Coral reef distribution identification method and system based on remote sensing image
By introducing the Multi-Scale Attention UNet model and preprocessing techniques, the subjectivity and confusion issues of traditional coral reef remote sensing identification were resolved, achieving high-precision coral reef distribution identification and improving the identification effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SURVEYING & MAPPING INST LANDS & RESOURCE DEPT OF GUANGDONG PROVINCE
- Filing Date
- 2025-12-10
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional coral reef remote sensing identification technology relies on expert experience, is highly subjective, inefficient, and struggles to identify details and internal structures. Furthermore, ordinary algorithms are unable to handle the problem of fragmented coral reef morphology being confused with seagrass beds and sandy areas.
The Multi-Scale Attention UNet model is adopted, which combines the hollow spatial pyramid pooling module and dual attention mechanism for remote sensing image preprocessing and model training. Noise interference is suppressed through atmospheric correction, flare removal and water body radiation correction. The model is optimized using a hybrid loss function to output a coral reef distribution map.
It significantly improves the accuracy and consistency of coral reef distribution identification, effectively suppresses interference from waves, uneven lighting, and water depth noise, and enhances the accuracy of identification results.
Smart Images
Figure CN121962885A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of therapeutic device control, and more specifically, to a method and system for identifying coral reef distribution based on remote sensing images. Background Technology
[0002] Coral reef ecosystems, hailed as "rainforests of the ocean," cover less than 0.1% of the ocean floor but nurture approximately 25% of the world's marine life, possessing immeasurable ecological value. They also provide fisheries, coastal protection, and tourism for coastal communities, sustaining the livelihoods of hundreds of millions globally. However, coral reefs face multiple threats, including climate change, ocean acidification, pollution, and overexploitation, with their degradation occurring at an alarming rate. Therefore, large-scale, high-frequency, and precise monitoring of coral reef distribution and health status is crucial for ecological protection, restoration, and management decisions. Traditionally, coral reef monitoring relies primarily on in-situ surveys by divers. While highly accurate, this method is time-consuming, labor-intensive, and costly, and struggles to cover large areas or remote sea regions. Remote sensing technology, due to its macroscopic, rapid, and periodic observation capabilities, has become a core tool for large-scale coral reef surveys and dynamic monitoring. Early developments in coral reef remote sensing identification technology relied mainly on experts visually interpreting low-to-medium resolution satellite imagery (such as Landsat) to delineate the approximate extent of coral reefs. This method relies heavily on expert experience, is highly subjective, inefficient, and struggles to identify details and internal structures, making it only suitable for large-scale surveys. In recent years, with the widespread availability of multispectral data (such as SPOT and IKONOS), the method has entered a digital phase. These methods treat each pixel as an independent sample, classifying it based on its spectral values in different bands. However, these methods primarily rely on manually selected and designed spectral features. While they perform reasonably well when dealing with areas with good spectral separation of ground features, they are less effective in complex environments like coral reefs. Due to the fragmented morphology of coral reefs and their tendency to be confused with seagrass beds and sandy areas, ordinary algorithms struggle to extract deep semantic features, resulting in poor recognition performance. Summary of the Invention
[0003] Therefore, it is necessary to provide a method and system for identifying coral reef distribution based on remote sensing images that has better recognition performance in response to the above-mentioned technical problems.
[0004] One method for identifying coral reef distribution based on remote sensing imagery includes: S1: Acquire the original remote sensing images and preprocess them to obtain preprocessed data. Construct a coral reef sample dataset based on the preprocessed data. S2: Construct a coral reef distribution identification model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-Scale AttentionUNet model; S3: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; S4: Acquire the image to be tested and identify the coral reef distribution based on the trained coral reef distribution recognition model.
[0005] Furthermore, in step S1, the preprocessing specifically includes: The original remote sensing images were subjected to atmospheric correction, flare removal, and water body radiation correction.
