An automatic extraction method for river width based on multi-strategy fusion of remote sensing data

Abstract

The application discloses a kind of river width automatic extraction method based on remote sensing data multi-strategy fusion, it is related to remote sensing data processing and river geomorphology monitoring technical field, first to multispectral remote sensing image is pretreated;Then three-way water body extraction is executed in parallel: first way is based on water body spectral index threshold and logic combination fast preliminary screening, second way is based on random forest etc.supervision classifier and performs pixel classification, third way is based on image segmentation and object level rule discrimination;Pixel level decision fusion is carried out to three-way result, and majority vote or weighted majority vote strategy is used to generate high confidence water body mask;Finally, distance conversion and thinning are carried out to fusion water body mask, and river center line is extracted, cross section is adaptively laid along center line, and the intersection length of cross section and river boundary is obtained to obtain river width;The application is through multi-strategy fusion and decision optimization, effectively suppresses shadow etc.object misclassification, improves water body boundary stability and river width measurement precision under complex background.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing and river geomorphology monitoring technology, and in particular to an automatic method for extracting river width based on multi-strategy fusion of remote sensing data. Background Technology

[0002] River width is a key parameter in river geomorphology and hydrology research, and can be used for river evolution analysis, flood risk assessment, ecological environment monitoring and water resource management. Traditional methods of obtaining river width rely on field surveys or manual interpretation, which are inefficient, costly, have limited coverage and are difficult to support long-term dynamic monitoring. With the development of remote sensing technology and cloud computing platforms, automatic water extraction based on satellite imagery has gradually become the mainstream. Existing methods mainly include: (1) spectral index thresholding method, which is simple but sensitive to thresholds and prone to misjudgment in complex backgrounds; (2) supervised classification method, which has high accuracy but depends on training samples and scene generalization; (3) object-oriented analysis method, which can utilize spatial context information but is computationally complex and parameter sensitive. Under complex conditions such as mountain shadows, dark features and riverbank vegetation, the stability of a single method is insufficient, and the water boundary jitter is easily amplified in river width measurement, making it difficult to guarantee the consistency and reliability of large-scale applications. Summary of the Invention

[0003] The purpose of this invention is to provide an automatic method for extracting river width based on multi-strategy fusion of remote sensing data, which solves the problems of high misclassification rate, unstable results, and imperfect river width measurement link in the existing single water body extraction method under complex background.

[0004] To achieve the above objectives, this invention provides an automatic method for extracting river width based on multi-strategy fusion of remote sensing data, comprising the following steps: Step 1: Acquire multispectral remote sensing images covering the study area and perform preprocessing. Preprocessing includes at least cloud and / or shadow masking and reflectance homogenization. Step 2: Perform three-channel water extraction in parallel based on the preprocessed image to obtain the first channel water result. Results of the second water body and the results of the third water body The process involves three main steps: First, water extraction uses a water spectral index threshold and logical combination rules to identify water bodies in pixels; Second, water extraction constructs feature vectors based on multi-band reflectance and water index, and uses a supervised classifier to identify water bodies in pixels; Third, water extraction segments the image to obtain homogeneous objects, and determines water bodies based on the spectral, geometric, and texture features of the objects according to a set of preset rules. Step 3, for , , Perform pixel-level decision fusion to obtain a fused water mask. Among them, the meta-level decision fusion adopts a majority voting strategy or a weighted majority voting strategy; Step 4: Based on the fusion water mask Extract the centerline of the river channel and lay out cross sections along the centerline. Calculate the length of the intersection between the cross sections and the river channel boundary to obtain the river channel width.

[0005] Preferably, the water spectral indices include at least the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), and the Normalized Vegetation Index (NDVI), wherein... , , ,in It is in the green light band. It is in the near-infrared band. It is a shortwave infrared band 1.

[0006] Preferably, the first water extraction route also includes: setting up Thresholds were used to exclude riparian vegetation, and blue and green light reflectance range constraints were introduced to reduce misclassification of shadow or dark features.

