A multi-strategy, integrated, automated river channel width extraction method based on remote sensing data.

Abstract

This invention provides a multi-strategy, integrated, automated channel width extraction method based on remote sensing data, which improves the stability of water body boundaries and the accuracy of channel width measurement under complex background conditions. [Solution] The method performs preprocessing on multispectral remote sensing images. Three parallel water body extraction processes are performed. The first process performs high-speed screening based on a water body spectral index threshold and logical combinations. The second process performs pixel classification based on a supervised classifier such as a random forest. The third process performs discrimination based on image segmentation and object-level rules. Pixel-level decision fusion is performed on the results of the three processes, and a high-confidence water body mask is generated by employing a majority vote or weighted majority vote strategy. Finally, distance transformation and thinning are performed on the fused water body mask to extract the river channel centerline, self-adaptively position cross-sections along the centerline, and calculate the length of the intersection segments between the cross-sections and the river channel boundary to obtain the river channel width.

Description

[Technical Field]

[0001] The present invention relates to the technology of remote sensing data processing and river topography monitoring, and more particularly to a multi-strategy integrated automatic river channel width extraction method based on remote sensing data. [Background technology]

[0002] River channel width is an important parameter in river geomorphology and hydrology research, and can be used for river change analysis, flood risk assessment, ecological environment monitoring, and water resource management. Traditional methods of obtaining river channel width often rely on field surveys or manual interpretation, resulting in low efficiency, high costs, limited applicability, and difficulty in supporting long-term dynamic monitoring. With the development of remote sensing technology and cloud computing platforms, automated extraction of water bodies based on satellite imagery is becoming mainstream. Existing methods mainly include: (1) spectral index thresholding, which has a simple algorithm but is sensitive to thresholds and prone to misclassification in complex backgrounds; (2) supervised classification, which has high accuracy but depends on training samples and scene generalization; and (3) object-oriented analysis, which can utilize spatial contextual information but is computationally complex and sensitive to parameters. Under complex conditions such as mountain shading, dark-colored terrain, and riparian vegetation, a single method lacks stability. Furthermore, fluctuations in the water body boundary tend to accumulate and amplify in the measurement of river channel width, making it difficult to ensure consistency and reliability for wide-area application. [Overview of the Initiative]

[0003] The objective of the present invention is to provide a multi-strategy integrated automatic river channel width extraction method based on remote sensing data, which solves the problems of conventional single-water body extraction methods, such as high misclassification rates under complex background conditions, unstable results, and incomplete processes in the river channel width measurement.

[0004] To achieve the above objective, the present invention provides a multi-strategy fusion type automatic river channel width extraction method based on remote sensing data, which includes the following steps. Step 1: Acquire multispectral remote sensing images covering the research area and perform preprocessing, which includes at least cloud and / or shadow masking and reflectance matching. Step 2: Based on the preprocessed image, three parallel water body extraction processes are performed to obtain the first water body result W1(x), the second water body result W2(x), and the third water body result W3(x). Here, the first water body extraction process uses a water body spectral index threshold and logical combination rules to perform water body discrimination on a pixel-by-pixel basis. The second water body extraction process constructs a feature vector based on multiband reflectance and water body index, and uses a supervised classifier to perform water body discrimination on a pixel-by-pixel basis. The third water body extraction process divides the image to obtain homogeneous objects and determines water body objects based on a pre-defined set of rules using the spectral, geometric, and texture features of the objects. Step 3: Perform a pixel-level decision fusion on W1(x), W2(x), and W3(x), and create a fused aqueous mask W f (x) is obtained. Here, the pixel-level decision fusion employs either a majority vote strategy or a weighted majority vote strategy. Step 4, Fusion Water Mask W f Based on (x), the centerline of the river channel is extracted, a cross section is placed along the centerline, the length of the line segment where the cross section intersects with the river channel boundary is calculated, and the river channel width is obtained.

[0005] Preferably, the water body spectral index includes at least the normalized differential water body index NDWI, the improved normalized differential water body index MNDWI, and the normalized vegetation index NDVI, where NDWI = (Green - NIR) / (Green + NIR), MNDWI = (Green - SWIR1) / (Green + SWIR1), and NDVI = (NIR - Red) / (NIR + Red), where Green is the green band, NIR is the near-infrared band, and SWIR1 is the short-wave infrared 1 band.

