A method for remote sensing extraction of offshore floating rafts and deep water farming areas

Abstract

The present application provides a kind of offshore floating raft and deep water aquaculture area remote sensing extraction method, comprising the following operating steps: S1, constructs and optimizes depth random rotating forest model;S2, based on the remote sensing data multi-feature extraction of aquaculture area identification.S1 mainly includes two aspects: one is based on the feature combination random forest, improves mixed node split, improves forest scale by multi-input adaptive model;Two is to combine multi-granularity scanning and deep cascade, starting from multi-granularity scanning feature, through parallel multiple rotating forest scanning, the context information of spatial structure and texture attribute is optimized, and the optimization result of multi-granularity scanning is multi-level series again, to generate a deep random rotating forest with strong generalization ability.The present application maximizes the recognition ability of depth rotating forest to typical target in high-resolution remote sensing image, can improve the efficiency and precision of seawater aquaculture area automatic identification.

Description

Technical Field

[0001] This invention relates to a remote sensing image extraction method, specifically a remote sensing image extraction method for nearshore floating rafts and deep-water aquaculture areas. Background Technology

[0002] Aquaculture is a typical human activity in coastal areas. Over the past decade or so, the global aquaculture industry has been expanding. Therefore, the rapid and accurate extraction of the distribution and area of ​​nearshore floating raft aquaculture areas provides fisheries management departments with decision-making information and scientific basis for rationally planning aquaculture use, controlling aquaculture density, curbing environmental degradation, and preventing aquaculture diseases.

[0003] Current research indicates that, given the complex water color environment near the coast, higher spatial resolution requirements are placed on remote sensing imagery for accurately extracting nearshore floating raft aquaculture areas against complex water color backgrounds. To meet these requirements, new remote sensing information extraction methods need to be designed. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a remote sensing extraction method for nearshore floating rafts and deep-water aquaculture areas, which fully utilizes the advantages of deep cascaded forests and rotating forests, and can efficiently and accurately extract nearshore aquaculture floating raft aquaculture areas and deep-water aquaculture areas.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a remote sensing extraction method for nearshore floating rafts and deep-water aquaculture areas, characterized by comprising the following operational steps:

[0006] S1. Input high-resolution remote sensing images and extract texture features;

[0007] S2. Extract local autocorrelation spatial structure features and establish a spatial context information model;

[0008] S3. Optimize spectral features based on principal component transformation and independent component transformation of multispectral information;

[0009] S4. Perform cross-sampling and sparse representation reconstruction on the texture features, spatial structure features and optimized spectral features in S1-S3, and generate new features through repeated sampling, representation and reconstruction.

[0010] S5. Construct a deep cascaded rotating forest model, and use multi-granularity scanning and forest cascade parallel processing to deeply express spectral features, spatial structure features and texture features to obtain deep information in the image.

[0011] S6. Based on the parallel connection, the rotating forest model is progressively connected in series and then updated using a deep adaptive iterative optimization method to achieve high-precision extraction of remote sensing information.

[0012] Preferably, the specific method of S1 is as follows:

[0013] The wavelet transform is improved into an adaptive multi-scale transform. Based on the opening and closing operations of mathematical morphology, mathematical morphological sequences at different scales are extracted to reconstruct profiles, fully exploring the texture information of remote sensing images. The specific formula is as follows:

[0014] Equation (1)

[0015] Equation (2)

[0016] Equation (3)

[0017] Equation (4)

[0018] Equations (1) and (2) represent the set theory expression of the erosion and dilation operation of B on A, which is the basis of the mathematical morphological filtering filling operation. The mathematical model for measuring erosion and dilation is shown in Equations (3) and (4). In the formulas, S is the labeled image and T is the template image. When n=0, D(S)=S and E(S)=S. The mathematical morphological reconstruction of measuring erosion and dilation is achieved through the above iteration.

[0019] Preferably, the specific method of S2 is as follows:

[0020] Based on spatial autocorrelation, neighborhood analysis is enhanced, and the local spatial autocorrelation formula is improved as follows:

[0021] Equation (5)

[0022] Equation (6)

[0023] Equation (7)

[0024] In equation (5), x i It is the attribute value of spatial unit i, w ij The spatial weight matrix represents the degree of influence between spatial units i and j; I i It is the MORAN index, with a value range of [-1, 1]. A positive value indicates that the spatial cell has similar attribute values ​​to its neighboring cells, and the spatial autocorrelation is positive; a negative value indicates that the spatial cell has dissimilar attribute values ​​to its neighboring cells, and the spatial autocorrelation is negative; 0 indicates that there are no spatially correlated attributes.

