A remote sensing refined identification method and system for aquaculture areas based on U-net framework

Abstract

The present invention proposes a remote sensing fine identification method and system for breeding areas based on the U-net framework, which obtains breeding area features by effectively extracting features from spatial remote sensing image data (multispectral or hyperspectral) containing target breeding areas, constructs a breeding area remote sensing fine model based on the breeding area features and corresponding target breeding area label data features, inputs the spatial remote sensing image data to be predicted into the breeding area remote sensing fine model, obtains the target breeding area label data of the spatial remote sensing image to be predicted, and realizes intelligent extraction of breeding areas. Through the effective processing of spatial remote sensing image data, the breeding area identification accuracy of the breeding area remote sensing fine model is improved, thereby assisting in intelligent judgment of breeding areas, thereby assisting in intelligent breeding, reducing the time cost of manual inspections, improving the accuracy of breeding area image prediction, and realizing personalized image recognition survey services for specific breeding areas.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent image recognition, and in particular relates to a remote sensing refined recognition method and system for aquaculture areas based on a U-net framework. Background Art

[0002] The breeding industry is an important part of the modern agricultural industry system. Accurately identifying breeding farms and obtaining accurate spatial images of breeding farms are of great significance for dynamic monitoring of breeding, accurate estimation of production, and good management of breeding environment. Breeding farms are the main places for human breeding activities and are a kind of functional spatial environment. The extraction of spatial images is a hot research issue in the field of surveying and remote sensing. Traditional information extraction methods mainly use artificially designed features to extract information. Specifically, by comprehensively considering information such as spectrum, texture, shape, and spatial relationship, the rules for extracting target information are constructed.

[0003] This method of extracting spatial images relies on past experience and may not fully utilize spatial images. In addition, when factors such as lighting, atmospheric conditions, and sensor quality vary, the same environment will change significantly in remote sensing images. Therefore, the above-mentioned method of designing features based on experience can often only process specific data and cannot play a good role in the extraction of complex aquaculture environments.

[0004] To solve the above problems, the existing technology usually extracts artificially designed features and then combines them with computers to extract remote sensing space images. For example, support vector machines, random forest classifiers and decision trees use computer algorithms to overcome the shortcomings of empirical methods by inputting a large number of training samples, while the method of using the U-net model to identify target farms needs further development.

[0005] Specifically for marine farms, there are generally no obvious contrast objects around them, so it is necessary to carry out fine identification of the farms. At the same time, marine farms are usually composed of multiple breeding areas, which are arranged regularly, with small spacing between sheds and clear boundaries. As for the collection of spatial images of marine farms, the existing technology usually focuses on manual field surveys. However, according to the above content, it can be seen that the spatial distribution of marine farms is relatively wide, the images are relatively simple, and the optical conditions are complex. Manual survey methods require high time and labor costs, and it is difficult to obtain accurate and complete information. In addition, the extraction method of such spatial images is mainly based on low-level or mid-level features, relying on artificial design to describe the properties of the farms. The extraction of spatial images of marine farms is often limited by artificial knowledge and experience, and there are problems such as low segmentation accuracy and blurred segmentation boundaries.

[0006] In summary, there are limitations in using existing methods to identify spatial images of marine farms. Specifically, there are currently the following difficulties in identifying marine farms: farms are of different sizes, small farms have unobtrusive appearance features, and are easily confused with the surrounding environment; although the appearance designs are similar, the specific features are different; the appearance features are mainly square shapes, which are easy to confuse; there are serious pixel mixing problems under complex backgrounds, and the extraction accuracy is low. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a remote sensing fine identification method of aquaculture areas based on the U-net framework.

[0008] Containing first spatial remote sensing image data of a target breeding area, and obtaining corresponding label data of the target breeding area;

[0009] Extract effective features from the first spatial remote sensing image data to obtain breeding area features; and construct a breeding area remote sensing refined prediction model based on the breeding area features and the target breeding area label data;

[0010] Receive the second spatial remote sensing image data to be predicted, perform effective feature processing on the second spatial remote sensing image data to generate features of the breeding area to be detected, input the features of the breeding area to be detected into the breeding area remote sensing refined prediction model to generate current target breeding area label data, and help users identify the target breeding area according to the current target breeding area label data.