[0006] Preferably, the first water body extraction system further includes setting a threshold for NDVI to exclude riparian vegetation and introducing reflectance interval constraints for the blue and green bands to reduce misclassification of shaded or dark-colored objects.

[0007] Preferably, the supervised classifier is a random forest classifier. Training samples are generated by manually interpreting aquatic / non-aquatic vector boundaries or by referencing authoritative aquatic products, and the reference boundaries are projected onto multispectral remote sensing images to form pixel-level labels. Hierarchical random sampling is employed to construct training and validation sets, where the ratio of aquatic to non-aquatic samples is 1:1 to 1:3, and a hierarchy is established for each different feature type (vegetation, bare ground, buildings, shaded areas, etc.) to improve generalization ability. Number of trees in the random forest: N t The range is 100-500, maximum depth D max The number of leaf node samples is determined by 10-30, or by a grid search of the validation set, with a minimum number of leaf node samples L. min The values ​​are set from 1 to 5, and bootstrap sampling and feature random selection (max features Training is performed using the formula =p / 3 (where p is the input feature dimension).

[0008] Preferably, the pre-defined set of rules in the third aqueous extraction system includes spectral index threshold rules, geometric constraint rules, and texture constraint rules.

[0009] Preferably, the process for extracting the third water body is as follows: For each segmented object, the object mean values ​​MNDWI, NDWI, NDVI, and the self-adaptive water body enhancement index AWEI, as well as the object compactness C and texture entropy E are calculated. Here, compactness C = 4π × area / perimeter 2be defined as follows. If (MNDWI>t1)∧(NDWI>t2)∧(NDVI<t3)∧(AWEI>t4)∧(C>t5)∧(E<t6) is satisfied, the corresponding object is determined to be a water body object. Here, t1 is the modified normalized difference water index threshold, which ranges from 0.10 to 0.20; t2 is the normalized difference water index threshold, which ranges from 0 to 0.05; t3 is the normalized difference vegetation index threshold, which ranges from 0.15 to 0.25; t4 is the adaptive water body enhancement index threshold, which ranges from 0 to 0.10; t5 is the object compactness threshold, which ranges from 0.10 to 0.30; t6 is the texture entropy threshold, which takes a value between 2.0 and 6.0 depending on the image quantization level. AWEI adopts the shadow-free form of AWEI nsh =Blue+2.5×Green-1.5×(NIR+SWIR1)-0.25×SWIR2, and the texture entropy E is a gray level co-occurrence matrix (GLCM) entropy statistic, where Blue is the blue band, and SWIR2 is the short-wave infrared 2 band.

[0010] Preferably, when a majority voting strategy is adopted for pixel-level decision fusion, the content is as follows: If and only if W1(x)+W2(x)+W3(x)≧2, pixel x is determined as a water body, and a fused water body mask that can be used stably can be obtained even when there is no verification set or no weight evaluation condition.

[0011] Preferably, when a weighted majority voting strategy is adopted for pixel-level decision fusion, the content is as follows: Assign accuracy weights w1, w2, and w3 to the water body extraction results of the first system, the second system, and the third system respectively, such that w1+w2+w3=1 is satisfied. Calculate the weighted score S(x)=w1W1(x)+w2W2(x)+w3W3(x), and determine pixel x as a water body when S(x)>T. Here, the threshold T is determined by self-adaptation of the verification set: construct a verification set V generated from manual interpretation or authoritative products, and within the interval [0.45, 0.70], generate candidate thresholds T with a step size ΔT (0.01~0.02) i by enumeration, and for each T i , calculate the F1 metric M(T of the fusion result and the verification set V i ), calculate M(T i ) the T that maximizes* =argmax M(T i Select ) as the threshold T. Stop when the increase in the maximum index falls below ε, or when enumeration is complete, with the value range of ε being 0.001 to 0.005. Under conditions where a validation set built by manual interpretation or authoritative products exists, combine self-adaptation of weights and thresholds to improve fusion accuracy and robustness and enhance the effect of generating a high-confidence aqueous mask.