[0007] Preferably, the supervised classifier is a random forest classifier; training samples are generated using manually interpreted water / non-water body vector boundaries or authoritative water body products as references, and the reference boundaries are projected onto multispectral remote sensing images to form pixel-level labels; stratified random sampling is used to construct training and validation sets, where the ratio of water body samples to non-water body samples is 1:1 to 1:3, and stratification is set separately for different land cover types (such as vegetation, bare land, buildings, and shadows) to improve generalization ability; the number of trees in the random forest... The maximum depth is between 100 and 500. The minimum number of leaf node samples is between 10 and 30, or determined by a grid search of the validation set. The values ​​are 1 to 5, using bootstrap sampling and random feature selection. Training was conducted, among which The input feature dimension.

[0008] Preferably, the preset rule set for the third water body extraction includes spectral index threshold rules, geometric constraint rules, and texture constraint rules.

[0009] Preferably, the extraction process for the third water body is as follows: calculate the mean value for each segmented object. , , With Adaptive Water Enhancement Index and object compactness and texture entropy Among them, compactness When ( The target of the determination is a water body, and the threshold for the improved normalized difference water index is... The normalized differential water index threshold ranges from 0.10 to 0.20. The normalized vegetation index threshold is between 0 and 0.05. The adaptive water enhancement index threshold ranges from 0.15 to 0.25. The object compactness threshold is between 0 and 0.10. The texture entropy threshold ranges from 0.10 to 0.30. The value is taken between 2.0 and 6.0 based on the image quantization level; Use no shadow Texture entropy E is the gray-level co-occurrence matrix (GLCM) entropy statistic, where Blue is the blue light band and SWIR2 is the shortwave infrared band 2.

[0010] Preferably, when pixel-level decision fusion adopts a majority voting strategy, the content is as follows: if and only if At that time, the pixel If identified as a water body, a stable and usable fused water body mask can still be obtained even when a validation set is lacking or weight evaluation conditions are not available.

[0011] Preferably, when a weighted majority voting strategy is adopted for pixel-level decision fusion, the following content is provided: Precision weights are set for the water extraction results of the first, second, and third paths in sequence. , , and satisfy Calculate the weighted score ,when Time determination of pixels For water bodies; where the threshold is... Adaptive determination of the validation set: Construct a validation set generated by human interpretation or authoritative products. In the interval Inner step length Enumerate candidate thresholds , Calculate each value between 0.01 and 0.02. Fusion results and validation set F1 indicator Select to make The largest as a threshold When the maximum indicator improvement is less than Or stop when the enumeration is complete. The value range is 0.001-0.005. Under the condition of having a validation set constructed by human interpretation or authoritative products, the fusion accuracy and robustness are improved by combining weight and threshold adaptively, thereby enhancing the generation effect of high-confidence water body mask.

[0012] Preferably, the extraction of the river centerline involves performing a distance transformation on the merged water body mask to obtain the distance field. ,in For non-water body pixels, Representing water body pixels; using the ridge region of the distance field as a candidate centerline, and thinning the centerline to a single pixel width through a thinning operation, the thinning operation is implemented using morphological skeleton extraction or iterative thinning algorithm, removing redundant boundary pixels while maintaining the connectivity of the centerline, so that the centerline width is a single pixel; preferably using a skeleton extraction method based on the local maximum constraint of the distance field; and performing short branch culling (length less than 3–5 pixels) on the skeleton to eliminate noise bifurcation.

[0013] Preferably, the process of obtaining the river width in step 4 is as follows: along the centerline at fixed intervals Set sampling points Fixed spacing Based on image spatial resolution Compared with the local river width estimate Confirmed, among which ， From 3r to 10r, From 50r to 200r, The value ranges from 0.1 to 0.5; for each sampling point Based on adjacent centerline points and Calculate the tangent vector and determine the normal vector. ,in, This represents the component of the tangent vector in the x-direction (horizontal coordinate direction). This represents the component of the tangent vector in the y-direction (vertical coordinate direction); with a preset length. Generate cross-sectional line segments in the normal direction. ,in and It ranges from 2 to 4. Indicate the length magnification factor (between 2 and 4); calculate the intersection of the cross-section and the river boundary. , and the length of the intersection of the two intersection points This represents the width of the river channel at that sampling point.