[0011] Furthermore, the effective feature processing is performed by using a remote sensing segmentation method combined with a remote sensing image updating method to perform feature processing.

[0012] Furthermore, the image refinement formula is:

[0013]

[0014] Where C n+1 is the updated remote sensing data, G is the first spatial remote sensing image data, A is the parameter matrix, W i is the i-dimensional deformation operator, r i is a fixed value parameter greater than 0, is an artificial variable, A T is the transposed matrix of the parameter matrix, is the transposed matrix of the i-dimensional deformation operator, ∈ is the deviation image data, and the first spatial remote sensing image data after refinement is the first spatial remote sensing refined image data.

[0015] Furthermore, the aquaculture area remote sensing refined prediction model includes a classifier based on Fisher criterion, a support vector machine and a U-net neural network model.

[0016] Furthermore, before using the aquaculture area remote sensing refined prediction model, effective feature processing is performed on the first spatial remote sensing refined image data to separate the target aquaculture area portion from other remote sensing image portions, and the steps are as follows:

[0017] S1. Acquire the first spatial remote sensing refined image data;

[0018] S2. Calculate a histogram of the first spatial remote sensing refined image data, where the first spatial remote sensing refined image data is D-dimensional data;

[0019] S3, using an improved search algorithm to perform image segmentation; the M-dimensional data F=(F1,..., F i ) but F D The improved search algorithm is as follows:

[0020]

[0021]

[0022] In the formula, F i Belong to [a i ,b i ],a i and b i is the upper and lower limits of the image grayscale value, i=1, 2, 3, ..., M;

[0023] S4, select a jump rate to perform a reverse learning strategy on the first spatial remote sensing refined image data, generate a change counterpoint, perform greedy selection, and retain the point with the smallest fitness as the optimal threshold point. The fitness calculation method adopts the entropy calculation method. The entropy E(t1, t2, ..., t m ) is calculated as follows:

[0024] E(t1, t2, ..., t m )=H 0+...+ H m

[0025]

[0026]

[0027]

[0028] Where P iis the probability of occurrence of a pixel with gray value i, M*N is the total number of pixels of the first spatial remote sensing refined image data, h(i) is the number of pixels with gray value i, the gray value range of the image is {0,1,...,y-1}, and m thresholds divide the image into m+1 different categories, H m represents the entropy of the mth image after classification, and the image threshold vector is [t1, ..., t m ];

[0029] S5. Utilize step S3 to find the optimal image threshold vector to segment the first spatial remote sensing refined image data, and output the segmented first spatial remote sensing refined segmented image data.

[0030] Furthermore, the first spatial remote sensing refined segmentation image data and target breeding area label data are used to train a U-net neural network model.

[0031] Furthermore, the first spatial remote sensing refined segmentation image data is cut into a plurality of image blocks of 256×256 size, and then the training set and the test set are divided into a ratio of 4:1 to train the U-net neural network model.

[0032] A remote sensing refined identification system for aquaculture areas based on the U-net framework is also provided, including a data acquisition terminal and a remote sensing refined prediction terminal for aquaculture areas, which is characterized by:

[0033] The data acquisition terminal includes a breeding remote sensing image acquisition module and a data collection module, and the breeding area remote sensing refined prediction terminal includes a breeding area remote sensing refined module and a breeding area survey module, wherein:

[0034] Aquaculture remote sensing image acquisition module: obtains first spatial remote sensing image data containing the target aquaculture area, obtains corresponding target aquaculture area label data, and obtains second spatial remote sensing image data to be predicted;

[0035] Data collection module: extracting effective features from the first spatial remote sensing image data to obtain breeding area features; receiving the second spatial remote sensing image data to be predicted, and performing effective feature processing on the second spatial remote sensing image data to be predicted to obtain breeding area features to be detected;

[0036] Aquaculture area remote sensing refinement module: receiving the aquaculture area characteristics and the target aquaculture area label data to jointly construct an aquaculture area remote sensing refinement prediction model, receiving the aquaculture area characteristics input to be detected to generate the current target aquaculture area label data;

[0037] Breeding area survey module: helps users identify target breeding areas based on the target breeding area label data.

