Coral reef health state assessment method and system based on deep learning
Through deep learning-based image recognition and convolutional neural network technology, the health status of coral reefs is automatically identified and quantified, and the problems of strong subjectivity, inefficiency and poor real-time performance in traditional methods are solved, and a fast and accurate coral reef health assessment is achieved.
Patent Information
- Application Number
- CN202510475190.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2025-07-29
AI Technical Summary
The traditional coral reef health status assessment methods have problems such as strong subjectivity, inefficiency, lack of real-timeness and difficulty in quantification, and it is difficult to accurately and quickly evaluate the health status of large areas of coral reefs.
Using a deep learning-based method, the image target recognition model and coral reef health status model are used, and the concentration of coral pathogenic bacteria is combined as an evaluation indicator to automatically identify and quantify the health status of coral reefs, and image feature extraction and classification are used using convolutional neural networks.
It improves the accuracy and efficiency of coral reef health status assessment, and can quickly and in real time monitor the health status of large areas of coral reefs, providing a scientific basis for protection and restoration.
Smart Images
Figure CN120388228A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of coral reef health status assessment, and particularly to a method and system for assessing the health status of coral reefs based on deep learning. Background Art
[0002] Coral reefs are extremely important components of marine ecosystems, with extremely high biodiversity and ecological value. However, in recent years, due to factors such as marine pollution, climate change, and overfishing, coral reefs worldwide are facing serious degradation problems. Accurately assessing the health status of coral reefs is crucial for formulating effective protection measures and restoration strategies.
[0003] Traditional methods for assessing the health status of coral reefs mainly rely on means such as manual observation and water sample detection. These methods have the following limitations:
[0004] Strong subjectivity: Manual observation is easily affected by factors such as the experience and perspective of observers, resulting in inaccurate and non-objective assessment results.
[0005] Low efficiency: Detecting large areas of coral reef regions one by one is time-consuming and laborious, and it is difficult to quickly obtain comprehensive health status information.
[0006] Lack of real-time nature: Methods such as water sample detection take a certain amount of time to complete and are difficult to reflect the health changes of coral reefs in real time.
[0007] Difficulty in quantification: Traditional methods are difficult to accurately quantify the health status of coral reefs and cannot provide a scientific basis for protection and restoration work.
[0008] With the development of computer technology, especially the wide application of deep learning technology in the field of image recognition and classification, it provides a new idea for solving the above problems. Deep learning models can automatically learn features from a large amount of data and have powerful image analysis capabilities, which are expected to improve the efficiency and accuracy of coral reef health status assessment. Summary of the Invention
[0009] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a method for assessing the health status of coral reefs based on deep learning, which can improve the efficiency and accuracy of coral reef health status assessment.
[0010] To achieve the above purpose, the present invention provides the following solutions:
[0011] A method for assessing the health status of coral reefs based on deep learning, comprising:
[0012] Obtain a coral reef image, input the coral reef image into an image target recognition model, and obtain a target recognition result; the image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels;
[0013] According to the target recognition result, perform a cropping process on the coral reef image, input the processed image into a coral reef health status model, and obtain a health assessment result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates the classification label corresponding to the original coral reef image by extracting the concentration of coral pathogenic bacteria in the water sample in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
[0014] Optionally, the image target recognition model includes:
[0015] Multiple multi-branch perception units arranged in parallel and capable of adding target perception units, the multi-branch perception units are sequentially connected to an image similarity calculation unit and an output unit;
[0016] A target perception unit, configured to perceive the target in the coral reef image in a positive feedback loop manner, obtain an output image and the number of iterations of the positive feedback loop corresponding to each output image;
[0017] The image similarity calculation unit is configured to calculate the image similarity between the coral reef image and each output image, and obtain multiple image similarities;
[0018] The output unit is configured to determine whether the image similarity corresponding to the smallest number of iterations among all the number of iterations is the largest image similarity among all the image similarities. If it is determined that the image similarity corresponding to the smallest number of iterations is the largest image similarity, then use the category of the output image corresponding to the smallest number of iterations as the target recognition result.