[0014] Therefore, the present invention employs the above-mentioned method for automatic river width extraction based on multi-strategy fusion of remote sensing data, which has the following beneficial effects: (1) By extracting water bodies in parallel using three complementary strategies—spectral index thresholding, supervised classification, and object-oriented approaches—and fusing them at the pixel scale, the misjudgment of a single method can be significantly reduced, and the stability of water body boundaries in complex backgrounds can be improved. (2) A weighted voting mechanism of majority voting and precision weight is adopted to improve the robustness of the fusion results and reduce the cumulative amplification effect of boundary jitter on river width measurement; (3) Construct an end-to-end automated process from remote sensing images to river width parameters, which is suitable for large-scale, long-term river monitoring tasks and can achieve efficient batch processing on a cloud computing platform.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of a method for automatically extracting river width based on multi-strategy fusion of remote sensing data according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of three-way water extraction and pixel-level decision fusion according to an embodiment of the present invention; Figure 3 This is a schematic diagram of centerline extraction and cross-sectional river width measurement based on distance transformation according to an embodiment of the present invention; Figure 4 The following is a diagram showing the water mask and river width measurement results of an example area according to an embodiment of the present invention, wherein (a) is the input image, (b) is the merged water mask image, and (c) is the spatial distribution map of the river width. Detailed Implementation

[0017] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0018] Please see Figures 1-4 An automatic method for extracting river width based on multi-strategy fusion of remote sensing data is proposed. This method employs a technical approach of "data preprocessing—multi-strategy water body extraction—pixel-level decision fusion—centerline / cross-section geometric measurement," and includes the following steps: Step 1: Acquire multispectral remote sensing images covering the study area and perform preprocessing. Preprocessing includes at least cloud and / or shadow masking and reflectance homogenization. Step 2: Perform three-channel water extraction in parallel based on the preprocessed image to obtain the first channel water result. Results of the second water body and the results of the third water body The process involves three main steps: First, water extraction uses a water spectral index threshold and logical combination rules to identify water bodies in pixels; Second, water extraction constructs feature vectors based on multi-band reflectance and water index, and uses a supervised classifier to identify water bodies in pixels; Third, water extraction segments the image to obtain homogeneous objects, and determines water bodies based on the spectral, geometric, and texture features of the objects according to a set of preset rules. Step 3, for , , Perform pixel-level decision fusion to obtain a fused water mask. Among them, the meta-level decision fusion adopts a majority voting strategy or a weighted majority voting strategy; Step 4: Based on the fusion water mask Extract the centerline of the river channel and lay out cross sections along the centerline. Calculate the length of the intersection between the cross sections and the river channel boundary to obtain the river channel width.

[0019] Example This embodiment automatically extracts the river channel width based on Sentinel-2 imagery, as detailed below: (1) Data acquisition and preprocessing: Sentinel-2 multispectral images of the study area were acquired, and the blue light B2, green light B3, red light B4, near-infrared B8, short-wave infrared B11 and B12 bands were selected; cloud / shadow masking was performed and atmospheric correction was carried out to unify the images to surface reflectance.

[0020] (2) First water extraction: calculation , and Indices such as [insert indices here] are used for initial screening of water bodies based on experience or statistical thresholds; further [introduction of indices here]... Constraining and excluding vegetation, and combining blue / green bands with The reflectivity relationship of the bands suppresses shadow misclassification, and yields... .

[0021] (3) Second-stage water body extraction (supervised classification): Constructing a multi-band reflectance... and water index eigenvectors Training samples are derived from pixel labels obtained by projecting manually interpreted water / non-water body vector boundaries (or authoritative water body products) onto the current image. Hierarchical random sampling is used to construct training and validation sets (e.g., hierarchically stratified by "water, vegetation, bare land, buildings, shadow / dark features"). The ratio of water body samples to non-water body samples is set to 1:1 to 1:3, and samples are further divided into training / validation sets at a ratio of 7:3 or 8:2. Example of a random forest hyperparameter: number of trees. (Parameters can be adjusted within the range of 100-500), maximum depth Minimum number of leaf node samples , , The optimal parameters can be determined by grid search on the validation set using F1 as the metric, resulting in pixel-level classification results. and to Perform opening operations and small patch removal to suppress noise.

[0022] (4) Third-path water body extraction (object-oriented rule discrimination): Multi-scale segmentation (based on region growing or SLIC superpixels) is used to form a homogeneous set of objects; the mean spectrum and index of each object are calculated: , , , And calculate geometric features (area) ,perimeter Compactness With texture features (GLCM entropy) Homogeneity To avoid insufficient disclosure of "preset rules", a set of executable rules is provided. :like( If it is, then it is determined to be a body of water; among which Alternatively, the rules can be written as object water body scores. ,in For each rule, a 0 / 1 indicator function. As weight and ;when It was determined to be a body of water. Use no shadow The object-level discrimination results are rasterized to obtain .

