Raft culture area extraction method based on high-resolution optical image
By employing a raft aquaculture area extraction method based on high-resolution optical imagery, and utilizing discrete wavelet transform and iterative training of the SAM model, combined with genetic algorithm optimization of topology rules, the problem of automated aquaculture area extraction in traditional methods is solved, achieving high-precision and low-cost aquaculture area extraction.
Patent Information
- Application Number
- CN202511140110.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-14
- Publication Date
- 2025-11-25
AI Technical Summary
Traditional methods are difficult to extract raft aquaculture areas efficiently and automatically, especially due to the difficulty in distinguishing between the aquaculture area and the seawater background in high-resolution optical images, and deep learning models require a lot of manual annotation, which is costly.
A method for extracting raft aquaculture areas based on high-resolution optical imagery is adopted. This method integrates RGB and panchromatic images through discrete wavelet transform, uses the SAM model for annotation-guided iterative training, and combines genetic algorithms to optimize topology rules, thereby automatically extracting aquaculture areas.
It achieves high-precision, low-cost automated extraction of aquaculture areas, adapts to changes in aquaculture areas at different times, reduces the need for manual annotation, and improves the versatility of data.
Smart Images
Figure CN121010773A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of intelligent maritime affairs, in particular, especially relates to a raft culture area extraction method based on high-resolution optical image. BACKGROUND
[0002] In the past three decades, according to the data of the United Nations Food and Agriculture Organization, China's aquaculture has developed rapidly, and the output ranks first in the world, accounting for more than 60% of the world's aquaculture area and output. Raft culture in shallow sea area is the main way of aquaculture in China, which has developed rapidly due to its high economic value and high income. Although the industry has made a significant contribution to the supply of aquatic products, there are various problems such as poor layout planning and blind expansion of aquaculture. Economic benefits have driven the uncontrolled development of raft culture in many aspects, and it has entered an unordered development stage, which has a long-term impact on marine ecology and coastal shipping.
[0003] The monitoring method of raft culture area mainly depends on field investigation, high-resolution optical remote sensing image and synthetic aperture radar image. However, the traditional field investigation method is time-consuming and laborious, and it is difficult to meet the needs of large-area raft culture area identification and dynamic monitoring. Satellite remote sensing technology has the advantages of short detection period, strong real-time performance, low cost and large coverage, and has become an important means of raft culture extraction. However, synthetic aperture radar image is affected by coherent wave interference, and there are a large number of speckle noises and complex data processing process, which affects the classification accuracy of ground objects. In addition, high-resolution optical image not only has clear detail information, but also has rich spectral information. In addition, the data processing of high-resolution optical image is simple, and the ground object features are clear, which is easier for artificial understanding.
[0004] Raft culture area has periodicity in shape. Raft culture area is mainly composed of floating raft at the edge of culture area block and connecting rope, and the culture object is attached to the rope and mostly plant. Because the growth of plants and the harvest of farmers have periodicity, when the plants grow in the middle and late stages, the obvious difference between the culture area and the sea water background can be observed in the remote sensing image, especially in the mature period before harvest; after harvest, only floating raft and rope are left in the culture area, which is difficult to distinguish from the sea water background. This periodicity greatly limits the extraction of raft culture area, while the invariability of culture net in the whole cycle provides a new idea for the extraction of culture area, and high-resolution optical image provides data support.
[0005] In recent years, as an important branch of machine learning, deep learning has made great progress in the field of computer vision, and has excellent performance in tasks such as target recognition and image segmentation due to its powerful feature learning and nonlinear modeling capabilities. Although it requires large-scale data and powerful computing power to meet the training requirements, its task accuracy is close to or even surpasses the level of artificial intelligence, making it a huge potential. At the same time, with the development of "large model + transfer learning", the application difficulty of deep learning is greatly alleviated, among which the SAM model is particularly representative. SUMMARY
[0006] Therefore, the purpose of the present application is to provide a raft culture area extraction method based on high-resolution optical images, which aims to extract culture nets in an automated manner and then gradually achieve the automatic extraction of raft culture areas.