[0019] Optionally, after performing the cropping process on the coral reef image, it includes: performing denoising and normalization operations on the cropped coral reef image to improve the image quality.
[0020] Optionally, establishing the coral reef health status model includes:
[0021] Use the Sequential model of the Keras library to construct a convolutional neural network model. The convolutional neural network model is composed of a convolutional layer, a max pooling layer, a Flatten layer, and a fully connected layer. Each max pooling layer is sequentially located after each convolutional layer. The Flatten layer flattens the output of the last pooling layer into a one-dimensional vector and connects it to the fully connected layer;
[0022] The convolutional layer is used to extract image features represented by each pixel in the image in a combined or independent manner. The extracted image features include texture features and color features.
[0023] The max pooling layer is used to reduce the spatial size of the feature map to speed up training, reduce the risk of overfitting, and add an aggregation operation to merge the feature values in a local area into a representative feature value, suppressing noise and redundant information while retaining important feature information.
[0024] The fully connected layer is used to convert the original output into a probability distribution based on the Softmax function and take the highest probability distribution as the health assessment result.
[0025] Optionally, the expression of the Softmax function is:
[0026]
[0027] where \(z_i\) i is the \(i\)-th element in the input vector, \(K\) is the total number of classes. Through the Softmax function, the output value of each class is converted into a probability value between 0 and 1, and the sum of the probabilities of all classes is 1.
[0028] Optionally, training the convolutional neural network model using the second training set includes:
[0029] Training the convolutional neural network model using the second training set and using categorical cross-entropy to define the loss function to measure the difference between the probability distribution predicted by the model and the true label:
[0030] BCE(p,y)=-[ylog(p)+(1 - y)log(1 - p)]
[0031] where \(y\) is the true label and \(p\) is the positive class probability predicted by the model.
[0032] Optionally, the classification labels include: healthy, sub-healthy, and damaged.
[0033] To achieve the above object, the present invention also provides a coral reef health status assessment system based on deep learning, including:
[0034] A target recognition module, configured to obtain a coral reef image, input the coral reef image into an image target recognition model, and obtain a target recognition result; the image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels.
[0035] An image processing module for cropping the coral reef image according to the target recognition result;
[0036] A health status assessment module for inputting the processed image into a coral reef health status model to obtain a health assessment result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates the classification labels corresponding to the original coral reef images by extracting the concentration of coral pathogenic bacteria in the water samples in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
[0037] Optionally, the image processing module includes:
[0038] An image processing sub-module for cropping the coral reef image according to the target recognition result, performing denoising and normalization operations on the cropped coral reef image to improve the image quality.
[0039] The beneficial effects of the present invention are:
[0040] By analyzing the coral reef image through a deep learning model, the present invention can automatically identify and extract key features in the image, reduce manual intervention, and reduce the influence of subjective factors on the evaluation result, thereby improving the accuracy and reliability of the evaluation.
[0041] The present invention can quickly process a large number of coral reef images, automatically complete the entire process from image acquisition to health status assessment, greatly improve the evaluation efficiency, and can monitor the health status of large areas of coral reef areas in a short time.
[0042] By using the concentration of coral pathogenic bacteria as an evaluation index and combining the classification function of the deep learning model, the present invention can quantify the health status of coral reefs into specific categories such as healthy, sub-healthy, and damaged, providing a more scientific and specific basis for the protection and restoration of coral reefs.
[0043] The present invention can flexibly adapt to various evaluation scenarios by adjusting the training set and model parameters according to different coral reef areas and environmental conditions, and has strong versatility and scalability.
[0044] The present invention can combine modern image acquisition devices to realize real-time monitoring of the health status of coral reefs, timely discover coral reef health problems, provide early warnings for taking protection measures, and contribute to better protecting the coral reef ecosystem. Description of the Drawings
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0046] Figure 1 It is a flowchart of a method for evaluating the health status of coral reefs based on deep learning according to an embodiment of the present invention. Detailed implementation manners
[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.
