A waterlogged space distribution identification method and system based on multi-classifier cooperation

Abstract

This invention discloses a method and system for identifying the spatial distribution of accumulated water based on multi-classifier collaboration. The method includes: constructing a hierarchical classifier library comprising a preliminary screening layer, an arbitration layer, a re-identification layer, and a verification layer, with each layer containing multiple classifiers of different types. First, the preliminary screening layer performs coarse classification on UAV images, and the results of each classifier are fused using a logical OR operation to determine suspicious areas. Then, the arbitration layer and the re-identification layer sequentially perform refined identification of ambiguous areas, using a logical AND operation to obtain high-confidence accumulated water areas. For the remaining ambiguous areas, the verification layer is activated for final discrimination. The high-confidence accumulated water areas identified by each layer are merged to form a preliminary accumulated water distribution map. Pixel confidence is optimized through weighted fusion, and low-confidence pixels are removed. Finally, post-processing is performed to output the final spatial distribution result of the accumulated water. This application can effectively improve the identification accuracy and robustness of small and irregular accumulated water areas in complex urban scenes.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method and system for identifying the spatial distribution of water accumulation based on multi-classifier collaboration. Background Technology

[0002] In recent years, urban waterlogging monitoring technology has made some progress in the fields of remote sensing and computer vision, providing support for urban management in the face of extreme weather. In urban environments, waterlogging typically manifests as small areas with irregular shapes, highly intertwined with roads, buildings, and other features. In complex situations such as slippery surfaces, dark paving, or vehicle shadows, it is easily confused with non-waterlogged areas in terms of spectral or textural variations, affecting the final identification results. Current mainstream methods mostly rely on satellite remote sensing imagery to extract large-scale water bodies, but in rainy weather, cloud cover reduces image availability, affecting the timeliness and completeness of waterlogging information acquisition. Furthermore, existing methods are mostly designed for natural water bodies or large-scale water areas, with limited coverage of small waterlogged areas within cities, restricting their applicability in refined management scenarios. At the same time, single classification models often exhibit limited judgment ability, insufficient robustness, and unstable generalization performance when facing diverse scenarios, making it difficult to balance accuracy and adaptability.

[0003] In conclusion, improving the accuracy and adaptability of water accumulation spatial distribution identification in complex urban scenarios is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for identifying the spatial distribution of water accumulation based on multi-classifier collaboration, so as to solve the problems of low accuracy, poor robustness, and insufficient generalization ability of existing technologies in complex urban scenarios when identifying small and irregular water accumulation areas. The specific solution is as follows:

[0005] In a first aspect, the present invention provides a method for identifying the spatial distribution of accumulated water based on multi-classifier collaboration, comprising:

[0006] Construct a hierarchical classifier library, which includes at least an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer, with each layer containing at least one classifier;

[0007] Acquire drone images, classify and identify drone images through a preliminary screening layer, and fuse the identification results of each classifier in the preliminary screening layer to determine suspicious areas;

[0008] The suspicious areas are classified and identified by the arbitration layer, and the identification results of each classifier in the arbitration layer are merged to obtain the first high-confidence water accumulation area and the first ambiguous area.

[0009] The first ambiguous region is classified and identified by the re-identification layer, and the identification results of each classifier in the re-identification layer are fused to obtain the second high-confidence water accumulation region and the second ambiguous region.

[0010] The second ambiguous region is classified and identified by the verification layer, and the identification results of each classifier in the verification layer are fused to obtain the third high-confidence water accumulation region.

[0011] By merging the first, second, and third high-confidence water accumulation areas, a preliminary water accumulation distribution map is obtained;

[0012] The preliminary water accumulation distribution map is weighted and fused for optimization. The confidence level of each pixel is calculated according to the preset weights of each level classifier, and low-confidence pixels are removed based on a preset threshold to obtain the optimized water accumulation distribution map.

[0013] The optimized water accumulation distribution map is post-processed to output the final spatial distribution result of the water accumulation.

[0014] Furthermore, the initial screening layer includes a minimum distance classifier, a parallelepiped classifier, a spectral angle mapping classifier, and a maximum likelihood classifier; the fusion of the identification results from each classifier in the initial screening layer to determine suspicious regions includes:

[0015] Perform a pixel-by-pixel logical OR operation on the recognition results of each classifier in the initial screening layer;

[0016] If any classifier determines that a pixel is filled with water, then that pixel is marked as a suspicious area.

[0017] Furthermore, the arbitration layer includes an adaptive coherence estimation classifier, a constrained energy minimization classifier, and a Mahalanobis distance classifier;

[0018] The recognition results of each classifier in the fusion arbitration layer are used to obtain the first high-confidence water accumulation region and the first ambiguous region, including:

[0019] Perform a pixel-by-pixel logical AND operation on the recognition results of each classifier in the arbitration layer;

[0020] If all classifiers determine that a pixel is a water accumulation area, then the pixel is marked as the first high-confidence water accumulation area;

[0021] Pixels within the suspicious area, excluding the first high-confidence water accumulation area, are marked as the first blurred area.