[0012] Preferably, the extraction of the river channel centerline is as follows: A distance transformation is applied to the fused water mask, and the distance field D(x) = min || xy || (where y represents non-aqueous pixels and x represents aqueous pixels) is obtained. The ridge region of the distance field is used as a candidate centerline, and the centerline is made single-pixel wide by a thinning operation. The thinning operation is implemented by employing a morphological skeleton extraction or iterative thinning algorithm, and redundant boundary pixels are removed while maintaining the connectivity of the centerline, making the width of the centerline a single pixel. Preferably, a skeleton extraction method based on the local maximal constraint of the distance field is employed. Furthermore, short branch removal (length less than 3 to 5 pixels) is performed on the skeleton to remove noise branches.

[0013] Preferably, the process for obtaining the channel width in step 4 is as follows: Sampling points P along the centerline at fixed intervals d. i Install the fixed interval d, which is determined by the image spatial resolution r and the local river channel width estimate W^ i Based on this, d = max(d min ,min(d max ,α×W^ i )) is determined by, and here d min 3r~10r, d max The time interval is set to 50r to 200r, and α is set to 0.1 to 0.5. Each sampling point P i For this, the adjacent centerline point P i-1 , P i+1 Based on the tangent vector t i =normalize(P i-1 -P i+1 ) calculates and the normal vector n i =(-t{iy},t{i x}) is determined. Here t{i x} is the x-direction (horizontal axis direction) component of the tangent vector, t{i y} represents the y-direction (vertical axis direction) component. The predetermined length L(L≧β×W^ i β is the length scaling factor, and the cross-sectional line segment P is normal in the direction of the normal at 2-4). i ±(L / 2)·n i Generates the intersection Q of the cross section and the river channel boundary. i1 Q i2 Calculate the length W of the intersecting line segment between the two intersections. i = || Q i1 -Q i2|| This is defined as the channel width at the sampling point.

[0014] Therefore, by employing the above-mentioned automatic river channel width extraction method based on multi-strategy fusion of remote sensing data, the present invention has the following beneficial effects. (1) By extracting water bodies in parallel using three complementary strategies—spectral exponential thresholding, supervised classification, and object-oriented methods—and fusing them at the pixel scale, misclassification in a single method is significantly reduced, and the stability of water body boundaries under complex background conditions is improved. (2) A weighted voting mechanism based on majority rule and accuracy weighting is adopted to enhance the robustness of the fusion results and reduce the cumulative expansion effect of boundary changes on river channel width measurements. (3) An end-to-end automated process is constructed from remote sensing images to river channel width parameters, which is applicable to wide-area and long-term river channel width monitoring tasks and enables highly efficient batch processing on a cloud computing platform.

[0015] The technical solutions of the present invention will be described in more detail below with reference to the attached drawings and embodiments. [Brief explanation of the drawing]

[0016] [Figure 1]This is an overall flowchart of a multi-strategy integrated river channel width automatic extraction method based on remote sensing data according to an embodiment of the present invention. [Figure 2] This is a schematic diagram of the structure of three-system aqueous extraction and pixel-level determination fusion according to an embodiment of the present invention. [Figure 3] This is a schematic diagram illustrating centerline extraction and cross-sectional river channel width measurement based on distance conversion according to an embodiment of the present invention. [Figure 4] This diagram shows the water body mask and channel width measurement results for the target area in an embodiment of the present invention, where (a) is the input image diagram, (b) is the fused water body mask diagram, and (c) is the channel width spatial distribution diagram. [Modes for carrying out the invention]

[0017] The following detailed description of embodiments of the present invention shown in the accompanying drawings is not intended to limit the scope of the present invention as defined in the claims, but merely to illustrate selected embodiments of the present invention. All other embodiments that can be obtained based on embodiments of the present invention without creative work by a person skilled in the art are all within the scope of the present invention.