[0007] The technical means adopted by the present application are as follows: A raft culture area extraction method based on high-resolution optical images, comprising the following steps: S1, in the image library, screening out the pre-processed high-resolution optical image of the target area, using discrete wavelet transform to fuse R, G, B and panchromatic bands to obtain a high-resolution RGB fusion image; S2, using non-subsampled contourlet transform to enhance the RGB fusion image and the panchromatic image; S3, constructing a culture net extraction rule to segment the panchromatic image in S2 to obtain an initial culture net mask; S4, inputting the enhanced RGB fusion image and the initial culture net mask into the SAM model for annotation-guided prompt training, that is, generating a culture net prompt mask in the SAM model using the initial culture net mask, segmenting the culture net prompt mask to generate a culture net candidate mask, calculating the cosine similarity of the culture net prompt mask and the candidate mask, retaining the high similarity area as a pseudo culture net mask, and calculating the loss of the initial culture net mask and the pseudo culture net mask to optimize the SAM model; S5, evaluating the initial culture net mask and the pseudo culture net mask in S4, if the evaluation accuracy does not meet the standard, using the pseudo culture net mask as the initial culture net mask to re-execute S4 for iteration; if the evaluation accuracy meets the standard, outputting the pseudo culture net mask as the culture net mask to execute S6; S6, extracting the mask block center point of the culture raft mask, constructing a culture net topology rule to connect the center points to obtain a raft culture area; interpreting the culture net mask center points connected with a small amount of culture areas to obtain a raft culture area verification sample; based on the verification sample, using a genetic algorithm to automatically tune the topology rule to achieve the optimal effect.
[0008] Preferably, in S1, the image fusion of the R, G, and B bands with the panchromatic band includes the following steps: S11. Using a biorthogonal wavelet basis (bior4.4), perform N-level decomposition on the R, G, B, and panchromatic bands respectively, and obtain the approximate coefficients for each level. ) and detail factor ( , , The decomposition formula is as follows:
[0009]
[0010]
[0011]
[0012] in, k It is the number of decomposition levels. It is a two-dimensional scaling function of Bior 4.4. , , These are the horizontal, vertical, and diagonal wavelet functions of Bior 4.4. It is the original image. Represents downsampling, Represents discrete convolution operation. col, row Represents the convolution direction; S12. Use fusion rules to merge the detail information of the panchromatic band with the spectral information of the R, G, and B bands. This is represented by the fusion of approximation coefficients and detail coefficients. The fusion formula is as follows:
[0013]
[0014] in, These are the approximate coefficients after fusion. It is the detail coefficient after fusion. , , , These are the approximation coefficients for the panchromatic, R, G, and B bands, respectively. It is the detail coefficient of the panchromatic band. It is the detail coefficient of any band of R, G, or B. These are weighting coefficients; S13. Starting from the Nth layer, reconstruct the structure layer by layer upwards until the resolution is restored. The reconstruction formula is as follows:
[0015] in,LL These are approximate coefficients. , It's the reconstruction filter for Bior 4.4. This represents upsampling.
[0016] Furthermore, in S3, the extraction rules for aquaculture feed are as follows: S31. By statistically analyzing the pixel values, shape, and area features of the aquaculture baskets on the panchromatic image, the following rules for extracting the aquaculture baskets are constructed:
[0017] in, It is the initial aquaculture membrane. , , These are the pixel values, aspect ratio, and area of the aquaculture basket. A, B, C It refers to the range of characteristic statistics.
[0018] Furthermore, in S4, the cue training steps for the SAM model are as follows: S41. Input the initial aquaculture mat mask from S3 into the trained SAM model for annotation-guided cue training, then:
[0019] in, It is the initial SAM model. It's a fake aquaculture mat. It is the initial aquaculture mat mask input. These are the parameters of the cross-attention layer between the cue encoder and mask decoder in the SAM model. L It is the loss function.
[0020] L The mask loss and the IoU prediction loss are weighted and summed as follows:
[0021] in, , These are the mask loss and IoU Predicting losses, It's the weight.