[0022] Furthermore, the re-identification layer includes a neural network classifier and a support vector machine classifier; the fusion of the identification results of each classifier in the re-identification layer yields a second high-confidence water accumulation region and a second blurred region, including:

[0023] Perform a pixel-by-pixel logical AND operation on the recognition results of each classifier in the re-recognition layer;

[0024] If all classifiers determine that a pixel is a water accumulation area, then mark that pixel as the second high-confidence water accumulation area;

[0025] Pixels within the first blurred region, excluding the second high-confidence water accumulation region, are marked as the second blurred region.

[0026] Furthermore, before using the verification layer to classify and identify the second blurred region, the following steps are also included:

[0027] Determine whether the area of ​​the second blurred region is greater than a preset threshold;

[0028] If the value exceeds the preset threshold, the verification layer will be activated for classification and identification.

[0029] If the value is less than or equal to the preset threshold, the verification layer is skipped.

[0030] Furthermore, the verification layer includes an orthogonal subspace projection classifier and a spectral information divergence classifier; the classification and identification of the second fuzzy region using the verification layer includes:

[0031] Obtain the preset standard water body spectral characteristics;

[0032] The spectral vector of each pixel within the second blurred region is compared with the spectral features of the standard water body to calculate a similarity metric.

[0033] Based on the comparison result between the similarity metric and the preset threshold, it is determined whether the pixel is water accumulation;

[0034] The classifier in the verification layer does not rely on training samples during the discrimination process.

[0035] Furthermore, the preliminary waterlogging distribution map is weighted and optimized, including:

[0036] Assign preset weights to each classifier in the initial screening layer, arbitration layer, re-identification layer, and verification layer;

[0037] For each pixel, the confidence level of that pixel is calculated based on the weight of the layer in which the classifier that classifies it as water accumulation belongs.

[0038] Remove pixels with a confidence level below a preset threshold from the preliminary water accumulation distribution map.

[0039] Furthermore, the confidence score of each pixel is calculated based on the preset weights of each level of classifiers, including:

[0040] For each pixel, count the number of classifiers at each level that classify it as water accumulation;

[0041] The weights of the classifiers at each level that classifies a pixel as having water accumulation are summed to obtain the weighted confidence score for that pixel.

[0042] The preset threshold is a constant set based on empirical values ​​or experimental data.

[0043] Furthermore, the post-processing includes:

[0044] Morphological closure operation is performed on the optimized water accumulation distribution map;

[0045] The digital elevation model and land cover type map are acquired, and pixels located in areas with slopes greater than a preset threshold and on the top surface of buildings are identified and removed to obtain the final spatial distribution result of water accumulation.

[0046] Secondly, the present invention provides a spatial distribution identification system for accumulated water based on multi-classifier collaboration, comprising:

[0047] A classifier library construction module is used to build a hierarchical classifier library, which includes at least an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer, with each layer containing at least one classifier.

[0048] The first recognition and fusion module is used to acquire UAV images, classify and recognize the UAV images using the initial screening layer, and fuse the recognition results of each classifier in the initial screening layer to determine suspicious areas.

[0049] The second identification and fusion module is used to classify and identify suspicious areas using the arbitration layer, and to fuse the identification results of each classifier in the arbitration layer to obtain the first high-confidence water accumulation area and the first ambiguous area.

[0050] The third recognition and fusion module is used to classify and recognize the first ambiguous region using the re-recognition layer, and to fuse the recognition results of each classifier in the re-recognition layer to obtain the second high-confidence water accumulation region and the second ambiguous region.

[0051] The fourth recognition and fusion module is used to classify and recognize the second ambiguous region using the verification layer, and to fuse the recognition results of each classifier in the verification layer to obtain the third high-confidence water accumulation region.

[0052] The water accumulation area merging module is used to merge the first, second, and third high-confidence water accumulation areas to obtain a preliminary water accumulation distribution map;

[0053] The weighted fusion optimization module is used to perform weighted fusion optimization on the preliminary water accumulation distribution map. It calculates the confidence level of each pixel according to the preset weights of each level classifier and removes low-confidence pixels based on a preset threshold to obtain the optimized water accumulation distribution map.

[0054] The results output module is used to post-process the optimized water accumulation distribution map and output the final spatial distribution result of the water accumulation.