[0018] Referring to Figures 1 to 4, the multi-strategy fusion-type automatic channel width extraction method based on remote sensing data employs a technical path of "data preprocessing - multi-strategy water body extraction - pixel-level determination fusion - centerline / cross-sectional geometric measurement," and includes the following steps. Step 1: Acquire multispectral remote sensing images covering the research area and perform preprocessing, which includes at least cloud and / or shadow masking and reflectance matching. Step 2: Based on the preprocessed image, three parallel water body extraction processes are performed to obtain the first water body result W1(x), the second water body result W2(x), and the third water body result W3(x). Here, the first water body extraction process uses a water body spectral index threshold and logical combination rules to perform water body discrimination on a pixel-by-pixel basis. The second water body extraction process constructs a feature vector based on multiband reflectance and water body index, and uses a supervised classifier to perform water body discrimination on a pixel-by-pixel basis. The third water body extraction process divides the image to obtain homogeneous objects and determines water body objects based on a pre-defined set of rules using the spectral, geometric, and texture features of the objects. Step 3: Perform a pixel-level decision fusion on W1(x), W2(x), and W3(x), and create a fused aqueous mask W f (x) is obtained. Here, the pixel-level decision fusion employs either a majority vote strategy or a weighted majority vote strategy. Step 4, Fusion Water Mask W f Based on (x), the centerline of the river channel is extracted, a cross section is placed along the centerline, the length of the line segment where the cross section intersects with the river channel boundary is calculated, and the river channel width is obtained.

[0019] Examples This embodiment performs automatic extraction of river channel width based on Sentinel-2 images, and specifically, it is as follows: (1) Data acquisition and preprocessing: Sentinel-2 multispectral images of the research area are acquired, and the blue band B2, green band B3, red band B4, near-infrared band B8, and shortwave infrared bands B11 and B12 are selected. Cloud / shadow masking is performed, atmospheric correction is applied, and the images are standardized to ground reflectance. (2) Extraction of first system water bodies: Indices such as NDWI, MNDWI, and NDVI are calculated, and water bodies are screened using empirical or statistical thresholds. Furthermore, a constraint of NDVI < 0.2 is introduced to exclude vegetation, and the reflectance relationship between the blue / green band and the SWIR band is combined to suppress misclassification of shadows and obtain W1(x). (3) Second system water body extraction (supervised classification): Construct a feature vector F(x) including multi-band reflectance {B2,B3,B4,B8,B11,B12} and water body indices {NDWI,MNDWI,NDVI,AWEI}. The training samples are derived from pixel labels obtained by projecting the water / non-water vector boundaries (or authoritative water body products) from manual interpretation onto the current image. Adopt hierarchical random sampling to construct a training set and a validation set (for example, stratified by "water area, vegetation, bare land, building, shadow / dark feature"), set the ratio of water body samples to non-water body samples to 1:1 to 1:3, and split the samples at a ratio of 7:3 or 8:2 for training / validation. Examples of random forest hyperparameters: Number of trees N t =300 (parameters can be adjusted within the range of 100 to 500), maximum depth D max =20, minimum leaf node sample number L min =2, max features =p / 3, bootstrap=True, perform a grid search on the validation set using the F1 metric as the indicator, and the optimal parameters can be determined. Obtain the pixel-level classification result W2(x), and suppress noise for W2(x) by performing an opening operation and removing small spots. (4) Third system water body extraction (object-oriented rule discrimination): Adopt multi-scale segmentation (Region Based Growth or SLIC superpixel) to form a homogeneous object set. For each object, calculate the object average spectrum and indices MNDWI, NDWI, NDVI, AWEI, and calculate geometric features (area A, perimeter P, compactness C=4πA / P 2 ) and texture features (GLCM entropy E, homogeneity H). To avoid insufficient disclosure of the "preset rule", as a feasible rule set R, for example, an object is determined to be a water body object when satisfying (MNDWI>t1)∧(NDWI>t2)∧(NDVI<t3)∧(AWEI>t4)∧(C>t5)∧(E<t6). Here, t1=0.10~0.20, t2=0~0.05, t3=0.15~0.25, t4=0~0.10, t5=0.10~0.30, t6=2.0~6.0. In addition, the rule is defined as object water body score S=Σj=1 ^j(6α j· I j )(I j α is the 0 / 1 indicator function for each rule. j Σα is the weight j =1, “Σ j=1 Let ^j(...)'' represent the sum of ... from j=1 to j, and let S≧T s (T s It may be determined to be a water object if the value is between 0.6 and 0.8. AWEI is the AWEI without shading format. nsh The formula used is =Blue + 2.5 × Green - 1.5 × (NIR + SWIR1) - 0.25 × SWIR2, and the object-level discrimination result is rasterized to obtain W3(x). (5) Pixel-level determination fusion: First, adopt a majority vote, W fision Let (x) be a liquid body if W1(x)+W2(x)+W3(x)≧2. If there is verification data, further weighted voting is performed: calculate S(x)=w1W1(x)+w2W2(x)+w3W3(x), and set the threshold T to W f The determination is made that (x) = I(S(x) > T). The weights w1, w2, and w3 are obtained by normalizing the accuracy of each lineage on the validation set. The self-adaptive determination method for the threshold T is a candidate threshold T within the range [0.45, 0.70]. i The results are enumerated in step ΔT, and the F1 index M(T) of the fusion result and validation set are combined. i ) calculate T * =argmax M(T i ) The process stops when the increase in the maximum index falls below ε or when enumeration is complete. (6) Measurement of centerline and river channel width: Fusion water mask W f After filling in the gaps and removing small spots for (x), the distance field D(x) = min{y∈Ω} \ Wf} || xy || The distance field is calculated, the ridges are extracted as centerline candidates, and a single-pixel wide centerline is obtained by thinning. Sampling points P are taken along the centerline at intervals d. i Place d=max(d min ,min(d max ,α×W^ i )) and dmin is 3r to 10r, d max is 50r to 200r, where r is the image spatial resolution, Ŵ i is the local approximate river channel width value obtained based on the distance field or neighborhood statistics. For sampling point P i , the tangent vector t is calculated using adjacent points i =normalize(P i-1 -P i+1 ), and the normal vector n i =(-t_i y , t_i x ) is calculated. Here, t_i y denotes the y-direction (vertical axis direction) component of the tangent vector, and t_i x denotes the x-direction (horizontal axis direction) component. To ensure intersection with the boundaries on both sides, a cross-sectional line segment P with length L (L≧β×Ŵ i (β=2-4)) in the normal direction i ±(L / 2)·n i is generated. The intersection points Q between the cross-section and the water body boundary i1 , Q i2 are determined, the width W i = || Q i1 -Q i2|| is calculated, and the river channel width sequence {(P i , W i )} is output.