[0038] A control device for fine-grained remote sensing identification of aquaculture areas based on a U-net framework is also provided, comprising a memory and a processor, characterized in that a control program for fine-grained remote sensing identification of aquaculture areas based on a U-net framework is stored in the memory, and the processor is configured to run the control program for intelligent image processing of fine-grained remote sensing identification of aquaculture areas to run the fine-grained remote sensing identification method of aquaculture areas based on a U-net framework.

[0039] A computer-readable storage medium is also provided, characterized in that it includes: computer executable instructions, which, when executed, execute the remote sensing refined identification method of aquaculture areas based on the U-net framework.

[0040] The present invention utilizes image quality optimization and refinement methods combined with background segmentation methods to process image features, making remote sensing images more accurate and target feature vectors clearer. The remote sensing image data after such feature processing is closer to the actual situation of the farm, and makes the remote sensing refinement prediction model of the breeding area more accurate. Secondly, when training the model, the image is cut into multiple image blocks, so that the model's prediction is more in line with the characteristic performance of marine remote sensing images, thereby providing users with more accurate image recognition and research services.

[0041] More embodiments and improved effects of the present invention will be further introduced in conjunction with the drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a method for remote sensing fine identification of aquaculture areas based on a U-net framework of the present invention;

[0043] Figure 2 It is a schematic diagram of a remote sensing refined identification system for aquaculture areas based on a U-net framework of the present invention;

[0044] Figure 3 is a schematic diagram of a target marine aquaculture farm identified by the present invention;

[0045] Figure 4 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0047] In order to solve the above technical problems, the present invention proposes a remote sensing fine identification method of aquaculture areas based on the U-net framework.

[0048] Containing first spatial remote sensing image data of a target breeding area, and obtaining corresponding label data of the target breeding area;

[0049] Extract effective features from the first spatial remote sensing image data to obtain breeding area features; and construct a breeding area remote sensing refined prediction model based on the breeding area features and the target breeding area label data;

[0050] Receive the second spatial remote sensing image data to be predicted, perform effective feature processing on the second spatial remote sensing image data to generate features of the breeding area to be detected, input the features of the breeding area to be detected into the breeding area remote sensing refined prediction model to generate current target breeding area label data, and help users identify the target breeding area according to the current target breeding area label data.

[0051] In this embodiment, the label data of the target breeding area is set as the label vector L is 1, and the label vector of the non-target breeding area is 0.

[0052] Furthermore, the effective feature processing is performed by using a remote sensing segmentation method combined with a remote sensing image updating method to perform feature processing.

[0053] Furthermore, the image refinement formula is:

[0054]

[0055] Where C n+1 is the updated remote sensing data, G is the first spatial remote sensing image data, A is the parameter matrix, W i is the i-dimensional deformation operator, r i is a fixed value parameter greater than 0, is an artificial variable, A T is the transposed matrix of the parameter matrix, is the transposed matrix of the i-dimensional deformation operator, ∈ is the deviation image data, and the first spatial remote sensing image data after refinement is the first spatial remote sensing refined image data.

[0056] The parameter matrix in this embodiment is a unit matrix, the deformation operator can be set according to the needs of those skilled in the art, and the deviation image data is a deviation average gray value image matrix of the erroneous image data collected in the first spatial remote sensing image data.

[0057] Furthermore, the aquaculture area remote sensing refined prediction model includes a classifier based on Fisher criterion, a support vector machine and a U-net neural network model.