[0055] The beneficial effects of the technical solution of the present invention, which is a method and system for identifying the spatial distribution of water accumulation based on multi-classifier collaboration, include at least the following:

[0056] By constructing a hierarchical classifier library comprising an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer, this method performs progressively refined discrimination on UAV imagery. The initial screening layer employs multiple traditional spectral classifiers for broad screening, quickly locating potential waterlogged areas. The arbitration layer filters out obvious false positives from the initial screening results through high consistency requirements. The re-identification layer introduces a machine learning model for deep discrimination of ambiguous areas. The verification layer uses an unsupervised spectral matching mechanism for final confirmation when necessary. Results are fused between layers through logical operations, avoiding the limitations of a single model. In the result optimization stage, a hierarchical weighting mechanism is introduced to quantitatively evaluate pixel confidence and combines prior knowledge of terrain and land features for physical plausibility verification. This method, through a multi-level, multi-mechanism, and multi-criteria collaborative strategy, effectively addresses the challenges of small-scale waterlogged areas, complex backgrounds, and severe spectral obfuscation in urban environments, significantly improving the accuracy and scene adaptability of waterlogging identification. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of a method for identifying the spatial distribution of water accumulation based on multi-classifier collaboration, provided as an embodiment of this specification.

[0059] Figure 2 This diagram illustrates the initial screening layer identification in a multi-classifier collaborative method for identifying the spatial distribution of accumulated water, as provided in the embodiments of this specification.

[0060] Figure 3 This is a cutout diagram of a water accumulation spatial distribution identification method based on multi-classifier collaboration provided in the embodiments of this specification.

[0061] Figure 4 This diagram illustrates a high-confidence arbitration method for identifying the spatial distribution of accumulated water based on multi-classifier collaboration, as provided in an embodiment of this specification.

[0062] Figure 5 This diagram illustrates the re-identification process in a multi-classifier collaborative method for identifying the spatial distribution of accumulated water, as provided in an embodiment of this specification.

[0063] Figure 6This is a schematic diagram of the final verification of the residual area in a water accumulation spatial distribution identification method based on multi-classifier collaboration provided in the embodiments of this specification.

[0064] Figure 7 This diagram illustrates matrix merging in a multi-classifier collaborative method for identifying the spatial distribution of accumulated water, as provided in an embodiment of this specification.

[0065] Figure 8 This diagram illustrates the weighted calculation in a multi-classifier collaborative method for identifying the spatial distribution of accumulated water, as provided in the embodiments of this specification.

[0066] Figure 9 This is a schematic diagram of a water accumulation spatial distribution identification system based on multi-classifier collaboration, provided as an embodiment of this specification. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. In addition, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0068] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to facilitate the description of the embodiments and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this specification.

[0069] This invention provides a method for identifying the spatial distribution of water accumulation based on multi-classifier collaboration. By constructing a hierarchical classifier library and executing a multi-stage progressive identification process, it achieves accurate identification of water accumulation areas in UAV imagery. In this embodiment, urban flood monitoring is used as the application scenario. A UAV equipped with a multispectral imaging device is used to take aerial photographs of the target area, acquiring high-resolution multispectral image data. Based on the constructed hierarchical classifier library, the image data is processed by progressive discrimination and fusion, ultimately outputting the spatial distribution results of water accumulation that conform to physical laws and actual geographical conditions.

[0070] like Figure 1 As shown, the overall process of the present invention includes at least the following steps:

[0071] Step S1: Construct a hierarchical classifier library. In this step, a hierarchical classifier library containing four functional levels is pre-constructed. Based on the significant differences and characteristics of classifiers in terms of principles, assumptions, applicable scenarios, and sensitivity to noise, illumination, and reflection, they are divided into an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer. Each level contains at least one classifier, and each classifier distinguishes pixels in the image based on different principles. Specifically, the classifiers contained in each level are as follows:

[0072] The initial screening layer includes the minimum distance classifier (MinD), parallelepiped classifier (PP), spectral angle mapping classifier (SAM), and maximum likelihood classifier (ML). These four classifiers have the advantages of low computational complexity and parallel processing of the entire image, making them suitable for rapid initial screening of UAV imagery.

[0073] Arbitration layer: Includes Adaptive Coherence Estimation (ACE), Constrained Energy Minimization (CEM), and Mahalanobis Distance (MD) classifiers for high-confidence verification. Re-identification layer includes neural network classifiers and support vector machine classifiers. These three classifiers are highly resistant to background interference, sensitive to weak targets, and suitable for high-confidence verification of suspicious regions.

[0074] Validation layer: Includes Orthogonal Subspace Projection Classifier (OSP) and Spectral Information Divergence Classifier (SID). These two classifiers are unsupervised classifiers that do not rely on training samples and directly make judgments based on the degree of matching between pixel spectral features and water spectral absorption characteristics, and are suitable for final validation.

[0075] This hierarchical design allows different models to perform their respective functions, complement each other's strengths, and build a collaborative identification mechanism that progresses from coarse to fine, filters at each level, and verifies in multiple dimensions.