[0006] Furthermore, the original remote sensing images undergo atmospheric correction, flare removal, and water body radiation correction specifically as follows: Atmospheric correction: Using the FLAASH model, the radiance of the original remote sensing image is converted to reflectance above the sea surface; thus obtaining the pre-processed original remote sensing image. Solar flare removal: The land, clouds, and obvious solar flare areas in the pre-processed original remote sensing image are masked out, and the focus is on the ocean area; thus, a further pre-processed original remote sensing image is obtained. Water body radiation correction: Based on the original remote sensing images that have undergone further preprocessing, the multi-band reflectance images are processed using the Lyzenga model; a depth-invariant index image is generated or the seabed reflectance image is directly derived; the depth-invariant index image or the directly derived seabed reflectance image is used as the preprocessed data.
[0007] Furthermore, the original remote sensing images are Sentinel-2 high-resolution remote sensing images or WorldView high-resolution remote sensing images; The land, clouds, and obvious solar flare areas in the initial preprocessed original remote sensing image are masked out, and the focus is on the ocean area; the original remote sensing image for further preprocessing is obtained by: using the strong absorption characteristics of the near-infrared band in deep water areas, a linear regression model is established to achieve flare removal.
[0008] Further, in step S2, the Multi-Scale Attention UNet model includes: Encoder: ResNet-50 is used as the backbone network. It consists of 5 stages. In each stage, the feature map size is reduced to half of the original size and the number of channels is doubled, which is used to extract multi-level features from shallow water color to deep reef texture. Decoder: It includes four upsampling operations. After each upsampling, the feature map is concatenated with the attention-weighted encoder feature map to gradually restore the image resolution to the original image size.
[0009] Furthermore, in step S2, the void space pyramid pooling module specifically comprises: This is set after the deepest output of the encoder. This module contains 5 parallel branches: A 1x1 convolutional layer.
[0010] Three 3x3 hollow convolutional layers were used, with dilation rates set to 6, 12, and 18, respectively.
[0011] A globally averaged pooling layer: capable of adapting to the varying sizes of coral reef patches and expanding the receptive field.
[0012] In step S2, the dual attention mechanism is specifically as follows: An attention module is embedded before the encoder features are passed to the decoder. Channel attention: Suppress channel responses representing deep-sea or turbid water bodies by learning channel weights through global pooling and MLP; Spatial attention: Spatial weights are learned through convolutional layers to focus on the complex coral reef area while ignoring the smooth sand background.
[0013] Further, step S3 includes: Using a hybrid loss function to address the imbalance between positive and negative samples:
[0014] in, Used to optimize the overlap between the predicted region and the real region; Used to increase the weight of hard-to-classify samples.
[0015] Further, in step S4, specifically: The image to be tested is input into the trained coral reef distribution recognition model, which outputs a probability map. A binary mask is obtained by thresholding the probability map, and morphological opening is used to remove isolated noise points, resulting in the final coral reef distribution map. Furthermore, this invention also provides a coral reef distribution identification system based on remote sensing imagery, comprising: Acquisition module: Acquires raw remote sensing images and performs preprocessing to obtain preprocessed data, and constructs a coral reef sample dataset based on the preprocessed data; Model building module: Constructs a coral reef distribution recognition model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-ScaleAttention UNet model; Training module: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; Recognition module: Acquires the image to be tested and realizes coral reef distribution recognition based on the trained coral reef distribution recognition model.
[0016] This invention utilizes a hollow spatial pyramid pooling module to effectively capture multi-scale features, resulting in a significant improvement in IoU. Furthermore, this invention effectively suppresses noise interference from ocean waves, uneven lighting, and varying water depths through preprocessing, leading to more accurate recognition results. Attached Figure Description
[0017] Figure 1 This is a flowchart of a method for identifying coral reef distribution based on remote sensing imagery in one embodiment; Figure 2 This is a block diagram of a coral reef distribution identification system based on remote sensing imagery in one embodiment; Detailed Implementation To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0018] Example 1: This embodiment provides, as follows: Figure 1 The method for identifying coral reef distribution based on remote sensing imagery, as shown, includes: S1: Acquire the original remote sensing images and preprocess them to obtain preprocessed data. Construct a coral reef sample dataset based on the preprocessed data. S2: Construct a coral reef distribution identification model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-Scale AttentionUNet model; S3: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; S4: Acquire the image to be tested and identify the coral reef distribution based on the trained coral reef distribution recognition model.