[0025] Preferably, the specific operations of S4-S6 are as follows:

[0026] S01, Sample set Perform random sampling without replacement, and divide the entire dataset evenly using this method to obtain... A training set. The training set, partitioned using a sampling method without replacement, effectively improves the diversity of base classifiers. Furthermore... The training set has feature, ;

[0027] S02, to To prevent base classifiers from selecting the same training set for training, thus increasing inter-class differences between base classifiers, a subset of the training set is selected for PCA feature transformation. The method for selecting this subset is as follows: The training set is resampled by 75% to produce a training subset. in The base classifier number, Number the training subset, for this PCA feature transformation is performed on a subset of training data, where the generated principal component coefficients are: , length is At the same time, features whose feature vectors are 0 are removed;

[0028] S03, Rotation Matrix Processing: Obtained through multiple steps S02 A training subset is used to generate principal component coefficients that are input to a sparse rotation matrix. :

[0029] Equation (8)

[0030] Rotation matrix The dimension is After rotation matrix After the process is completed, a base classifier can be generated. training dataset The specific operation process of rotating a matrix involves adjusting the array... The column vectors are arranged to correspond to the order of the original feature set, and the rotation matrix is ​​adjusted accordingly. Size is When testing the final prediction results, the test dataset is used. ,Will Dot product Input base classifier Chinese Representation base classifier predict Belongs to class The probability, and calculate Confidence level belonging to each category :

[0031] Equation (9);

[0032] S04. Compare the confidence scores of each category and obtain the category with the highest confidence score, which is the final category.

[0033] Preferably, the specific method of S04 is as follows: assuming that there are 1000 (hyperparameter) completely random trees in each completely random tree forest in the cascaded structure, a feature from the data is randomly selected as a discrimination condition for segmentation, and child nodes are generated according to this discrimination condition until each leaf node in the forest has only instances of the same class left or the number of instances does not exceed 10. Multi-granularity scanning is used to preprocess the input image for features. A sampling sliding window obtains the original features through scanning.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] 1. This invention employs resampling and reconstruction, multi-granularity scanning, and parallel processing methods to deeply and adaptively iteratively optimize the processing of spectral, spatial structure, texture, and contextual information, updating the model to achieve high-precision extraction of nearshore floating raft aquaculture areas. Given the complex nearshore water quality environment surrounding aquaculture areas, characterized by severe "same object, different spectrum" and "different objects, same spectrum" phenomena, this invention utilizes multi-scale wavelet transform, morphological opening and closing operations, and reconstruction operations from the field of pattern recognition to extract spatial structural features, establishing corresponding texture feature decomposition scale models and contextual information models. This maximizes the ability of deep rotating forests to identify typical targets in high-resolution remote sensing images, improving the efficiency and accuracy of automated identification of marine aquaculture areas.

[0036] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0038] Figure 2 This is a diagram of the architecture of the rotating forest model in this invention.

[0039] Figure 3 This is a schematic diagram of class vector generation in this invention.

[0040] Figure 4 This is a schematic diagram of the deep rotating forest cascade structure in this invention.

[0041] Figure 5 This is a schematic diagram of the nearshore floating raft aquaculture extraction results obtained by various algorithms in this embodiment.

[0042] Figure 6This is a schematic diagram of the nearshore deep-sea aquaculture extraction results obtained by various algorithms in this embodiment. Detailed Implementation

[0043] like Figure 1 and Figure 2 As shown, the present invention includes the following operational steps:

[0044] S1. Input the high-resolution remote sensing image of Jilin-1 and extract the texture features. Calculate the spatial correlation index of MORAN, GEARY and GETIS-ORD in the study area using Rook's case proximity rule statistics, and convert it into a gray level of G=256.

[0045] S2. Extract local autocorrelation spatial structure features and establish a spatial context information model; segment the spatial index and perform spatial analysis using morphology, calculating and extracting regions with positive and negative autocorrelation with building type height. Remove isolated small-area objects and use mathematical morphology to perform growth calculations on meaningful objects. Perform intersection operations on the optimized object graphics to identify and extract regions with both positive and negative autocorrelation.

[0046] S3. Optimize spectral features based on principal component transformation and independent component transformation of multispectral information: Optimize spectral information by performing principal component transformation and independent component transformation on multispectral information; for N-dimensional random variables, traditional principal component transformation involves finding all directions W1, ..., W... n This maximizes the variance. Principal component transformation (PCT) and independent component transformation (ICT) require performing PCT first, then labeling the resulting components to obtain the final result. PCT reduces noise or irrelevant data, retaining the principal component signals.

[0047] S4. Cross-sampling and sparse representation reconstruction are performed on the texture features, spatial structure features and optimized spectral features in S1-S3. Through repeated sampling, expression and reconstruction, new features of Jilin No. 1 that are beneficial for extraction from aquaculture areas are generated.