[0076] Step S2: Initial screening and fusion to identify suspicious areas. Image data of the area to be identified is acquired from drone images. This image data is input into the initial screening layer, where four classifiers (MinD, PP, SAM, ML) are used to independently classify and identify each pixel in the drone image. All four classifiers have advantages such as no need for iterative optimization, low computational complexity, and the ability to process the entire image in parallel. They are suitable for the initial rapid scanning of large-scale remote sensing images. Furthermore, they identify the spectral characteristics of water accumulation from different perspectives, including geometric distance, threshold range, spectral direction consistency, and statistical distribution assumptions, ensuring that at least one model can effectively respond to real water accumulation even under interference such as changes in lighting and complex surface materials. Please refer to [link to relevant documentation]. Figure 2Each classifier outputs a binary preliminary recognition result matrix, M1, M2, M3, and M4, based on its own discrimination rules. Pixels identified as water accumulation are assigned a value of 1, while pixels in other regions are assigned a value of 0. After obtaining the four preliminary recognition result matrices, a pixel-by-pixel logical OR operation is performed on them. For any pixel (p, q) in the image, if at least one of the four classifiers identifies it as water accumulation, then M1 is considered a valid binary matrix. n (p,q) = 1, otherwise M n (p,q) = 0. Here, (p,q) represents the pixel position in the matrix, n∈{1,2,3,4}. A pixel-wise logical OR operation is performed on the first-round recognition result matrices M1, M2, M3, and M4 to generate the initial screening matrix M. a The formula is: In this context, the symbol ⋁ represents the logical OR operation, a binary logical operation. For two Boolean values ​​A and B, the rule for the logical OR operation is: if at least one of A and B is true, the result is true; if both A and B are false, the result is false. Therefore, if any one of the classifiers identifies a certain area as having accumulated water, it is temporarily marked as a suspicious area. That is, if any matrix M1, M2, M3, or M4 in the identification result matrix is ​​identified as such... n In (n∈{1,2,3,4}), the pixel M n If (p, q) is assigned a value of 1, then the initial screening matrix M... a The pixel M a (p,q) = 1. M is considered non-flooded only when all four classifiers classify it as such, i.e., M1(p,q) = M2(p,q) = M3(p,q) = M4(p,q) = 0. a (p,q) are then assigned a value of 0. A mask for suspicious areas is generated through logical OR fusion to ensure that real water accumulation is not missed even under conditions of changing lighting or complex surface materials, thus guaranteeing a high recall rate for the initial screening layer.

[0077] Then based on the initial screening matrix M a Spatial domain cropping of the input image: Please refer to [link / reference]. Figure 3 Only retain the regions with a pixel value of 1 (i.e., satisfying M). a The region (p, q) = 1 is considered the effective processing region and sent to the next stage, the arbitration classifier layer; the remaining pixels (i.e., M) a The region (p,q) = 0 is explicitly marked as a non-waterlogged region and will no longer participate in any subsequent classification, fusion or optimization operations, which significantly reduces the computational load. After pruning, it can avoid repeatedly executing high-complexity models on large areas of dry surface, greatly improve the overall processing efficiency, and further reduce the risk of misjudgment propagation. The regions that have been excluded in the initial screening are completely isolated to prevent them from being incorrectly judged in subsequent identification due to model errors or noise interference.

[0078] Step S3: Arbitration layer identification and fusion to obtain the first high-confidence water accumulation region and the first ambiguous region. Using the suspicious region identified in Step S2 as input, the three classifiers (ACE, CEM, MD) within the arbitration layer are used to independently classify and identify each pixel within that region. These three classifiers all possess strong resistance to background interference and sensitivity to small targets. ACE, CEM, and MD focus on three different aspects: target-background coherence, energy constraint optimization, and multi-dimensional covariance structure modeling, respectively, avoiding systematic misjudgments due to bias in a single criterion.

[0079] Please see the appendix Figure 4 Each classifier outputs a binary discrimination result matrix M5, M6, and M7. The second-round recognition result matrix M... 5, The pixels in M6 and M7 are binary assigned, with water-filled pixels marked as 1 and non-water-filled pixels marked as 0.

[0080] Perform a pixel-by-pixel logical AND operation on the discrimination result matrices output by the three classifiers in the arbitration layer to generate the arbitration matrix M. b The formula is: The symbol ⋀ represents the logical AND operation. For two Boolean values ​​A and B, the rule of the logical AND operation is: if both A and B are true, the result is true; otherwise, the result is false. Therefore, for any pixel in the suspicious area, it is only marked as 1 in the logical AND operation result when all three classifiers classify it as water accumulation, that is, when the pixel is marked as 1 in all three result matrices. Specifically, if pixels M5(p,q), M6(p,q), and M7(p,q) in the two-round recognition result matrices M5, M6, and M7 are all assigned the value 1, then the arbitration matrix M... b The pixel M b (p, q) = 1. All other pixels are assigned a value of 0. Pixels with a logical AND operation result of 1 are marked as the first high-confidence water accumulation region, and the output is the first high-confidence water accumulation region matrix M. confirm1 ; Mark the pixels within the suspicious area, excluding the first high-confidence water accumulation area, as the first blurred area, and output the first blurred area matrix M. ambiguous1 Then proceed to the next round of recognition. The formula is as follows:

[0081] ;

[0082] These areas failed to reach a consensus at the arbitration level and require further, more refined judgment in the next stage. By extracting all consistently confirmed areas from the models through logic and operations, high-confidence water accumulation results are obtained, effectively suppressing false alarms.