[0020] Proposal for key parameters and value selection: When Sentinel-2 (r=10m) is adopted, d min can be 30 to 100m, and d max can be 500 to 2000m. If the river channel has a large curvature, d can be appropriately reduced to increase the sampling density along the line. The segmentation scale parameter can be determined by conducting a small-range test on the image to achieve the state of "minimum internal dispersion of the object and optimal boundary alignment".

[0021] Validation Set Construction Instructions: Validation set V can be obtained by manually sampling and marking typical areas (e.g., extracting multiple river sections and typical non-aquatic areas from each scene image). It can also be generated based on higher-resolution images or existing aquatic vector data. To avoid the effects of categorical imbalance, it is recommended to sample hierarchically, ensuring that the sample amounts for aquatic and non-aquatic areas are roughly equal.

[0022] The method provided by the present invention is executed on a cloud computing platform and enables batch processing of wide-area, multi-phase remote sensing images to extract river channel widths. The cloud computing platform includes a readable storage medium on which a computer program is stored, and the method provided by the present invention is implemented when the computer program is executed by a processor.

[0023] The present invention further includes an automatic channel width extraction system, specifically comprising a preprocessing module, a water body extraction module, a decision fusion module, and a channel width measurement module. The water body extraction module includes exponential thresholding, supervised classification, and object-oriented branching. The channel width measurement module performs distance transformation and centerline extraction based on a fused water body mask, and outputs the channel width by generating a cross section along the centerline.