[0058] In this embodiment, since only target aquaculture areas and non-target aquaculture areas need to be classified in the marine aquaculture areas, the above classifier models are calculated in the form of a binary classification algorithm. The classifier based on the Fisher criterion can be:

[0059]

[0060] Where A is the first spatial remote sensing refined image data, g(S) is the target breeding area label data, and W T is the normal vector perpendicular to the hyperplane. Since marine aquaculture farms will experience the impact of waves, the image features will be affected. The classifier is improved for waves. W(C) is the period value of the periodic function of the waves, and S(C) is the average water flow rate under calm sea level conditions.

[0061] This embodiment considers the influence of sea waves to modify the classifier function to obtain a more accurate remote sensing image data classification model. The periodic function of sea waves is specifically:

[0062]

[0063] Its period value is W(C), where h is the wind speed at a height of 7.5 meters above the sea surface, and the constant R is 3.05m / s 2 , g is the acceleration of gravity, C is the frequency of the wave, and the value of S(C) is generally 0.5m / s-2m / s.

[0064] Furthermore, before using the aquaculture area remote sensing refined prediction model, effective feature processing is performed on the first spatial remote sensing refined image data to separate the target aquaculture area portion from other remote sensing image portions, and the steps are as follows:

[0065] S1. Acquire the first spatial remote sensing refined image data;

[0066] S2. Calculate a histogram of the first spatial remote sensing refined image data, where the first spatial remote sensing refined image data is D-dimensional data;

[0067] S3, using an improved search algorithm to perform image segmentation; the M-dimensional data F=(F1,..., F i ) but F D The improved search algorithm is as follows:

[0068]

[0069]

[0070] In the formula, F i Belong to [a i ,b i ],a i and b i is the upper and lower limits of the image grayscale value, i=1, 2, 3, ..., M;

[0071] S4, select jump rate Jv=0.65 to perform reverse learning strategy on the first spatial remote sensing refined image data, generate change counterpoints, perform greedy selection, and retain the one with the smallest fitness as the optimal threshold point. The fitness calculation method adopts entropy calculation method. The entropy E(t1, t2, ..., t m ) is calculated as follows:

[0072] E(t1, t2, …, t m )=H 0+...+ H m

[0073]

[0074]

[0075]

[0076] Where P i is the probability of occurrence of a pixel with gray value i, M*N is the total number of pixels of the first spatial remote sensing refined image data, h(i) is the number of pixels with gray value i, the gray value range of the image is {0,1,...,y-1}, and m thresholds divide the image into m+1 different categories, H m represents the entropy of the mth image after classification, and the image threshold vector is [t1, ..., t m ];

[0077] S5. Utilize step S3 to find the optimal image threshold vector to segment the first spatial remote sensing refined image data, and output the segmented first spatial remote sensing refined segmented image data.

[0078] Furthermore, the first spatial remote sensing refined segmentation image data and target breeding area label data are used to train a U-net neural network model.

[0079] Furthermore, the first spatial remote sensing refined segmentation image data is cut into multiple image blocks of 256×256 size. In this embodiment, it is preferably cut into 30 256X256 pixel image blocks, and then the training set and the test set are divided into a ratio of 4:1 to train the U-net neural network model.

[0080] A remote sensing refined identification system for aquaculture areas based on the U-net framework is also provided, including a data acquisition terminal and a remote sensing refined prediction terminal for aquaculture areas, which is characterized by:

[0081] The data acquisition terminal includes a breeding remote sensing image acquisition module and a data collection module, and the breeding area remote sensing refined prediction terminal includes a breeding area remote sensing refined module and a breeding area survey module, wherein:

[0082] Aquaculture remote sensing image acquisition module: obtains first spatial remote sensing image data containing the target aquaculture area, obtains corresponding target aquaculture area label data, and obtains second spatial remote sensing image data to be predicted;

[0083] Data collection module: extracting effective features from the first spatial remote sensing image data to obtain breeding area features; receiving the second spatial remote sensing image data to be predicted, and performing effective feature processing on the second spatial remote sensing image data to be predicted to obtain breeding area features to be detected;

[0084] Aquaculture area remote sensing refinement module: receiving the aquaculture area characteristics and the target aquaculture area label data to jointly construct an aquaculture area remote sensing refinement prediction model, receiving the aquaculture area characteristics input to be detected to generate the current target aquaculture area label data;

[0085] Breeding area survey module: helps users identify target breeding areas based on the target breeding area label data.