[0083] Step S4: Re-identification and fusion at the re-identification layer yields the second high-confidence water accumulation region and the second blurred region. (See appendix for details.) Figure 4The first fuzzy region matrix M obtained in step S3 ambiguous1 As input, each pixel in the region is independently classified and identified using two classifiers, a neural network and a support vector machine, within the re-identification layer, generating two three-round recognition result matrices, M8 and M9.

[0084] These two classifiers are supervised learning models with strong nonlinear fitting capabilities and the ability to abstract complex spectral patterns. They automatically learn the boundary features between water accumulation and non-water accumulation using training data. With the support of representative training samples, they can significantly improve the ability to identify blurred regions. Each classifier outputs a binary discrimination matrix, with water accumulation pixels labeled as 1 and non-water accumulation pixels labeled as 0.

[0085] Please see the appendix Figure 5 The resolution matrices output by the two classifiers in the re-identification layer are subjected to a pixel-by-pixel logical AND operation to generate the re-identification matrix M. c The formula is as follows: For any pixel within the first fuzzy region, it is only marked as 1 in the logical AND operation result if both classifiers classify it as water accumulation. Pixels with a logical AND operation result of 1 are marked as the second high-confidence water accumulation region matrix M. confirm2 That is, if pixels M8(p,q) and M9(p,q) in the three-round recognition result matrices M8 and M9 are both assigned a value of 1, then the re-recognition matrix M... c The pixel M c (p, q) = 1. The pixels in the remaining areas are assigned a value of 0.

[0086] The pixels within the first blurred region, excluding the second high-confidence water accumulation region, are labeled as the second blurred region matrix M. ambiguous2 This step can improve the ability to distinguish easily confused targets such as shadows and slippery surfaces.

[0087] Step S5: Determine whether to activate the verification layer. Obtain the second fuzzy region matrix M obtained in step S4. ambiguous2 The area of ​​the second blurred region is calculated. This area is compared with a preset area threshold (e.g., 50 pixels). If the area of ​​the second blurred region is greater than the preset threshold, proceed to step S6 to initiate the verification layer for final verification; if the area of ​​the second blurred region is less than or equal to the preset threshold, skip the verification layer and proceed directly to step S7. This judgment step can effectively avoid unnecessary complex calculations for small areas, improving overall processing efficiency.

[0088] Step S6: Verification layer identification and fusion to obtain the third high-confidence water accumulation area.

[0089] If the result of step S5 is to activate the verification layer, then the second fuzzy region matrix M will be... ambiguous2As input, each pixel in the region is independently classified and identified using two unsupervised classifiers in the validation layer: an orthogonal subspace projection classifier and a spectral information divergence classifier, generating two four-round recognition result matrices M. 10 and M 11 The classifier in the validation layer does not rely on training samples during the discrimination process; it directly makes judgments based on the degree of matching between pixel spectral features and water spectral absorption characteristics.

[0090] The specific identification process is as follows: First, the preset standard water body spectral features are obtained, which represent the typical spectral response curves of accumulated water in multiple bands. For each pixel in the second fuzzy region, its spectral vector is compared with the standard water body spectral features: The orthogonal subspace projection classifier, based on the orthogonal subspace projection principle, calculates the similarity metric between the pixel spectrum and the target spectrum after removing interference signals. This metric reflects the projection intensity of the pixel spectrum in the direction of the water body features.

[0091] The spectral information divergence classifier is based on the principle of spectral information divergence. It calculates the degree of distribution difference between pixel spectra and standard water body spectra. This metric reflects the information-theoretic distance between the two spectral vectors.

[0092] Each classifier independently outputs a binary discrimination result matrix M based on the comparison result of the calculated similarity metric and its own preset threshold. 10 and M 11 .

[0093] The orthogonal subspace projection classifier sets a first threshold. If the similarity metric reaches the threshold, the pixel is marked as 1 in the orthogonal subspace projection classifier result matrix; otherwise, it is marked as 0. The spectral information divergence classifier sets a second threshold. If the similarity metric reaches the threshold, the pixel is marked as 1 in the spectral information divergence classifier result matrix; otherwise, it is marked as 0.

[0094] Please see the appendix Figure 6 The discrimination results M output by the two classifiers in the validation layer are used to determine the outcome. 10 and M 11 The matrix is ​​subjected to a pixel-by-pixel logical AND operation. For any pixel within the second fuzzy region, it is marked as 1 in the logical AND result only if both classifiers classify it as water accumulation. Pixels with a logical AND result of 1 are marked as the third high-confidence water accumulation region matrix M. confirm3 .