[0048] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific implementation manners.
[0049] As Figure 1 shown, this embodiment discloses a method for evaluating the health status of coral reefs based on deep learning, including: obtaining a coral reef image, inputting the coral reef image into an image target recognition model to obtain a target recognition result; the image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels; according to the target recognition result, performing cropping processing on the coral reef image, and inputting the processed image into a coral reef health status model to obtain a health evaluation result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates classification labels corresponding to the original coral reef images by extracting the concentration of coral pathogenic bacteria in the water samples in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
[0050] Further, the image target recognition model includes: multiple multi-branch perception units arranged in parallel and capable of adding target perception units, the multi-branch perception units are sequentially connected with an image similarity calculation unit and an output unit; a target perception unit, configured to perceive the targets in the coral reef image in a positive feedback loop manner, obtain an output image and the number of iterations of the positive feedback loop corresponding to each output image; an image similarity calculation unit, configured to calculate the image similarity between the coral reef image and each output image, and obtain multiple image similarities; an output unit, configured to determine whether the image similarity corresponding to the smallest number of iterations among all the numbers of iterations is the largest image similarity among all the image similarities, and if it is determined that the image similarity corresponding to the smallest number of iterations is the largest image similarity, then use the category of the output image corresponding to the smallest number of iterations as the target recognition result.
[0051] Specifically, the coral reef image is input into a pre-trained image target recognition model capable of continuous learning. The image target recognition model recognizes the targets in the coral reef image and outputs the image target recognition result. Among them, the image target recognition model includes multiple multi-branch perception modules arranged in parallel and capable of adding target perception modules, a first image similarity calculation module connected to the multiple multi-branch perception modules arranged in parallel, and an output module connected to the first image similarity calculation module. Each of the multiple multi-branch perception modules includes multiple target perception modules of the same type, and the multiple target perception modules corresponding to each multi-branch perception module perceive the targets in the coral reef image through a positive feedback loop. Moreover, the newly added target perception modules of the same type in each multi-branch perception module are obtained after being trained with the training samples misjudged by all other target perception modules in the corresponding multi-branch perception module. Each multi-branch perception module is used to perceive the coral reef image to obtain the output image of each multi-branch perception module and the iteration times of the positive feedback loop corresponding to each output image. The first image similarity calculation module is used to calculate the image similarity between the coral reef image and each output image to obtain multiple image similarities. The output module is used to perform target recognition based on the iteration times and the multiple image similarities to obtain the image target recognition result. The output module is further used to determine whether the image similarity corresponding to the smallest iteration time among all iteration times is the largest image similarity among all image similarities. If it is determined that the image similarity corresponding to the smallest iteration time is the largest image similarity, the category of the output image corresponding to the smallest iteration time is used as the image target recognition result. Each multi-branch perception module is provided with a first competition index for representing the winning ability of the corresponding multi-branch perception module when competing with all multi-branch perception modules. The output module is further used to determine whether the image similarity corresponding to the smallest iteration time is not the largest image similarity. If so, it is determined whether the first competition index corresponding to the smallest iteration time is the largest first competition index among all first competition indexes. If the first competition index corresponding to the smallest iteration time is the largest first competition index, the category of the output image corresponding to the smallest iteration time is used as the image target recognition result. The determination process of the first competition index includes: in the current training process of the image target recognition model, obtaining the first training image similarity corresponding to each multi-branch perception module in the current training process; sorting all the first training image similarities in descending order to determine the largest first training image similarity; adding 1 to the value of the first competition index corresponding to the largest first training image similarity.
