SWIR , preprocess the original data matrix to obtain I' Green 、I' NIR and I' SWIR .

[0055] The preprocessing operation of the original data matrix is: the original data matrix I is processed by median filtering or Gaussian filtering Green , I NIR and I SWIR , remove noise and outliers, then perform atmospheric and radiation correction to make it consistent with the physical spectral reflectance characteristics, and remove clouds and shadows. After correction, we get I' Green 、I' NIR and I' SWIR .

[0056] S2, through I' Green 、I' NIR and I' SWIR The water index matrix W is calculated.

[0057] Select the appropriate water index according to the application scenario. Commonly used ones are NDWI and MNDWI. The formulas are as follows:

[0058]

[0059] W(x,y) represents the water index value of the pixel.

[0060] S3, multi-scale GLCM texture feature extraction, GLCM is the gray level co-occurrence matrix, and the elements of GLCM are defined as:

[0061] P(i,j|d,θ)

[0062] Where i, j represents the gray level, d is the pixel distance, and θ is the direction;

[0063] The following three texture features are extracted from GLCM:

[0064] Moment of inertia:

[0065]

[0066] Angular second moment:

[0067]

[0068] entropy:

[0069]

[0070] The above three texture features are calculated by kernels of different scales. The kernel scales used in GLCM calculation are 3×3, 5×5 and 7×7 respectively, and the multi-scale texture feature matrix T is obtained. Inertia 、T ASM and T Entropy .

[0071] S4, the water index matrix W and the multi-scale texture feature T Inertia 、T ASM 、T Entropy Weighted fusion to generate a comprehensive feature matrix S(x,y):

[0072]

[0073] Where α is the water index weight coefficient; β, γ, and δ are the weight coefficients of each texture feature respectively; w k is the comprehensive weight coefficient of multi-scale texture features.

[0074] S5, parameter space definition, define α, β, γ, δ and w k The parameter space is constructed based on the value range of , and the Monte Carlo method is used to randomly sample the parameter combination η in the parameter space. The comprehensive feature matrix S(x,y) is calculated based on the parameter combination η.

[0075] Set the value range of the parameter space so that the best feature weight combination can be automatically found through Monte Carlo optimization. The specific parameter space can be adjusted according to needs. In general, α, β, γ, δ and w k The value range is [0.5, 3.5], with a step size of 0.5. This set of parameter spaces can be applied to water extraction for general needs. Through Monte Carlo iterative optimization and random sampling weight combination, the optimal parameter combination and feature matrix are returned by using the statistical characteristics of Monte Carlo sampling and large sample theory.

[0076] For specific application scenarios, when we extract large lakes based on remote sensing images, we hope to filter out small water bodies (such as small reservoirs and small tributaries). However, the traditional Otsu-based method does not have room for adjustment, but automatically sets the maximum inter-class variance as the threshold. Therefore, when faced with the situation of small water bodies that need to be filtered, it has obvious limitations. By setting the value of the parameter space in a larger range, α, β, γ, δ and w k The value range of is [1,5], with a step size of 1, which can filter out small water bodies. For parameter settings, the Monte Carlo method is used to perform random exploration in the parameter space without giving specific parameter values, such as Figure 2 These are the water body extraction results of four selected random parameters with obvious characteristics in this parameter space. The method can effectively filter small water bodies.

[0077] When we need to capture small water bodies, we set the parameter space range to a smaller space and reduce the step size, α, β, γ, δ and w at the same time. k The value range is [0.1, 1], and the step size is 0.1. When the parameter space and step size are set small, it is suitable for extracting small water bodies and improving the performance in detail processing and boundary recognition, such as Figure 3 The results of small water body extraction are compared with the results of traditional water body extraction method based on water body index.

[0078] S6, Otsu method automatic threshold segmentation, applies Otsu method to the comprehensive feature matrix S of the determined parameter combination η, and automatically determines the optimal segmentation threshold of binarization. The formula is as follows:

[0079]

[0080] in represents the between-class variance.

[0081] By adjusting the parameter combination, the optimal segmentation threshold can be automatically adjusted.

[0082] S7, generate a binary water matrix, binarize the comprehensive feature matrix S according to the threshold τ, and generate a water mask matrix M(x,y):

[0083]

[0084] S8, iteration, set the number of iterations N, repeat steps S5 to S7 until the number of iterations is reached. In each iteration, the error E is calculated based on the generated water mask matrix. η :

[0085]

[0086] Among them, Atrue is the reference water area, A η is the water area calculated using the parameter combination η based on the water mask matrix M(x,y).