[0024] Therefore, the present invention employs a multi-strategy fusion-type automatic channel width extraction method based on the aforementioned remote sensing data. First, preprocessing is performed on the multispectral image with cloud masking and reflectance matching. Subsequently, three systems of water body extraction are performed in parallel: high-speed screening using exponential thresholding, high-precision discrimination using random forest supervised classification, and discriminant analysis using object-oriented rules. The results of the three systems are generated through pixel-level majority voting or weighted self-adaptive fusion to produce a highly reliable water body mask. On this basis, distance transformation and thinning are performed on the mask to extract the channel centerline, cross-sections are self-adaptively placed along the centerline, and the channel width is obtained by calculating the length of the intersection segment between the cross-sections and the channel boundary. Through the mutual complementarity and decision fusion of multiple strategies, the method effectively suppresses interference from features such as shading, improves the stability of the water body boundary, and realizes automated and high-precision extraction of channel width over a wide area.

[0025] Finally, the above embodiments are used solely to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and that such modifications or equivalent substitutions will not cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-strategy integrated automatic river channel width extraction method based on remote sensing data, Step 1: Acquire a multispectral remote sensing image covering the target area and perform preprocessing, the preprocessing including at least cloud and / or shadow masking and reflectance matching. Step 2: Based on the preprocessed image, three parallel water body extraction processes are performed to obtain the first water body result W1(x), the second water body result W2(x), and the third water body result W3(x). Here, the first water body extraction process uses a water body spectral index threshold and logical combination rules to perform water body discrimination on a pixel-by-pixel basis. The second water body extraction process constructs a feature vector based on multiband reflectance and water body index, and uses a supervised classifier to perform water body discrimination on a pixel-by-pixel basis. The third water body extraction process divides the image to obtain homogeneous objects and determines water body objects based on a pre-defined set of rules using the spectral, geometric, and texture features of the objects. Step 3: Perform pixel-level decision fusion on W1(x), W2(x), and W3(x) to obtain the fused aqueous mask Wf(x), where the pixel-level decision fusion employs either a majority voting strategy or a weighted majority voting strategy. Step 4: A multi-strategy fusion-type automatic river channel width extraction method based on remote sensing data, characterized by extracting the river channel centerline based on the fused water body mask Wf(x), arranging a cross section along the centerline, calculating the length of the intersection segment between the cross section and the river channel boundary, and obtaining the river channel width.

2. The water body spectral index comprises at least a normalized differential water body index NDWI, an improved normalized differential water body index MNDWI, and a normalized vegetation index NDVI, where NDWI = (Green - NIR) / (Green + NIR), MNDWI = (Green - SWIR1) / (Green + SWIR1), and NDVI = (NIR - Red) / (NIR + Red), where Green is the green band, NIR is the near-infrared band, and SWIR1 is the short-wave infrared 1 band, characterized in that the multi-strategy fusion type automatic channel width extraction method based on remote sensing data according to claim 1.

3. The multi-strategy integrated automatic channel width extraction method based on remote sensing data according to claim 2, further comprising setting a threshold for NDVI to exclude riparian vegetation and introducing reflectance interval constraints for blue and green bands to reduce misclassification of shaded or dark-colored objects.

4. The supervised classifier is a random forest classifier, and the training samples are generated by referencing manually interpreted aquatic / non-aquatic vector boundaries or authoritative aquatic products, and the reference boundaries are projected onto multispectral remote sensing images to form pixel-level labels, and hierarchical random sampling is employed to construct training and validation sets, where the ratio of aquatic samples to non-aquatic samples is 1:1 to 1:3, and a hierarchy is set for each different feature type to improve generalization ability, as described in claim 3, a multi-strategy fusion type automatic channel width extraction method based on remote sensing data.

5. The multi-strategy fusion type automatic channel width extraction method based on remote sensing data according to claim 4, characterized in that the pre-set rule set in the third system of water body extraction includes spectral index threshold rules, geometric constraint rules, and texture constraint rules.