[0052] Each target perception module of each multi-branch perception module is provided with a second competition index for representing the winning ability of the corresponding target perception module when competing with all target perception modules of the corresponding multi-branch perception module; the output module is further configured to, if the first competition index corresponding to the minimum number of iterations is not the largest first competition index, determine the largest second competition index among all the second competition indexes of the multi-branch perception module corresponding to the minimum number of iterations, and determine whether the largest second competition index is unique. If it is determined that the largest second competition index is unique, the category of the output image of the target perception module corresponding to the largest second competition index is used as the image target recognition result.
[0053] The determination process of the second competition index includes: in the current training process of the image target recognition model, obtaining the second training image similarity between the output images of all target perception modules of the current multi-branch perception module in the current iteration process and the input image of the image target recognition model; sorting all the second training image similarities corresponding to the current multi-branch perception module in descending order to determine the largest second training image similarity corresponding to the current multi-branch perception module; adding 1 to the value of the second competition index of the target perception module corresponding to the largest second training image similarity in the current multi-branch perception module.
[0054] Each target perception module of each multi-branch perception module is further provided with a perception index for representing its target perception ability; the output module is further configured to, if it is determined that the largest second competition index among all the second competition indexes of the multi-branch perception module corresponding to the minimum number of iterations is not unique, determine the largest perception index among all the perception indexes of the multi-branch perception module corresponding to the minimum number of iterations, and determine whether the largest perception index is unique. If it is determined that the largest perception index is unique, the category of the output image of the target perception module corresponding to the largest perception index is used as the image target recognition result.
[0055] The training samples of the image target recognition model include current positive samples; the determination process of the perception index includes: during the current training of the image target recognition model using the current positive samples, obtaining the third training image similarity between the output images of all target perception modules in the current multi-branch perception module during this iteration and the input image of the image target recognition model; adding 1 to the value of the perception index corresponding to the third training image similarity greater than or equal to the preset first image similarity threshold among all the third training image similarities, subtracting 1 from the value of the perception index corresponding to the third training image similarity less than the first image similarity threshold and whose corresponding output image is not an all-zero pixel map, and keeping the value of the perception index corresponding to the third training image similarity whose corresponding output image is an all-zero pixel map unchanged. The training samples of the image target recognition model include current negative samples; the determination process of the perception index includes: during the current training of the image target recognition model using the current negative samples, obtaining the output images of all target perception modules in the current multi-branch perception module during this iteration, and determining whether each output image is an all-zero pixel map; adding 1 to the value of the perception index corresponding to all the output images that are all-zero pixel maps, and subtracting 1 from the value of the perception index corresponding to all the output images that are not all-zero pixel maps.
[0056] Further, after cropping the coral reef image, it includes: denoising and normalizing the cropped coral reef image to improve the image quality.
[0057] Further, establishing a coral reef health status model includes: using the Sequential model of the Keras library to construct a convolutional neural network model, which is composed of a convolutional layer, a max pooling layer, a Flatten layer, and a fully connected layer. Each max pooling layer is located after each convolutional layer in sequence. The Flatten layer flattens the output of the last pooling layer into a one-dimensional vector and connects it to the fully connected layer; the convolutional layer is used to extract image features reflected by each pixel in the image in a combined or independent manner, and the extracted image features include texture features and color features; the max pooling layer is used to reduce the spatial size of the feature map to accelerate the training speed, reduce the risk of overfitting, and add an aggregation operation to merge the feature values in the local area into a representative feature value, suppressing noise and redundant information while retaining important feature information; the fully connected layer is used to convert the original output into a probability distribution based on the Softmax function and take the highest probability distribution as the health assessment result.
[0058] Specifically, a convolutional neural network model is constructed. A convolutional neural network model is built using the Sequential model of the Keras library. The convolutional neural network model consists of 5 convolutional layers Conv2D, 5 max pooling layers MaxPooling2D, 1 Flatten layer, and 2 fully connected layers. Each max pooling layer is located after each convolutional layer in sequence. The Flatten layer flattens the output of the last pooling layer into a one-dimensional vector and then connects it to the fully connected layers. The constructed convolutional neural network model is used to evaluate the health status of coral reefs.