[0086] A control device for fine-grained remote sensing identification of aquaculture areas based on a U-net framework is also provided, comprising a memory and a processor, characterized in that a control program for fine-grained remote sensing identification of aquaculture areas based on a U-net framework is stored in the memory, and the processor is configured to run the control program for intelligent image processing of fine-grained remote sensing identification of aquaculture areas to run the fine-grained remote sensing identification method of aquaculture areas based on a U-net framework.

[0087] A computer-readable storage medium is also provided, characterized in that it includes: computer executable instructions, which, when executed, execute the remote sensing refined identification method of aquaculture areas based on the U-net framework.

[0088] The present invention utilizes image quality optimization and refinement methods combined with background segmentation methods to process image features, making remote sensing images more accurate and target feature vectors clearer. The remote sensing image data after such feature processing is closer to the actual situation of the farm, and makes the remote sensing refinement prediction model of the breeding area more accurate. Secondly, when training the model, the image is cut into multiple image blocks, so that the model's prediction is more in line with the characteristic performance of marine remote sensing images, thereby providing users with more accurate image recognition and research services.

[0089] Of course, it can be understood that each embodiment of the present invention can achieve one of the effects alone, and a combination of multiple embodiments of the present invention can achieve all of the above effects, but it is not required that every embodiment of the present invention achieve all of the above advantages and effects, because each embodiment of the present invention can constitute a separate technical solution and make one or more contributions to the prior art.

[0090] The present invention does not particularly specify the structure of some modules, which shall be subject to the contents recorded in the prior art. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters. The scope of protection of the present invention shall be subject to the contents actually recorded in the claims.

Claims

1. A remote sensing refined identification method for aquaculture areas based on the U-net framework, characterized in that: The method comprises: Acquire first spatial remote sensing image data including a target breeding area, and acquire corresponding target breeding area label data; Extract effective features from the first spatial remote sensing image data to obtain breeding area features; and construct a breeding area remote sensing refined prediction model based on the breeding area features and the target breeding area label data; Receiving second spatial remote sensing image data to be predicted, performing effective feature processing on the second spatial remote sensing image data to generate features of a breeding area to be detected, inputting the features of the breeding area to be detected into the breeding area remote sensing refined prediction model to generate label data of the current target breeding area, and helping a user to identify the target breeding area according to the label data of the current target breeding area; The effective feature processing is performed by using a remote sensing splitting method combined with a remote sensing image updating method to perform feature processing; The formula for image refinement is: Where C n+1 is the updated remote sensing data, G is the first spatial remote sensing image data, A is the parameter matrix, W i is the i-dimensional deformation operator, r i is a fixed value parameter greater than 0, is an artificial variable, A T is the transposed matrix of the parameter matrix, is the transposed matrix of the i-dimensional deformation operator, ∈ is the deviation image data, and the first spatial remote sensing image data after refinement is the first spatial remote sensing refined image data; Before using the aquaculture area remote sensing refined prediction model, the first spatial remote sensing refined image data is effectively processed with features to separate the target aquaculture area from other remote sensing image parts, and the steps are as follows: S1. Acquire the first spatial remote sensing refined image data; S2. Calculate a histogram of the first spatial remote sensing refined image data, where the first spatial remote sensing refined image data is M-dimensional data; S3, using an improved search algorithm to perform image segmentation; let the M-dimensional data F = (F1, ..., F M ) but F D The improved search algorithm is as follows: In the formula, F i Belong to [a i ,b i ],a i and b i is the upper and lower limits of the image grayscale value, i=1, 2, 3, ..., M; S4, select a jump rate to perform a reverse learning strategy on the first spatial remote sensing refined image data, generate a change counterpoint, perform greedy selection, and retain the point with the smallest fitness as the optimal threshold point. The fitness calculation method adopts the entropy calculation method. The entropy E(t1, t2, ..., t m ) is calculated as follows: E(t1,t2,...,t m )=H 0+...+ H m Where P i is the probability of occurrence of a pixel with gray value i, M*N is the total number of pixels of the first spatial remote sensing refined image data, h(i) is the number of pixels with gray value i, the gray value range of the image is {0,1,...,y-1}, and m thresholds divide the image into m+1 different categories, H m represents the entropy of the mth image after classification, and the image threshold vector is [t1, ..., t m ]; S5, using step S3 to find the optimal image threshold vector to split the first spatial remote sensing refined image data, and output the split first spatial remote sensing refined split image data; The aquaculture area remote sensing refined prediction model includes a classifier based on Fisher criterion, a support vector machine and a U-net neural network model; The classifier based on Fisher's criterion is: Where L is the first spatial remote sensing refined image data, g(S) is the target breeding area label data, and W T is the normal vector perpendicular to the hyperplane, the classifier is improved by waves, |W(C)| is the period value of the periodic function of the sea wave, |S(C)| is the average water flow rate under calm sea level conditions; The periodic function of the waves is: Its period value is |W(C)|, where h is the wind speed at a height of 7.5 meters above the sea surface, and the constant R is 3.05m / s 2 , g is the acceleration of gravity, C is the frequency of the ocean wave, and the value of |S(C)| is 0.5m / s-2m / s.