[0048] S5. Further obtain a new feature set F' by multi-granular scanning of the training set obtained through S4. Construct a deep rotating forest suitable for Jilin-1 high-resolution images by multi-layer cascading and iterative deepening of the network through rotating forest.

[0049] S6. Based on the parallel connection, the rotating forest model is progressively connected in series and then updated using a deep adaptive iterative optimization method to achieve high-precision extraction of remote sensing information.

[0050] In this embodiment, the specific method of S1 is as follows:

[0051] The wavelet transform is improved into an adaptive multi-scale transform. Based on the opening and closing operations of mathematical morphology, mathematical morphological sequences at different scales are extracted to reconstruct profiles, fully exploring the texture information of remote sensing images. The specific formula is as follows:

[0052] Equation (1)

[0053] Equation (2)

[0054] Equation (3)

[0055] Equation (4)

[0056] Equations (1) and (2) represent the set theory expression of the erosion and dilation operation of B on A, which is the basis of the mathematical morphological filtering filling operation. The mathematical model for measuring erosion and dilation is shown in Equations (3) and (4). In the formulas, S is the labeled image and T is the template image. When n=0, D(S)=S and E(S)=S. The mathematical morphological reconstruction of measuring erosion and dilation is achieved through the above iteration.

[0057] In this embodiment, the specific method of S2 is as follows:

[0058] Based on spatial autocorrelation, neighborhood analysis is enhanced, and the local spatial autocorrelation formula is improved as follows:

[0059] Equation (5)

[0060] Equation (6)

[0061] Equation (7)

[0062] In equation (5), x i It is the attribute value of spatial unit i, w ij The spatial weight matrix represents the degree of influence between spatial units i and j; I i It is the MORAN index, with a value range of [-1, 1]. A positive value indicates that the spatial cell has similar attribute values ​​to its neighboring cells, and the spatial autocorrelation is positive; a negative value indicates that the spatial cell has dissimilar attribute values ​​to its neighboring cells, and the spatial autocorrelation is negative; 0 indicates that there are no spatially correlated attributes.

[0063] In this embodiment, the specific operations of S4-S6 are as follows:

[0064] S01, Sample set Perform random sampling without replacement, and divide the entire dataset evenly using this method to obtain... A training set. The training set, partitioned using a sampling method without replacement, effectively improves the diversity of base classifiers. Furthermore... The training set has feature, ;

[0065] S02, to To prevent base classifiers from selecting the same training set for training, thus increasing inter-class differences between base classifiers, a subset of the training set is selected for PCA feature transformation. The method for selecting this subset is as follows: The training set is resampled by 75% to produce a training subset. in The base classifier number, Number the training subset, for this PCA feature transformation is performed on a subset of training data, where the generated principal component coefficients are: , length is At the same time, features whose feature vectors are 0 are removed;

[0066] S03, Rotation Matrix Processing: Obtained through multiple steps S02 A training subset is used to generate principal component coefficients that are input to a sparse rotation matrix. :

[0067] Equation (8)

[0068] Rotation matrix The dimension is After rotation matrix After the process is completed, a base classifier can be generated. training dataset The specific operation process of rotating a matrix involves adjusting the array... The column vectors are arranged to correspond to the order of the original feature set, and the rotation matrix is ​​adjusted accordingly. Size is When testing the final prediction results, the test dataset is used. ,Will Dot product Input base classifier Chinese Representation base classifier predict Belongs to class The probability, and calculate Confidence level belonging to each category :

[0069] Equation (9);

[0070] S04. Compare the confidence scores of each category and obtain the category with the highest confidence score, which is the final category.

[0071] In this embodiment, the specific method of S04 is as follows: assuming that there are 1000 (hyperparameter) completely random trees in each completely random tree forest in the cascaded structure, a feature from the data is randomly selected as a discrimination condition for segmentation, and child nodes are generated according to this discrimination condition, until each leaf node in the forest has only instances of the same class left or the number of instances does not exceed 10. The process of generating the class distribution vector of each layer of the rotating forest is as follows. Figure 3 As shown.

[0072] Multi-granularity scanning is used to preprocess the input image for features. A sampling sliding window acquires the original features through scanning. For example... Figure 4 As shown, the sampling sliding window acquires the original features through scanning. For a 400-dimm (dimensional) sequence of image data, a 100-dimm sampling window is used for sliding sampling to process the input image's features. This generates 301 feature vectors of size 100-dimm. These acquired feature vectors (or feature maps) are then sequentially input into a rotating forest A and a rotating forest B, respectively.