[0095] Step S7: Merge the high-confidence waterlogged areas to obtain a preliminary waterlogged distribution map. Please refer to the appendix. Figure 7 The first high-confidence water accumulation region matrix M obtained in step S3 is... confirm1 The second high-confidence water accumulation region matrix M obtained in step S4confirm2 And the third high-confidence water accumulation region matrix M obtained in step S6 confirm3 Merging is performed. Merging uses a pixel-by-pixel logical OR operation; that is, if a pixel is marked as water accumulation in any high-confidence water accumulation region, it is marked as water accumulation in the merged result. The merged region is the initial water accumulation matrix M. confirm The formula is as follows:

[0096] ;

[0097] Step S8: Weighted fusion optimization to obtain the optimized water accumulation distribution map. Although the preliminary water accumulation distribution map has high accuracy after the aforementioned multi-stage collaborative identification, there may still be individual low-confidence misclassified pixels caused by noise or interference. This step uses weighted fusion optimization to remove these low-confidence pixels. First, preset weights are assigned to classifiers within each level. The preset weights are dynamically allocated based on the historical verification accuracy or F1-score of each level's classifier, where the higher the accuracy, the greater the weight of the corresponding level. In a preferred embodiment, the weight of the initial screening layer is set to w1=0. This weight setting considers that the initial screening layer is only responsible for high-recall initial screening, and the reliability of a single judgment is low. The weight of the arbitration layer is set to w2=2, that is, if a pixel is identified as water accumulation by one of the classifiers, the confidence is increased by 2. This weight setting considers that the arbitration layer adopts a multiple consensus mechanism, and the discrimination result has high confidence. The re-identification classifier layer is used to finely distinguish easily confused targets, and its reliability is between that of the initial screening and arbitration, so the weight is set to w3=1. If a pixel is identified as water accumulation by one of the classifiers, the confidence level is incremented by 1. The validation classifier layer does not depend on the training samples, is sensitive to subtle spectral differences, and is suitable for finally confirming ambiguous regions; therefore, its weight is set to w4=1. That is, if a pixel is identified as water accumulation by one of the classifiers, the confidence level is incremented by 1.

[0098] Please see the appendix Figure 8 For each pixel in the preliminary water accumulation distribution map, the number of classifiers at each level that classify it as water accumulation is counted, and the weights of the levels to which these classifiers belong are accumulated to obtain the weighted confidence score t for that pixel. m For the initial water accumulation matrix M confirm For each pixel, if it is determined to be water accumulation in layer m, then the output of that pixel in layer m is defined as M. m (p,q) = 1, and the rest of the layers are 0. The confidence score S(p,q) for this pixel is calculated based on the set weights, using the following formula: Where: m represents the classifier layer, n∈{1,2,3,4}, t m Let be the confidence score of the m-th layer. If a pixel is classified as water accumulation by multiple classifier layers, its score will be accumulated with the corresponding weight.

[0099] Set the confidence threshold T = 6, and according to each pixel M m (p,q), the confidence score S(p,q), perform a filtering operation on the preliminary waterlogging matrix M confirm as follows:

[0100] If S(p,q)≥T, then finally determine that this pixel is waterlogged and output it to the result waterlogging matrix M final ;

[0101] If S(p,q)<T, then finally determine that this pixel is not waterlogged and filter it;

[0102] Step S9: Post-processing, output the final waterlogging spatial distribution result. In order to further improve the spatial coherence and physical rationality of the recognition result, post-process the optimized waterlogging distribution map. The post-processing of this embodiment includes the following two sub-steps:

[0103] First, perform morphological closing operation on the optimized waterlogging distribution map. Closing operation is an operation of first dilation and then erosion, which can effectively eliminate isolated noises, fill holes, and smooth boundaries. Thus, it enhances the overall connectivity and geometric integrity of the region. Secondly, combine the digital elevation model and the land cover type map to perform physical rationality verification on the result after the closing operation. Obtain the digital elevation model and the land cover type map of the area to be recognized. Based on the digital elevation model, use the third-order irregular plane method to calculate the slope value of each grid cell and generate a slope map; register the slope map and the optimized waterlogging distribution map spatially so that each waterlogged pixel is associated with its corresponding slope value;剔除所有关联坡度值大于预设阈值(如5°)的积水像素。这些区域在物理上不可能形成积水,因此将其从识别结果中剔除。经过上述后处理,最终输出的积水空间分布结果具有空间连续、语义准确且符合地理实际的特点,能够为城市内涝的应急响应与灾害评估提供可靠的数据支持。