[0087] At each iteration, A η is recorded and statistically analyzed in all iterations. Based on the maximum likelihood estimation principle, the calculated value A of the parameter combination η with the highest frequency is η It best reflects the area of ​​the water body.

[0088] When we need to accurately extract the area of ​​a water body, we can increase the number of Monte Carlo iterations. Figure 4 The results of 100 Monte Carlo parameter iterations show the evolution of the water area calculation results and their probability density estimates. The results indicate that over the 100 iterations, the calculated water area values ​​generally fluctuated around the reference value. Through random parameter adjustments, the algorithm was able to stabilize near the reference value. Based on the principle of maximum likelihood estimation, the calculated value with the highest frequency was selected as the result that best reflects the actual water area. The statistical properties of Monte Carlo sampling and large sample theory ensured the applicability and reliability of the parameter optimization process. These results provide a theoretical basis and data support for the further application of this method for large-scale water area monitoring.

[0089] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion, characterized by: The method comprises the following steps: S1, data collection, obtains remote sensing image data including green light, near infrared and short wave infrared bands from the radar image dataset of Sentinel-2 satellite. The original data matrices of remote sensing images of the three bands are I Green , I NIR and I SWIR , preprocess the original data matrix to obtain I′ Green , I′ NIR and I′ SWIR ; S2, through I′ Green , I′ NIR and I′ SWIR The water index matrix W is calculated; S3, multi-scale GLCM texture feature extraction, GLCM is the gray level co-occurrence matrix, and the elements of GLCM are defined as: P(i, j|d, θ) Where i, j represent gray levels, d is the pixel distance, and θ is the direction; The following three texture features are extracted from GLCM: Moment of inertia: Angular second moment: entropy: The above three texture features are calculated by kernels of different scales to obtain the multi-scale texture feature matrix T Inertia 、T ASM and T Entropy ; S4, the water index matrix W and the multi-scale texture feature T Inertia 、T ASM 、T Entropy Weighted fusion to generate a comprehensive feature matrix S(x, y): Where α is the water index weight coefficient; β, γ, and δ are the weight coefficients of each texture feature respectively; w k is the comprehensive weight coefficient of multi-scale texture features; S5, parameter space definition, define α, β, γ, δ and w k The parameter space is constructed based on the value range of , and the Monte Carlo method is used to randomly sample and generate parameter combinations η in the parameter space. The comprehensive feature matrix S(x, y) is calculated based on the parameter combination η; S6, 0tsu method automatic threshold segmentation, apply 0tsu method on the comprehensive feature matrix S of the determined parameter combination η, and automatically determine the optimal segmentation threshold of binarization. The formula is as follows: in represents the between-class variance; S7, generate a binary water matrix, binarize the comprehensive feature matrix S according to the threshold τ, and generate a water mask matrix M(x, y): S8, iteration, set the number of iterations, repeat steps S5 to S7 until the number of iterations is reached. In each iteration, the error E is calculated based on the generated water mask matrix. η : Among them, A true is the reference water area, A η is the water body area calculated using the parameter combination η based on the water body mask matrix M(x, y).

2. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 1 is characterized in that: In step S1, the pre-processing operation of the original data matrix is: processing the original data matrix I by median filtering or Gaussian filtering Green , I NIR and I SWIR , remove noise and outliers, then perform atmospheric and radiation corrections, and obtain I′ after correction Green , I′ NIR and I′ SWIR .

3. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 2 is characterized in that: In step S2, the calculation formula of the water index matrix W is as follows: W(x, y) represents the water index value of the pixel.

4. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 3 is characterized in that: In step S3, the kernel sizes used in the GLCM calculation are 3×3, 5×5, and 7×7 respectively.

5. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 4 is characterized in that: In step S5, α, β, γ, δ and w k The value range is [0.5, 3.5], and the step size is 0.

5.

6. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 4, characterized in that: In step S5, α, β, γ, δ and w k The value range is [1, 5], and the step size is 1.

7. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to claim 4, characterized in that: In step S5, α, β, γ, δ and w k The value range is [0.1, 1], and the step size is 0.

1.

8. The method for rapid water body extraction from remote sensing images using multi-scale GLCM feature fusion according to any one of claims 5 to 7, characterized in that: In step S8, each time it is iterated, A η is recorded and statistically analyzed in all iterations. Based on the maximum likelihood estimation principle, the calculated value A of the parameter combination η with the highest frequency is η It best reflects the area of ​​the water body.

Images