[0073] Table 1, Table 2, Figure 5 and Figure 6 This section compares the performance of various algorithms with that of the algorithm proposed in this invention.

[0074] Table 1 Comparison of recognition results between the invented algorithm and existing algorithms (results extracted from floating raft aquaculture)

[0075]

[0076] Table 2 Comparison of recognition results between the invented algorithm and existing algorithms (results extracted from deep-sea aquaculture)

[0077]

[0078] It is evident that the algorithm provided by this invention has higher accuracy and a higher Kappa coefficient, making it more suitable for remote sensing extraction in nearshore floating rafts and deep-sea aquaculture areas.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Any simple modifications, alterations, and equivalent changes made to the above embodiments based on the inventive essence shall still fall within the protection scope of the present invention.

Claims

1. A remote sensing extraction method for nearshore floating rafts and deep-water aquaculture areas, characterized in that, The following steps are included: S1. Input high-resolution remote sensing images and extract texture features; S2. Extract local autocorrelation spatial structure features and establish a spatial context information model; S3. Optimize spectral features based on principal component transformation and independent component transformation of multispectral information; S4. Perform cross-sampling and sparse representation reconstruction on the texture features, spatial structure features and optimized spectral features in S1-S3, and generate new features through repeated sampling, representation and reconstruction. S5. Construct a deep cascaded rotating forest model, and use multi-granularity scanning and forest cascade parallel processing to deeply express spectral features, spatial structure features and texture features to obtain deep information in the image. S6. On the basis of parallel connection, the rotating forest model is progressively connected in series and then updated by a deep adaptive iterative optimization method to achieve high-precision extraction of remote sensing information. The specific operations of S4-S6 are as follows: S01, Sample set Perform random sampling without replacement, and divide the entire dataset evenly using this method to obtain... A training set; the training set partitioned using a non-replacement sampling method can effectively improve the diversity of base classifiers; and The training set has Conquest, ; S02, to To prevent base classifiers from selecting the same training set for training, thus increasing inter-class differences between base classifiers, a subset of the training set is selected for PCA feature transformation. The method for selecting this subset is as follows: The training set is resampled by 75% to produce a training subset. in The base classifier number, Number the training subset, for this PCA feature transformation is performed on a subset of training data, where the generated principal component coefficients are: , length is At the same time, features whose feature vectors are 0 are removed; S03, Rotation Matrix Processing: Obtained through multiple steps S02 A training subset is used to generate principal component coefficients that are input to a sparse rotation matrix. : Equation (8) Rotation matrix The dimension is After rotation matrix After the process is completed, a base classifier is generated. training dataset The specific operation process of rotating a matrix involves adjusting the array... The column vectors are arranged to correspond to the order of the original feature set, and the rotation matrix is ​​adjusted accordingly. Size is When testing the final prediction results, the test dataset is used. ,Will Dot product Input base classifier Chinese Representation base classifier predict Belongs to class The probability, and calculate Confidence level belonging to each category : =1, 2, ..., Equation (9); S04. Compare the confidence scores of each category and obtain the category with the highest confidence score. The final category.

2. The remote sensing extraction method for nearshore floating rafts and deep-water aquaculture areas according to claim 1, characterized in that, The specific method of S1 is as follows: The wavelet transform is improved into an adaptive multi-scale transform. Based on the opening and closing operations of mathematical morphology, mathematical morphological sequences at different scales are extracted to reconstruct profiles, fully exploring the texture information of remote sensing images. The specific formula is as follows: Equation (1); Equation (2); Equation (3); Equation (4); Equations (1) and (2) represent the set theory expression of the erosion and dilation operation of B on A, which is the basis of the mathematical morphological filtering filling operation. The mathematical model for measuring erosion and dilation is shown in Equations (3) and (4). In the formulas, S is the labeled image and T is the template image. When n=0, D(S)=S and E(S)=S. The mathematical morphological reconstruction of measuring erosion and dilation is achieved through the above iteration.

3. The remote sensing extraction method for nearshore floating rafts and deep-water aquaculture areas according to claim 1, characterized in that, The specific method of S2 is as follows: Based on spatial autocorrelation, neighborhood analysis is enhanced, and the local spatial autocorrelation formula is improved as follows: Equation (5); Equation (6); Equation (7); In equation (5), x i It is the attribute value of spatial unit i, w ij The spatial weight matrix represents the degree of influence between spatial units i and j; I i It is the MORAN index, with a value range of [-1, 1]. A positive value indicates that the spatial cell has similar attribute values ​​to its neighboring cells, and the spatial autocorrelation is positive; a negative value indicates that the spatial cell has dissimilar attribute values ​​to its neighboring cells, and the spatial autocorrelation is negative; 0 indicates that there are no spatially correlated attributes.

Images