[0104] It should be noted that there is an unclear part in the original text where it says "剔除所有关联坡度值大于预设阈值(如5°)的积水像素。这些区域在物理上不可能形成积水,因此将其从识别结果中剔除。" which has been translated as best as possible while maintaining the original meaning, but the original text seems a bit incomplete in this regard. If you can provide more context or clarify this part, the translation can be further refined.Throughout the processing flow, data transfer between modules is achieved through two-dimensional arrays or image formats in memory. Classifier model parameters are pre-stored in the electronic device's memory, and the processor sequentially calls each layer of classifiers to process the input data during runtime. The spatial resolution of the UAV multispectral imagery is set to 0.1 meters per pixel, with spectral bands including blue, green, red, near-infrared, and two water-sensitive bands such as 900 nm and 1000 nm, ensuring sensitivity to small water accumulation areas. Standard water body spectral feature vectors are jointly calibrated through laboratory measurements and field sampling, covering water reflectance characteristics under different turbidity, depth, and illumination conditions to enhance the generalization ability of the validation layer. This method, through the collaborative work of multi-level classifiers, fully utilizes the spatial details and spectral features of high-resolution UAV imagery, effectively improving the accuracy and adaptability of water accumulation area detection, and is particularly suitable for rapid response and disaster assessment in dynamic and complex scenarios such as post-rain urban flooding.

[0105] This embodiment provides a spatial distribution identification system for accumulated water based on multi-classifier collaboration. This system corresponds to the method embodiment and is used to perform the steps in the above method embodiment. Please refer to the appendix. Figure 9 The system includes:

[0106] Classifier library construction module: used to build a hierarchical classifier library, which includes at least an initial screening layer, an arbitration layer, a re-identification layer and a verification layer, and each layer contains at least one classifier.

[0107] The first recognition and fusion module is used to acquire UAV images, classify and recognize the UAV images using the initial screening layer, and fuse the recognition results of each classifier in the initial screening layer to determine suspicious areas.

[0108] The second identification and fusion module is used to classify and identify suspicious areas using the arbitration layer, and to fuse the identification results of each classifier in the arbitration layer to obtain the first high-confidence water accumulation area and the first ambiguous area.

[0109] The third recognition and fusion module is used to classify and recognize the first ambiguous region using the re-recognition layer, and to fuse the recognition results of each classifier in the re-recognition layer to obtain the second high-confidence water accumulation region and the second ambiguous region.

[0110] The fourth recognition and fusion module is used to classify and recognize the second fuzzy region using the verification layer, and to fuse the recognition results of each classifier in the verification layer to obtain the third high-confidence water accumulation region.

[0111] Waterlogged Area Merging Module: This module merges the first, second, and third high-confidence waterlogged areas to obtain a preliminary waterlogged distribution map.

[0112] Weighted Fusion Optimization Module: This module is used to perform weighted fusion optimization on the initial water accumulation distribution map. It calculates the confidence level of each pixel based on the preset weights of each level classifier and removes low-confidence pixels based on a preset threshold to obtain the optimized water accumulation distribution map.

[0113] The results output module is used to post-process the optimized water accumulation distribution map and output the final spatial distribution result of the water accumulation.

[0114] The specific implementation methods of each module correspond one-to-one with the steps described in the method embodiment, and will not be repeated here.