6. The process for aqueous extraction of the third system is as follows: For each segmented object, calculate the object average value MNDWI, NDWI, NDVI, the adaptive water body enhancement index AWEI, as well as the object compactness C and texture entropy E, where compactness C = 4π × area / perimeter 2 where (MNDWI > t １ ) ∧ (NDWI > t ２ ) ∧ (NDVI < t ３ ) ∧ (AWEI > t ４ ) ∧ (C > t ５ ) ∧ (E < t ６ ), the object is determined to be a water body object, where t １ is the modified normalized difference water index threshold ranging from 0.10 to 0.20, t ２ is the normalized difference water index threshold ranging from 0 to 0.05, t ３ is the normalized difference vegetation index threshold ranging from 0.15 to 0.25, t ４ is the adaptive water body enhancement index threshold ranging from 0 to 0.10, t ５ is the object compactness threshold ranging from 0.10 to 0.30, t ６ is the texture entropy threshold that takes a value between 2.0 and 6.0 depending on the image quantization level, AWEI adopts the shadow-free form AWEI ｎｓｈ = Blue + 2.5 × Green - 1.5 × (NIR + SWIR1) - 0.25 × SWIR2, the texture entropy E is a gray level co-occurrence matrix entropy statistic, where Blue is the blue band, and SWIR2 is the shortwave infrared 2 band, the multi-strategy fusion automatic river channel width extraction method based on remote sensing data according to claim 5, characterized in that.

7. When employing a majority vote strategy for pixel-level decision fusion, The multi-strategy fused river channel width automatic extraction method based on remote sensing data according to claim 1, characterized in that, when W1(x) + W2(x) + W3(x) ≥ 2, and only in that case, pixel x is determined to be a water body, and a fused water body mask that can be reliably used even when no validation set exists or there are no weight evaluation conditions.

8. When employing a weighted majority voting strategy for pixel-level decision fusion, Precision weights w1, w2, and w3 are assigned to the water body extraction results of the first, second, and third systems, respectively, satisfying the condition w1 + w2 + w3 = 1. A weighted score S(x) = w1W1(x) + w2W2(x) + w3W3(x) is calculated, and if S(x) > T, pixel x is determined to be a water body. Here, the threshold T is determined by self-adaptation of the validation set. A multi-strategy fusion-type automatic channel width extraction method based on remote sensing data according to claim 1, characterized in that a validation set V is constructed from manual interpretation or generated from authoritative products, candidate thresholds Ti are enumerated within the interval [0.45, 0.70] with a step size ΔT (0.01 to 0.02), the fusion result at each Ti and the F1 index M(Ti) of the validation set V are calculated, the T* = argmax M(Ti) at which M(Ti) is maximized is selected as the threshold T, the process is stopped when the increase in the maximum index falls below ε or when enumeration is complete, and the range of ε is 0.001 to 0.

005.

9. The extraction of the river channel centerline is as follows: Apply distance transformation to the fused water mask, and the distance field D(x) = min || x - y || A multi-strategy fusion type automatic river channel width extraction method based on remote sensing data according to claim 1, characterized in that a (y) represents non-aquatic pixels and x represents aquatic pixels is obtained, the ridge region of the distance field is used as a candidate centerline, and the centerline is reduced to a single pixel width by a thinning operation.

10. The process for obtaining the river channel width in Step 4 is as follows: Sampling points P are set at fixed intervals d along the center line. ｉ The fixed interval d is set to the image spatial resolution r and the local river channel width estimate W^ ｉ Based on this, d = max(d ｍｉｎ ,min(d ｍａｘ , α×W^ ｉ )) is determined by, and here d ｍｉｎ 3r to 10r, d ｍａｘ The time interval is set to 50r to 200r, and α is set to 0.1 to 0.5, with each sampling point P ｉ For this, the adjacent centerline point P ｉ－１ , P ｉ＋１ Based on this, the tangent vector t ｉ =normalize(P ｉ－１ -P ｉ＋１ ) calculates and the normal vector n ｉ = (-t{i ｙ }, t{i ｘ }) is determined, and here t{i ｘ } is the x-direction (horizontal axis direction) component of the tangent vector, t{i ｙ } represents the y-direction (vertical axis direction) component of the tangent vector, and is a predetermined length L (L ≥ β × W^ ｉ β is the length scaling factor, and the cross-sectional line segment P is normal in the direction of the normal at 2-4). ｉ ±(L / 2)・n ｉ Generates the cross section and the intersection Q of the cross section and the river channel boundary. ｉ１ Q ｉ２ Calculate the length W of the intersecting line segment between the two intersections. ｉ = || Q ｉ１ - Q ｉ２|| The multi-strategy integrated automatic river channel width extraction method based on remote sensing data according to claim 1, characterized in that the river channel width at the sampling point is defined as follows.

Images