[0019] This embodiment effectively captures multi-scale features using a hollow spatial pyramid pooling module, resulting in a significant improvement in IoU. Furthermore, the invention effectively suppresses noise interference from ocean waves, uneven lighting, and varying water depths through preprocessing, leading to more accurate recognition results.
[0020] Example 2: This embodiment further discloses information based on Embodiment 1: In step S1, the preprocessing specifically involves: The original remote sensing images were subjected to atmospheric correction, flare removal, and water body radiation correction.
[0021] The specific steps for atmospheric correction, flare removal, and water body radiation correction of the original remote sensing images are as follows: Atmospheric correction: Using the FLAASH model, the radiance of the original remote sensing image is converted to reflectance above the sea surface; thus obtaining the pre-processed original remote sensing image. Solar flare removal: The land, clouds, and obvious solar flare areas in the pre-processed original remote sensing image are masked out, and the focus is on the ocean area; thus, a further pre-processed original remote sensing image is obtained. Water body radiation correction: Based on the original remote sensing images that have undergone further preprocessing, the multi-band reflectance images are processed using the Lyzenga model; a depth-invariant index image is generated or the seabed reflectance image is directly derived; the depth-invariant index image or the directly derived seabed reflectance image is used as the preprocessed data.
[0022] The original remote sensing images are Sentinel-2 high-resolution remote sensing images or WorldView high-resolution remote sensing images; The land, clouds, and obvious solar flare areas in the initial preprocessed original remote sensing image are masked out, and the focus is on the ocean area; the original remote sensing image for further preprocessing is obtained by: using the strong absorption characteristics of the near-infrared band in deep water areas, a linear regression model is established to achieve flare removal.
[0023] In step S2, the Multi-Scale Attention UNet model includes: Encoder: ResNet-50 is used as the backbone network. It consists of 5 stages. In each stage, the feature map size is reduced to half of the original size and the number of channels is doubled, which is used to extract multi-level features from shallow water color to deep reef texture. Decoder: It includes four upsampling operations. After each upsampling, the feature map is concatenated with the attention-weighted encoder feature map to gradually restore the image resolution to the original image size.
[0024] In step S2, the void space pyramid pooling module specifically works as follows: This is set after the deepest output of the encoder. This module contains 5 parallel branches: A 1x1 convolutional layer.
[0025] Three 3x3 hollow convolutional layers were used, with dilation rates set to 6, 12, and 18, respectively.
[0026] A globally averaged pooling layer: capable of adapting to the varying sizes of coral reef patches and expanding the receptive field.
[0027] In step S2, the dual attention mechanism is specifically as follows: An attention module is embedded before the encoder features are passed to the decoder. Channel attention: Suppress channel responses representing deep-sea or turbid water bodies by learning channel weights through global pooling and MLP; Spatial attention: Spatial weights are learned through convolutional layers to focus on the complex coral reef area while ignoring the smooth sand background.
[0028] Step S3 includes:
[0029] in, Used to optimize the overlap between the predicted region and the real region; Used to increase the weight of hard-to-classify samples.
[0030] In step S4, specifically: The image to be tested is input into the trained coral reef distribution recognition model, which outputs a probability map. The probability map is then segmented by thresholding to obtain a binary mask, and morphological opening is used to remove isolated noise points, resulting in the final coral reef distribution map.
[0031] This embodiment effectively captures multi-scale features using a hollow spatial pyramid pooling module, resulting in a significant improvement in IoU. Furthermore, the invention effectively suppresses noise interference from ocean waves, uneven lighting, and varying water depths through preprocessing, leading to more accurate recognition results.