[0023] (5) Pixel-level decision fusion: First, majority voting is used to obtain the result. That is, when It was initially identified as a body of water. Further weighted voting was conducted once verification data was available: calculation... and with threshold determination Weight The threshold is obtained by normalizing the precision of each branch on the validation set. The adaptive determination method is as follows: within [0.45, 0.70], with a step size of... Enumerate candidate thresholds Calculate the F1 score of the fusion result and the validation set. ,Pick When the maximum indicator improvement is less than Or stop when the enumeration is complete.

[0024] (6) Centerline and river width measurement: The water body mask will be integrated. After filling the pores and removing small patches, the distance field was calculated. The range field ridge line is extracted as a candidate center line, and the single-pixel center line is obtained through refinement. The center lines are then spaced along the center lines at intervals... Setting up sampling points ,in , For 3r-10r, For 50r-200r, For image spatial resolution, This is a rough estimate of the local river width based on distance field or neighborhood statistics. For sampling points... Calculate the tangent vector using adjacent points. normal vector ,in, This represents the component of the tangent vector in the y-direction (vertical coordinate direction). This represents the component of the tangent vector in the x-direction (horizontal coordinate direction); it generates the length in the normal direction. cross-sectional line segment ,and ( To ensure intersection with both side boundaries, find the intersection point of the cross-section and the water body boundary. , And calculate the width Output river width sequence .

[0025] Key parameters and recommended values: When using Sentinel-2 (r=10m), the following values ​​can be used: , If the river channel is highly meandering, the pressure can be appropriately reduced. To increase the sampling density along the route; the segmentation scale parameter can be determined on the image by conducting small-scale experiments based on the criteria of "minimum internal variance of the object and best boundary alignment".

[0026] Validation set construction instructions: The validation set V can be obtained by manually sampling and labeling typical areas (e.g., extracting several river sections and typical non-water areas from each image), or it can be generated by registering higher resolution images or historical water vector data. To avoid the impact of class imbalance, it is recommended to keep the sample size of water bodies and non-water bodies on the same order of magnitude and perform stratified sampling.

[0027] The method provided by this invention is executed on a cloud computing platform to achieve batch processing of river width extraction from large-scale, multi-temporal remote sensing images. The cloud computing platform includes a readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the method proposed in this invention.

[0028] The present invention also includes an automatic river width extraction system, specifically including a preprocessing module, a water body extraction module, a decision fusion module, and a river width measurement module; wherein the water body extraction module includes an exponential threshold branch, a supervised classification branch, and an object-oriented branch; the river width measurement module is used to perform distance transformation and centerline extraction based on the fused water body mask, and generate a cross-section along the centerline to output the river width.

[0029] Therefore, this invention employs the aforementioned method for automatic river width extraction based on multi-strategy fusion of remote sensing data. First, multispectral images undergo cloud masking and reflectance consistency preprocessing. Then, three parallel water body extraction methods are executed: rapid initial screening using an exponential threshold method, accurate identification using random forest supervised classification, and object-oriented rule-based discriminant analysis. The three results are then fused using pixel-level majority voting or weighted adaptive fusion to generate a high-confidence water body mask. Based on this, the mask is subjected to distance transformation and refinement to extract the river centerline. Cross-sections are adaptively deployed along the centerline, and the river width is obtained by calculating the intersection length of the cross-sections with the river boundary. This method, through multi-strategy complementarity and decision fusion, effectively suppresses interference from ground features such as shadows, improves water boundary stability, and achieves automated, high-precision extraction of river width over a large area.

[0030] Finally, it should be noted that the above embodiments are only used 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 should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatic extraction of river width based on multi-strategy fusion of remote sensing data, characterized in that, Includes the following steps: Step 1: Acquire multispectral remote sensing images covering the study area and perform preprocessing. Preprocessing includes at least cloud and / or shadow masking and reflectance homogenization. Step 2: Perform three-channel water extraction in parallel based on the preprocessed image to obtain the first channel water result. Results of the second water body and the results of the third water body The process involves three main steps: First, water extraction uses a water spectral index threshold and logical combination rules to identify water bodies in pixels; Second, water extraction constructs feature vectors based on multi-band reflectance and water index, and uses a supervised classifier to identify water bodies in pixels; Third, water extraction segments the image to obtain homogeneous objects, and determines water bodies based on the spectral, geometric, and texture features of the objects according to a set of preset rules. Step 3, for , , Perform pixel-level decision fusion to obtain a fused water mask. Among them, the meta-level decision fusion adopts a majority voting strategy or a weighted majority voting strategy; Step 4: Based on the fusion water mask Extract the centerline of the river channel and lay out cross sections along the centerline. Calculate the length of the intersection between the cross sections and the river channel boundary to obtain the river channel width.

2. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 1, characterized in that: Water spectral indices include at least the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), and the Normalized Vegetation Index (NDVI), among which... , , ,in It is in the green light band. It is in the near-infrared band. It is a shortwave infrared band 1.

3. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 2, characterized in that, The first-stage water extraction also includes: setting up... Thresholds were used to exclude riparian vegetation, and blue and green light reflectance range constraints were introduced to reduce misclassification of shadow or dark features.

4. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 3, characterized in that: The supervised classifier is a random forest classifier; training samples are generated using manually interpreted water / non-water body vector boundaries or authoritative water body products as references, and the reference boundaries are projected onto multispectral remote sensing images to form pixel-level labels; stratified random sampling is used to construct training and validation sets, in which the ratio of water body samples to non-water body samples is 1:1 to 1:3, and stratification is set for different land cover types to improve generalization ability.

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

6. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 5, characterized in that: The extraction process for the third water body is as follows: calculate the mean value for each segmented object. , , With Adaptive Water Enhancement Index and object compactness and texture entropy Among them, compactness When ( The target of the determination is a water body, and the threshold for the improved normalized difference water index is... The normalized differential water index threshold ranges from 0.10 to 0.

20. The normalized vegetation index threshold is between 0 and 0.

05. The adaptive water enhancement index threshold ranges from 0.15 to 0.

25. The object compactness threshold is between 0 and 0.

10. The texture entropy threshold ranges from 0.10 to 0.

30. The value is taken between 2.0 and 6.0 based on the image quantization level; Use no shadow Texture entropy E is the gray-level co-occurrence matrix entropy statistic, where Blue is the blue light band and SWIR2 is the shortwave infrared band 2.

7. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 1, characterized in that: When a majority voting strategy is adopted for meta-level decision fusion, the following applies: if and only if At that time, the pixel The water body is identified as a water body, and a stable fused water body mask is obtained when a validation set is lacking or when weight evaluation conditions are not available.

8. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 1, characterized in that: When a weighted majority voting strategy is adopted for pixel-level decision fusion, the following content is provided: Precision weights are assigned to the water extraction results of the first, second, and third paths in sequence. , , and satisfy Calculate the weighted score ,when Time determination of pixels For water bodies; where the threshold is... Adaptive determination of the validation set: Construct a validation set generated by human interpretation or authoritative products. In the interval Inner step length Enumerate candidate thresholds , Calculate each value between 0.01 and 0.

02. Fusion results and validation set F1 indicator Select to make The largest as a threshold ; When the maximum indicator improvement is less than Or stop when the enumeration is complete. The value range is 0.001-0.

005.

9. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 1, characterized in that: The extraction of the river centerline involves performing a distance transformation on the merged water body mask to obtain the distance field. ,in For non-water body pixels, Represents water body pixels; the ridge region of the distance field is used as a candidate for the center line, and the center line is made to be a single pixel width through a thinning operation.

10. The method for automatic extraction of river width based on multi-strategy fusion of remote sensing data according to claim 1, characterized in that, The process of obtaining the river width in step 4 is as follows: along the centerline at fixed intervals Set sampling points Fixed spacing Based on image spatial resolution Compared with the local river width estimate Confirmed, among which ， From 3r to 10r, From 50r to 200r, The value ranges from 0.1 to 0.5; for each sampling point Based on adjacent centerline points and Calculate the tangent vector and determine the normal vector. ,in, This represents the component of the tangent vector in the x-direction. This indicates that the tangent vector is in the y-direction; with a preset length. Generate cross-sectional line segments in the normal direction. ,in and It ranges from 2 to 4. Indicates the length magnification factor; calculates the intersection of the cross-section and the river channel boundary. , and the length of the intersection of the two intersection points This represents the width of the river channel at that sampling point.

Images