[0022] S42, It is achieved by calculating the cosine similarity between the candidate mask and the breeding field hint mask. The result, obtained after filtering highly similar regions, is calculated using the following formula:
[0023] in, It's a mask indicating the aquaculture area. It is a candidate mask.
[0024] Furthermore, the iteration conditions in S5 are as follows: S51. Calculate the initial culture bed mask and the pseudo culture bed mask. IoU , Dice The coefficient assesses the degree of overlap between the two, and the calculation formula is as follows:
[0025]
[0026] in, A, B These are the initial culture pan mask and the pseudo culture pan mask.
[0027] S52. Repeat S41, S42, and S51, continuously updating the model and the pseudo-culture mat mask, until the... k In the next iteration, we have:
[0028] S53. Set the iteration termination threshold to... The stopping condition for iteration is:
[0029] S54. Use the pseudo-culture floor mask that stops iterating when it meets the evaluation accuracy as the culture floor mask and execute S6.
[0030] Furthermore, in S6, the constructed aquaculture belt topology rules are as follows: S61. Baseline Extension Rule: Given a base point P ,search r Within distance ( n (Points = 1, 2, 3...) connect ;by As a new starting point, with and Direction Angle range and r Search for the next point within the distance range; continue this process, extending the baseline to the point where no points exist within the search range; iterate through all points not included in the baseline, executing the connection rules described above; based on the baseline length, discard points with fewer than [number missing] connection points in a single straight line trend. n The baseline for each.
[0031] S62. Baseline Closure Rule: Given the endpoints of an unclosed baseline E The angle between the lines of the search trend and the line of intersection is greater than 1. And the nearest other unclosed baseline endpoint FExtend the two baselines until they intersect to complete the baseline closure; and so on, traversing all the endpoints of the unclosed baselines to complete the baseline closure.
[0032] S63. Connect a small number of aquaculture plots through visual interpretation, create a validation set for the aquaculture area, and use a genetic algorithm to automatically optimize the parameters of the aquaculture plot topology rules.
[0033] S63. Connect a small number of aquaculture plots through visual interpretation, create a validation set for the aquaculture area, and use a genetic algorithm to automatically optimize the parameters of the aquaculture plot topology rules.
[0034] This application discloses a method for extracting raft aquaculture areas based on high-resolution optical imagery. Regarding the automatic extraction of aquaculture rafts: First, extraction rules are constructed to initially extract the rafts. Then, a pre-trained SAM model is used for annotation-guided training, iteratively obtaining the complete raft. This method has high accuracy and avoids extensive manual sample annotation, reducing model training costs. In terms of the spatiotemporal versatility of aquaculture area extraction: This invention first extracts the rafts, then constructs topological rules to connect them, and obtains the aquaculture area after closure processing. This method avoids the periodic heterogeneity of aquaculture area morphology in spatiotemporal terms, possessing higher data versatility and adapting to aquaculture area extraction tasks at any time period. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 The flowchart of the method of the present invention Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0037] like Figure 1 The method for extracting raft aquaculture areas based on high-resolution optical images, as shown, includes the following steps: S1. In the image library, select the high-resolution optical images of the target area after preprocessing, and use discrete wavelet transform to fuse the R, G, B bands and the panchromatic band to obtain a high-resolution RGB fused image; the preprocessing includes radiometric calibration, atmospheric correction and other steps.
[0038] In the remote sensing image database, 30 pre-processed optical images of a certain area of the sea from January to December were selected from the GF-2 satellite PMS sensor.