[0032] Example 3: This embodiment provides, as follows: Figure 2 The system shown is a coral reef distribution identification system based on remote sensing imagery, comprising: Acquisition module: Acquires raw remote sensing images and performs preprocessing to obtain preprocessed data, and constructs a coral reef sample dataset based on the preprocessed data; Model building module: Constructs a coral reef distribution recognition model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-ScaleAttention UNet model; Training module: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; Recognition module: Acquires the image to be tested and realizes coral reef distribution recognition based on the trained coral reef distribution recognition model.
[0033] This embodiment effectively captures multi-scale features using a hollow spatial pyramid pooling module, resulting in a significant improvement in IoU. Furthermore, the invention effectively suppresses noise interference from ocean waves, uneven lighting, and varying water depths through preprocessing, leading to more accurate recognition results.
[0034] Example 4: This embodiment provides a method for identifying coral reef distribution based on remote sensing imagery, including: Raw remote sensing images are obtained by imaging the target sea area using multispectral or hyperspectral sensors mounted on satellites or airborne platforms. However, while recording the potential distribution area of coral reefs, the raw remote sensing images also include multiple interference signals from the atmosphere, turbulent sea surface and complex water bodies. Therefore, atmospheric correction is first performed. Using a radiative transfer model, the scattering effects caused by atmospheric molecules and aerosols, as well as the absorption effects of gases such as ozone and water vapor, are accurately calculated and removed. This converts the top-level radiance value received by the sensor into a physical quantity that reflects the reflectivity of the sea surface.
[0035] To address the unique noise characteristics of the sea surface, a flare removal process is implemented. A calm sea surface will strongly reflect direct sunlight onto the sensor, forming locally oversaturated bright spots that completely obscure valuable underwater information. By combining imaging geometry models with image restoration techniques, these flare areas are detected and suppressed, restoring the obscured image details. At the same time, random noise such as whitecaps caused by wave breaking is also smoothed or filtered out.
[0036] Water body radiometric correction addresses the issue that light signals, during their reflection from the seabed and through the water column to reach the sensor, undergo strong absorption and scattering caused by water molecules, dissolved organic matter, and suspended particles. This attenuation is not only intense but also exhibits significant wavelength selectivity, resulting in drastically different spectral responses for the same substrate at different water depths, while different substrates may have similar spectra at specific water depths. Therefore, it is necessary to employ semi-empirical models such as Lyzenga or more complex semi-analytical models to establish light transmission equations in water bodies to inversely determine or compensate for the influence of the water column. The ultimate goal is to estimate a reflectivity signal that is as close as possible to the actual reflectivity signal determined solely by the properties of the seabed substrate itself, or to generate a feature representation that is insensitive to changes in water depth. This removes water depth, the biggest confounding variable, from the classification problem, allowing subsequent algorithms to focus on the spectral and spatial differences of the substrate type itself.
[0037] Once the image data has undergone the aforementioned rigorous physical corrections to obtain relatively "pure" seabed reflection information, it enters the data-driven intelligent recognition stage. Constructing a high-quality, large-scale, and precisely labeled coral reef sample dataset is fundamental to this stage. This requires collaboration between marine ecologists and remote sensing interpretation experts. Based on field survey data, historical nautical charts, and high-resolution imagery, they meticulously delineate the precise extent of the coral reefs pixel-by-pixel on the preprocessed images, generating a binary mask image serving as the "ground truth." Considering the complexity of the natural distribution of coral reefs and the thirst for massive training data in deep learning models, systematically augmenting these labeled sample pairs is an effective strategy to improve the model's generalization ability. By applying transformations such as random rotation, horizontal and vertical flipping, scale scaling, and color perturbation, data diversity can be greatly increased, simulating how coral reefs appear from different observation angles, spatial scales, and lighting conditions. This prevents the model from being limited by a finite number of samples during the learning process, thus enabling it to better adapt to diverse application scenarios in the real world.