[0115] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for identifying waterlogged space distribution based on multi-classifier cooperation, characterized in that, include: Construct a hierarchical classifier library, which includes at least an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer, with each layer containing at least one classifier; Acquire drone images, use a preliminary screening layer to classify and identify the drone images, and fuse the identification results of each classifier in the preliminary screening layer to determine suspicious areas; The suspicious areas are classified and identified by the arbitration layer, and the identification results of each classifier in the arbitration layer are merged to obtain the first high-confidence water accumulation area and the first ambiguous area. The first ambiguous region is classified and identified by the re-identification layer, and the identification results of each classifier in the re-identification layer are fused to obtain the second high-confidence water accumulation region and the second ambiguous region. The second ambiguous region is classified and identified by the verification layer, and the identification results of each classifier in the verification layer are fused to obtain the third high-confidence water accumulation region. By merging the first, second, and third high-confidence water accumulation areas, a preliminary water accumulation distribution map is obtained; The preliminary water accumulation distribution map is weighted and fused for optimization. The confidence level of each pixel is calculated according to the preset weights of each level classifier, and low-confidence pixels are removed based on a preset threshold to obtain the optimized water accumulation distribution map. The optimized water accumulation distribution map is post-processed to output the final spatial distribution result of the water accumulation. 2.The waterlogged space distribution recognition method based on multi-classifier cooperation according to claim 1, characterized in that, The initial screening layer includes a minimum distance classifier, a parallelepiped classifier, a spectral angle mapping classifier, and a maximum likelihood classifier; The identification results of each classifier in the fusion initial screening layer determine suspicious regions, including: Perform a pixel-by-pixel logical OR operation on the recognition results of each classifier in the initial screening layer; If any classifier determines that a pixel is filled with water, then that pixel is marked as a suspicious area. 3.The waterlogged space distribution recognition method based on multi-classifier cooperation according to claim 1, characterized in that, The arbitration layer includes an adaptive coherence estimation classifier, a constrained energy minimization classifier, and a Mahalanobis distance classifier; The recognition results of each classifier in the fusion arbitration layer are used to obtain the first high-confidence water accumulation region and the first ambiguous region, including: Perform a pixel-by-pixel logical AND operation on the recognition results of each classifier in the arbitration layer; If all classifiers determine that a pixel is a water accumulation area, then the pixel is marked as the first high-confidence water accumulation area; Pixels within the suspicious area, excluding the first high-confidence water accumulation area, are marked as the first blurred area. 4.The method of claim 1, wherein, The re-identification layer includes a neural network classifier and a support vector machine classifier; The recognition results of each classifier in the fusion re-identification layer are used to obtain the second high-confidence water accumulation region and the second blurred region, including: Perform a pixel-by-pixel logical AND operation on the recognition results of each classifier in the re-recognition layer; If all classifiers determine that a pixel is a water accumulation area, then mark that pixel as the second high-confidence water accumulation area; Pixels within the first blurred region, excluding the second high-confidence water accumulation region, are marked as the second blurred region. 5.The method of claim 1, wherein, Before using the verification layer to classify and identify the second fuzzy region, the following steps are also included: Determine whether the area of ​​the second blurred region is greater than a preset threshold; If the value exceeds the preset threshold, the verification layer will be activated for classification and identification. If the value is less than or equal to the preset threshold, the verification layer is skipped. 6.The method of claim 1, wherein, The verification layer includes an orthogonal subspace projection classifier and a spectral information divergence classifier; the classification and identification of the second fuzzy region using the verification layer includes: Obtain the preset standard water body spectral characteristics; The spectral vector of each pixel within the second blurred region is compared with the spectral features of the standard water body to calculate a similarity metric. Based on the comparison result between the similarity metric and the preset threshold, it is determined whether the pixel is water accumulation; The classifier in the verification layer does not rely on training samples during the discrimination process. 7.The waterlogged space distribution recognition method based on multi-classifier cooperation according to claim 1, wherein, The preliminary waterlogging distribution map was weighted and optimized, including: Assign preset weights to each classifier in the initial screening layer, arbitration layer, re-identification layer, and verification layer; For each pixel, the confidence level of that pixel is calculated based on the weight of the layer in which the classifier that classifies it as water accumulation belongs. Remove pixels with a confidence level below a preset threshold from the preliminary water accumulation distribution map. 8.The method of claim 1, wherein, The confidence score of each pixel is calculated based on the preset weights of each level of classifiers, including: For each pixel, count the number of classifiers at each level that classify it as water accumulation; The weights of the classifiers at each level that classifies a pixel as having water accumulation are summed to obtain the weighted confidence score for that pixel. The preset threshold is a constant set based on empirical values ​​or experimental data. 9.The method of claim 1, wherein, The post-processing includes: Morphological closure operation is performed on the optimized water accumulation distribution map; The digital elevation model and land cover type map are acquired, and pixels located in areas with slopes greater than a preset threshold and on the top surface of buildings are identified and removed to obtain the final spatial distribution result of water accumulation. 10.A waterlogged space distribution recognition system based on multi-classifier cooperation, characterized in that, include: A classifier library construction module is used to build a hierarchical classifier library, which includes at least an initial screening layer, an arbitration layer, a re-identification layer, and a verification layer, with each layer containing at least one classifier. The first recognition and fusion module is used to acquire UAV images, classify and recognize UAV images through a preliminary screening layer, and fuse the recognition results of each classifier in the preliminary screening layer to determine suspicious areas. The second identification and fusion module is used to classify and identify suspicious areas through the arbitration layer, and fuse the identification results of each classifier in the arbitration layer to obtain the first high-confidence water accumulation area and the first ambiguous area. The third recognition and fusion module is used to classify and recognize the first blurred region through the re-recognition layer, and fuse the recognition results of each classifier in the re-recognition layer to obtain the second high-confidence water accumulation region and the second blurred region. The fourth recognition and fusion module is used to classify and recognize the second ambiguous region through the verification layer, and fuse the recognition results of each classifier in the verification layer to obtain the third high-confidence water accumulation region. The water accumulation area merging module is used to merge the first, second, and third high-confidence water accumulation areas to obtain a preliminary water accumulation distribution map. The weighted fusion optimization module is used to perform weighted fusion optimization on the preliminary water accumulation distribution map. It calculates the confidence level of each pixel based on the preset weights of each level classifier and removes low-confidence pixels based on a preset threshold to obtain the optimized water accumulation distribution map. The results output module is used to post-process the optimized water accumulation distribution map and output the final spatial distribution result of the water accumulation.

Images