[0039] The image fusion process is as follows: S11. Using a biorthogonal wavelet basis (bior4.4), perform N-level decomposition on the R, G, B, and panchromatic bands respectively, and obtain the approximate coefficients for each level. ) and detail factor ( , , The decomposition formula is as follows:
[0040]
[0041]
[0042]
[0043] in, k It is the number of decomposition levels. It is a two-dimensional scaling function of Bior 4.4. , , These are the horizontal, vertical, and diagonal wavelet functions of Bior 4.4. It is the original image. Represents downsampling, Represents discrete convolution operation. col, row Represents the convolution direction; S12. Use fusion rules to merge the detail information of the panchromatic band with the spectral information of the R, G, and B bands. This is represented by the fusion of approximation coefficients and detail coefficients. The fusion formula is as follows:
[0044]
[0045] in, These are the approximate coefficients after fusion. It is the detail coefficient after fusion. , , , These are the approximation coefficients for the panchromatic, R, G, and B bands, respectively. It is the detail coefficient of the panchromatic band. It is the detail coefficient of any band of R, G, or B. This is the weighting coefficient, set to 0.5; S13. Starting from the Nth layer, reconstruct the structure layer by layer upwards until the resolution is restored. The reconstruction formula is as follows:
[0046] in, LL These are approximate coefficients. , It's the reconstruction filter for Bior 4.4. This represents upsampling.
[0047] S2. Use non-subsampled contourlet transform to enhance the RGB fused image and the panchromatic image; S3. Construct the aquaculture mat extraction rules, segment the panchromatic image in S2, and obtain the initial aquaculture mat mask; S31. By statistically analyzing the pixel values, shape, and area features of the aquaculture baskets on the panchromatic image, the following rules for extracting the aquaculture baskets are constructed:
[0048] in, It is the initial aquaculture membrane. , , These are the pixel values, aspect ratio, and area of the aquaculture basket. A, B, C It refers to the range of characteristic statistics.
[0049] The initial culture mat mask was extracted according to the culture mat extraction rules, wherein, A, B, C The statistical ranges are set as follows: , , .
[0050] S4. Input the enhanced RGB fused image and the initial aquaculture mat mask into the SAM model for annotation-guided cue training. Specifically, in the SAM model, the initial aquaculture mat mask is used to generate aquaculture mat cue masks. Based on the aquaculture mat cue masks, candidate aquaculture mat masks are generated by segmentation. The cosine similarity between the aquaculture mat cue masks and the candidate masks is calculated. High similarity regions are retained as pseudo aquaculture mat masks. The loss between the initial aquaculture mat mask and the pseudo aquaculture mat mask is calculated to optimize the SAM model. The suggested training steps for the SAM model are as follows: S41. Input the initial aquaculture mat mask from S3 into the trained SAM model for annotation-guided cue training, then:
[0051] in, It is the initial SAM model. It's a fake aquaculture mat. It is the initial aquaculture mat mask input. These are the parameters of the cross-attention layer between the cue encoder and mask decoder in the SAM model. L It is the loss function.
[0052] L The mask loss and the IoU prediction loss are weighted and summed as follows:
[0053] in, , These are the mask loss and IoU Predicting losses, This is the weight, set to 0.5.
[0054] S42, It is achieved by calculating the cosine similarity between the candidate mask and the breeding field hint mask. The result, obtained after filtering highly similar regions, is calculated using the following formula:
[0055] in, It's a mask indicating the aquaculture area. It is a candidate mask.
[0056] S5. Use the initial culture mat mask and the pseudo culture mat mask from S4 for evaluation. If the evaluation accuracy does not meet the standard, use the pseudo culture mat mask as the initial culture mat mask and re-execute S4 for iteration. If the evaluation accuracy meets the standard, output the pseudo culture mat mask as the culture mat mask and execute S6. The steps for iterative extraction of aquaculture feed are as follows: S51. Calculate the initial culture bed mask and the pseudo culture bed mask. IoU , Dice The coefficient assesses the degree of overlap between the two, and the calculation formula is as follows:
[0057]
[0058] in, A, B These are the initial culture pan mask and the pseudo culture pan mask.
[0059] S52. Repeat S41, S42, and S51, continuously updating the model and the pseudo-culture mat mask, until the... k In the next iteration, we have:
[0060] S53. Set the iteration termination threshold. , If we take 0.9, then the stopping condition for iteration is:
[0061] S54. Use the pseudo-culture floor mask that stops iterating when it meets the evaluation accuracy as the culture floor mask and execute S6.