[0059] The convolutional layer is used to extract image features reflected by each pixel in the image either by combination or independently. The extracted image features include texture features and color features. The number of filters in the 5 convolutional layers Conv2D gradually decreases from 32 to 8, so that the abstract features of the image can be gradually captured during the reduction process. The kernel size of each convolutional layer is (3, 3), that is, the size of the filter used for feature extraction in the convolution operation is (3, 3). Each kernel is a 3×3 matrix, and this 3×3 matrix slides over the entire input image to extract different features of the image through the convolution operation. The stride of the convolutional layer is (1, 1). A stride of (1, 1) means that the kernel moves one pixel in both the horizontal and vertical directions each time, and the kernel will calculate at each position of the input image in sequence. The pooling layer reduces the spatial dimension of the feature map to speed up the training speed, reduce the risk of overfitting, and adds an aggregation operation to merge the feature values in the local area into a representative feature value, suppressing noise and redundant information while retaining important feature information to extract more robust and representative features. The first and second fully connected layers use the ReLU function and the Softmax function respectively. The ReLU function is used to introduce non-linearity and allow the CNN network model to learn more complex features. The Softmax function is used to convert the original output of the CNN network model into a probability distribution. During the training process, the CNN network model uses the binary cross-entropy loss function to compare the difference between its output and the true label to optimize the CNN network model.
[0060] The first fully connected layer in the 2 fully connected layers takes the one-dimensional vector output by the Flatten layer as input, then performs a non-linear transformation using 128 neurons, and finally outputs a 128-dimensional feature vector. The second fully connected layer is the output layer of the convolutional neural network. The second fully connected layer maps and transforms the 128-dimensional feature vector output by the first fully connected layer into the final classification probability. The classifier uses the Softmax function. The Softmax function is a 5-class classifier. The classification labels corresponding to the classifier are: healthy, sub-healthy, and damaged, which are represented by 1, 2, and 3 respectively. The calculation formula of the classifier is as follows:
[0061]
[0062] In the formula, j represents the category, T is the total number of categories, T = 5; α j and α k respectively represent the values of the feature vectors, and P j represents the probability that the image belongs to the j-th category;
[0063] A 128-dimensional image feature vector as input, after passing through the Softmax function, outputs a 5×1 vector, and the category corresponding to the number with the largest value in the 5×1 vector is taken as the predicted label of this input data.
[0064] Furthermore, the expression of the Softmax function is:
[0065]
[0066] where z i is the i-th element in the input vector, K is the total number of categories. Through the Softmax function, the output value of each category is converted into a probability value between 0 and 1, and the sum of the probabilities of all categories is 1.
[0067] Furthermore, training the convolutional neural network model using the second training set includes:
[0068] Training the convolutional neural network model using the second training set, and using the categorical cross-entropy to define the loss function to measure the difference between the probability distribution predicted by the model and the true label:
[0069] BCE(p, y) = -[ylog(p) + (1 - y)log(1 - p)]
[0070] where y is the true label and p is the positive class probability predicted by the model.
[0071] Furthermore, the classification labels include: healthy, sub-healthy, and damaged.
[0072] This embodiment also provides a coral reef health status assessment system based on deep learning, including: a target recognition module for obtaining a coral reef image, inputting the coral reef image into an image target recognition model, and obtaining a target recognition result; the image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels; an image processing module for cropping the coral reef image according to the target recognition result; a health status assessment module for inputting the processed image into a coral reef health status model and obtaining a health assessment result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates classification labels corresponding to the original coral reef images by extracting the concentration of coral pathogenic bacteria in the water samples in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
[0073] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for evaluating the health status of coral reefs based on deep learning, characterized in that, Including: Obtain a coral reef image, input the coral reef image into an image target recognition model, and obtain a target recognition result; The image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels; According to the target recognition result, perform a cropping process on the coral reef image, input the processed image into a coral reef health status model, and obtain a health assessment result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates the classification label corresponding to the original coral reef image by extracting the concentration of coral pathogenic bacteria in the water sample in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
2. The method for evaluating the health status of coral reefs based on deep learning according to claim 1, characterized in that The image target recognition model includes: Multiple multi-branch perception units arranged in parallel and capable of adding target perception units, and the multi-branch perception units are sequentially connected to an image similarity calculation unit and an output unit; A target perception unit, configured to perceive the target in the coral reef image in a positive feedback loop manner, obtain an output image and the number of iterations of the positive feedback loop corresponding to each output image; The image similarity calculation unit is configured to calculate the image similarity between the coral reef image and each output image to obtain multiple image similarities; The output unit is configured to determine whether the image similarity corresponding to the smallest number of iterations among all the number of iterations is the largest among all the image similarities. If it is determined that the image similarity corresponding to the smallest number of iterations is the largest image similarity, then use the category of the output image corresponding to the smallest number of iterations as the target recognition result.