2. The method for remote sensing fine identification of aquaculture areas based on the U-net framework as claimed in claim 1, characterized in that: The first spatial remote sensing refined segmentation image data and the target breeding area label data are used to train the U-net neural network model.

3. A method for remote sensing fine identification of aquaculture areas based on a U-net framework as claimed in claim 2, characterized in that: The first spatial remote sensing refined segmentation image data is cut into multiple image blocks of 256×256 size, and then divided into a training set and a test set in a ratio of 4:1 to train the U-net neural network model.

4. A remote sensing fine identification system for aquaculture areas based on a U-net framework, used to implement a remote sensing fine identification method for aquaculture areas based on a U-net framework as claimed in claim 1, comprising a data acquisition terminal and a remote sensing fine prediction terminal for aquaculture areas, characterized in that: The data acquisition terminal includes a breeding remote sensing image acquisition module and a data collection module, and the breeding area remote sensing refined prediction terminal includes a breeding area remote sensing refined module and a breeding area survey module, wherein: Aquaculture remote sensing image acquisition module: obtains first spatial remote sensing image data containing the target aquaculture area, obtains corresponding target aquaculture area label data, and obtains second spatial remote sensing image data to be predicted; Data collection module: extracting effective features from the first spatial remote sensing image data to obtain breeding area features; receiving the second spatial remote sensing image data to be predicted, and performing effective feature processing on the second spatial remote sensing image data to be predicted to obtain breeding area features to be detected; Aquaculture area remote sensing refinement module: receiving the aquaculture area characteristics and the target aquaculture area label data to jointly construct an aquaculture area remote sensing refinement prediction model, receiving the aquaculture area characteristics input to be detected to generate the current target aquaculture area label data; Breeding area survey module: helps users identify target breeding areas based on the target breeding area label data.

5. A control device for remote sensing fine identification of aquaculture areas based on a U-net framework, comprising a memory and a processor, characterized in that: The memory stores a control program for remote sensing refined identification of aquaculture areas based on a U-net framework, and the processor is configured to run the control program for intelligent image processing of remote sensing refined identification of aquaculture areas to run the method described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: The method comprises: computer executable instructions, which, when executed, execute the method according to any one of claims 1 to 3.

Images