[0062] S6. Extract the center points of the mask blocks of the aquaculture rafts, construct the aquaculture raft topology rules, connect the center points to obtain the raft aquaculture area; visually interpret the center points of the aquaculture raft masks connecting a small number of aquaculture areas to obtain the raft aquaculture area verification samples; based on the verification samples, use a genetic algorithm to automatically adjust the topology rules to achieve the best results.
[0063] The center point of each aquaculture trough in the extracted results is calculated, and then aquaculture trough topology rules are constructed to extract the aquaculture area. The rules and parameter adjustment strategies are as follows: S61. Baseline Extension Rule: Given a base point P ,search r Within distance ( n (Points = 1, 2, 3...) connect ;by As a new starting point, with and Direction Angle range and r Search for the next point within the distance range; continue this process, extending the baseline to the point where no points exist within the search range; iterate through all points not included in the baseline, executing the connection rules described above; based on the baseline length, discard points with fewer than [number missing] connection points in a single straight line trend. n The baseline for each.
[0064] S62. Baseline Closure Rule: Given the endpoints of an unclosed baseline E The angle between the lines of the search trend and the line of intersection is greater than 1. And the nearest other unclosed baseline endpoint F Extend the two baselines until they intersect to complete the baseline closure; and so on, traversing all the endpoints of the unclosed baselines to complete the baseline closure.
[0065] S63. Connect a small number of aquaculture plots through visual interpretation, create a validation set for the aquaculture area, and use a genetic algorithm to automatically optimize the parameters of the aquaculture plot topology rules.
[0066] The optimal aquaculture area extraction result was obtained by adjusting the baseline extension and closure rules with the genetic algorithm; among them, , r、n、 The value ranges are respectively set as follows: , , , .
[0067] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for extracting raft aquaculture areas based on high-resolution optical images, characterized in that, Includes the following steps: S1: Select the high-resolution optical images of the target aquaculture area after preprocessing, and use discrete wavelet transform to fuse the R, G, B bands and panchromatic band to obtain a high-resolution RGB fused image; S2: Use non-subsampled contourlet transform to perform image enhancement processing on the RGB fused image and the panchromatic image to obtain an enhanced RGB fused image; S3: Construct the rules for extracting aquaculture baskets and segment the panchromatic image to obtain the initial aquaculture basket mask; S4: Input the enhanced RGB fused image and the initial aquaculture mat mask into the SAM model for annotation-guided cue training. In the SAM model, generate aquaculture mat cue masks using the initial aquaculture mat mask, segment and generate candidate aquaculture mat masks based on the cue masks, calculate the cosine similarity between the cue masks and the candidate masks, retain high similarity regions as pseudo aquaculture mat masks, and calculate the loss between the initial aquaculture mat mask and the pseudo aquaculture mat mask to optimize the SAM model. S5: Evaluate the initial culture mat mask and the pseudo culture mat mask. If the evaluation accuracy does not meet the standard requirements, use the pseudo culture mat mask as the initial culture mat mask and re-execute S4 for iteration. If the evaluation accuracy meets the standard requirements, output the pseudo culture mat mask as the culture mat mask and execute S6. S6: Extract the center points of the mask blocks of the aquaculture raft, construct the aquaculture raft topology rules, connect the center points to obtain the raft aquaculture area, interpret the center points of the aquaculture raft mask connecting a small number of aquaculture areas, obtain the raft aquaculture area verification sample, and use a genetic algorithm to automatically tune the parameters of the topology rules to control the parameters of the topology rules to achieve the optimal effect.
2. The method for extracting raft aquaculture areas based on high-resolution optical images according to claim 1, characterized in that: When fusing R, G, B bands and panchromatic bands using discrete wavelet transform: S11: Using a biorthogonal wavelet basis, perform N-level decomposition on the R, G, and B panchromatic bands respectively to obtain the approximate coefficients of each level. and detail coefficient , , The decomposition formula is as follows: in, k It is the number of decomposition levels. It is a two-dimensional scaling function of Bior 4.