3. The method for evaluating the health status of coral reefs based on deep learning according to claim 1, wherein After performing the cropping process on the coral reef image, it includes: performing denoising and normalization operations on the cropped coral reef image to improve the image quality.
4. The method for evaluating the health status of coral reefs based on deep learning according to claim 1, characterized in that Building the coral reef health status model includes: Using the Sequential model of the Keras library to construct a convolutional neural network model, and the convolutional neural network model is composed of a convolutional layer, a max pooling layer, a Flatten layer, and a fully connected layer. Each max pooling layer is sequentially located after each convolutional layer. The Flatten layer flattens the output of the last pooling layer into a one-dimensional vector and connects it to the fully connected layer; The convolutional layer is configured to extract image features reflected by each pixel in the image in a combined or independent manner, and the extracted image features include texture features and color features; The max pooling layer is configured to reduce the spatial size of the feature map to accelerate the training speed, reduce the risk of overfitting, and add an aggregation operation to merge the feature values in the local area into a representative feature value, suppressing noise and redundant information while retaining important feature information; The fully connected layer is configured to convert the original output into a probability distribution based on the Softmax function and use the highest probability distribution as the health assessment result.
5. The method for evaluating the health status of coral reefs based on deep learning according to claim 4, wherein The expression of the Softmax function is: where z i is the i-th element in the input vector, K is the total number of classes. Through the Softmax function, the output value of each class is converted into a probability value between 0 and 1, and the sum of the probabilities of all classes is 1.
6. The method for evaluating the health status of coral reefs based on deep learning according to claim 4, wherein, Training the convolutional neural network model using the second training set includes: Train the convolutional neural network model using the second training set, and define the loss function using categorical cross-entropy to measure the difference between the probability distribution predicted by the model and the true labels: BCE(p,y)= -[ylog(p)+(1 - y)log(1 - p)] where y is the true label and p is the probability of the positive class predicted by the model.
7. The coral reef health status assessment system based on deep learning according to claim 1, characterized in that The classification labels include: healthy, sub-healthy, and damaged.
8. A coral reef health status assessment system based on deep learning, characterized in that, It includes: A target recognition module, configured to obtain a coral reef image, input the coral reef image into an image target recognition model, and obtain a target recognition result; The image target recognition model is trained using a first training set, and the first training set includes: first original coral reef images and image position labels; An image processing module, configured to perform cropping processing on the coral reef image according to the target recognition result; A health status evaluation module, configured to input the processed image into a coral reef health status model to obtain a health evaluation result; the coral reef health status model is trained using a second training set, and the second training set includes: second original coral reef images and classification labels; the second training set generates the classification label corresponding to the original coral reef image by extracting the concentration of coral pathogenic bacteria in the water sample in the coral reef area and using the concentration of coral pathogenic bacteria as an evaluation index.
9. The coral reef health status assessment system based on deep learning according to claim 8, characterized in that, The image processing module includes: An image processing sub-module, configured to perform cropping processing on the coral reef image according to the target recognition result, and perform denoising and normalization operations on the cropped coral reef image to improve the image quality.