4. , , These are the horizontal, vertical, and diagonal wavelet functions of Bior 4.
4. It is the original image. Represents downsampling, Represents discrete convolution operation. col、 row Represents the convolution direction; S12: Use fusion rules to merge the detail information of the panchromatic band with the spectral information of the R, G, and B bands, and fuse the approximation coefficient and detail coefficient. The fusion formula is as follows: in, These are the approximate coefficients after fusion. It is the detail coefficient after fusion. , , , These are the approximation coefficients for the panchromatic, R, G, and B bands, respectively. It is the detail factor for the panchromatic band. It is the detail coefficient of any band of R, G, or B. These are weighting coefficients; S13: Reconstruct the resolution layer by layer starting from the Nth layer until the resolution is restored. The reconstruction formula is as follows: in, LL These are approximate coefficients. , It's the reconstruction filter for Bior 4.
4. This represents upsampling.
3. The method for extracting raft aquaculture areas based on high-resolution optical images according to claim 1, characterized in that: The following method is used when constructing the rules for extracting aquaculture feed: The following rules for extracting aquaculture baskets were constructed by statistically analyzing the pixel values, shape, and area features of the baskets in the panchromatic image: in, It is the initial aquaculture membrane. , , These are the pixel values, aspect ratio, and area of the aquaculture basket. A, B, C It refers to the range of characteristic statistics; The initial aquaculture mat was obtained by using the statistical characteristic range as the screening parameter.
4. The method for extracting raft aquaculture areas based on high-resolution optical images according to claim 1, characterized in that: When inputting the enhanced RGB fused image and the initial aquaculture mat mask into the SAM model for annotation-guided cue training: S41: Input the initial aquaculture plot mask into the trained SAM model for annotation-guided training: in, It is the initial SAM model. It's a fake aquaculture mat. It is the initial aquaculture mat mask input. These are the parameters of the cross-attention layer between the cue encoder and mask decoder in the SAM model. L It is a loss function; L Using mask loss and IoU The predicted losses are weighted and summed, as shown in the following formula: in, , These are the mask loss and IoU Predicting losses, It is weight; S42: False aquaculture mat mask It is achieved by calculating the cosine similarity between the candidate mask and the breeding field hint mask. The result, obtained after filtering highly similar regions, is calculated using the following formula: in, It's a mask indicating the aquaculture area. It is a candidate mask.
5. The method for extracting raft aquaculture areas based on high-resolution optical images according to claim 4, characterized in that: When re-executing S4 for iteration using the pseudo-culture floor mask as the initial culture floor mask: S51: Calculate the initial culture bed mask and the pseudo culture bed mask. IoU , Dice The coefficient assesses the degree of overlap between the two, and the calculation formula is as follows: in, A, B These are the initial culture tray mask and the pseudo culture tray mask, respectively. S52: Repeat S41, S42, and S51, continuously updating the SAM model and the pseudo-culture trough mask, until the... k In the next iteration, we have: S53: Set the iteration termination threshold to... The stopping condition for iteration is: S54: Use the pseudo-culture pan mask that stops iterating when it meets the evaluation accuracy as the culture pan mask and execute S6.
6. The method for extracting raft aquaculture areas based on high-resolution optical images according to claim 1, characterized in that: The specific topology rules for the aquaculture basket include: Baseline extension rules: Set base points P ,search r Within distance ( n (Points = 1, 2, 3...) connect ,by As a new starting point, with and Direction Angle range and r Search for the next point within the distance range, and so on, extending the baseline to the point where no points exist within the search range. Iterate through all points not included in the baseline, applying the connection rules described above. Based on the baseline length, discard points with fewer than 10 connection points in a single straight line trend. n The baseline for each; Baseline Closure Rule: Given the endpoints of an unclosed baseline E The angle between the lines of the search trend and the line of intersection is greater than 1. And the nearest other unclosed baseline endpoint F Extend the two baselines until they intersect to complete the baseline closure; and so on, traversing all the endpoints of the unclosed baselines to complete the baseline closure.