[0038] The core agent of the entire process is a meticulously designed deep learning model—MSA-UNet. It's a powerful feature extraction and segmentation network built upon the classic U-Net encoder-decoder segmentation architecture, deeply integrating advanced concepts such as residual connections, multi-scale context aggregation, and attention mechanisms. The model's training process is essentially its learning of the essential features of coral reefs: First, the input training image blocks enter the encoder path, which consists of a series of convolutional blocks embedded with residual connections. Each stage is accompanied by a downsampling operation, gradually increasing the network depth and expanding the receptive field. This allows the encoder to act like a hierarchical information filter, progressively abstracting increasingly complex features from the original pixels—from the most basic edges and textures, to medium-sized seagrass clusters or coral reef shapes, and then to the larger-scale overall reef structure and environmental context. The introduction of residual connections ensures that gradients can be effectively backpropagated even when the network is very deep, avoiding the vanishing gradient problem and guaranteeing the stability of training and the depth of feature representation capabilities. When features are compressed and refined to the lowest resolution, most semantically abstract bottleneck layer by the encoder, an innovative multi-scale feature aggregation module (typically employing a dilated spatial pyramid pooling (ASPP) structure) is activated. This module deploys multiple dilated convolutional layers with different dilation rates in parallel, enabling exponential expansion of the receptive field while maintaining the spatial resolution of the feature map. This allows for the simultaneous capture of multi-scale contextual information, from local details and medium-sized objects to the global scene. Simultaneously, a global average pooling branch is responsible for summarizing the overall semantics at the image level. These features from different scale perspectives are further processed and then stitched together and fused, allowing the model to consider both the subtle textures of its surroundings and the macroscopic layout at a distant distance when making decisions at any point. This design is crucial for addressing the scale effects and fragmented morphological features of coral reefs caused by changes in water depth.
[0039] Subsequently, the decoder path begins its work, gradually restoring the spatial resolution of the feature map through symmetrical upsampling operations. The goal is to map the high-level semantics learned by the encoder back to the original image size, achieving pixel-level accurate classification. The decoder does not work in isolation; its success heavily relies on skip connections from corresponding layers of the encoder. These connections pass on high-resolution detail features preserved during the encoding process, combining them with the deep semantic features upsampled by the decoder. In this model, this skip connection process is further intelligently implemented by integrating a Convolutional Block Attention (CBAM) module. Before feature fusion, the features from the encoder undergo "attention" recalibration: CBAM sequentially performs channel attention and spatial attention operations. The channel attention mechanism automatically evaluates the importance of each feature channel, strengthening those closely related to the coral reef's spectral and texture responses while suppressing irrelevant or noisy channels. The spatial attention mechanism acts like a focusing lens, calculating a weight map in the spatial dimension of the feature map, guiding the model to focus more on pixel locations that may contain coral reef edges or complex texture structures, rather than uniform seawater or sandy areas. This attention-guided skip connections ensure that when reconstructing high-resolution segmentation maps, the model can accurately and adaptively recall and fuse the most relevant details from the encoder, thereby significantly improving the clarity and coherence of segmentation boundaries and the ability to identify small coral patches.
[0040] The output of the entire network maps the number of channels to the number of categories through a 1x1 convolutional layer, and then uses a sigmoid activation function to generate a probability map for each pixel belonging to the coral reef. The training process of the model is driven and optimized by a carefully designed hybrid loss function that combines binary cross-entropy loss and Dice similarity coefficient loss. The former rigorously measures the difference between the predicted probability and the true label at the pixel level, while the latter optimizes based on the overall overlap of the regions, thus handling segmentation tasks such as coral reefs where the foreground and background areas are imbalanced.
[0041] Once the model has been fully trained and converged, the image to be detected is standardized by a preprocessing procedure identical to that of the training data, and then input into the trained MSA-UNet model. The network can output an initial probability map of coral reef distribution through a single forward propagation.
[0042] This embodiment effectively captures multi-scale features using a hollow spatial pyramid pooling module, resulting in a significant improvement in IoU. Furthermore, the invention effectively suppresses noise interference from ocean waves, uneven lighting, and varying water depths through preprocessing, leading to more accurate recognition results.
Claims
1. A method for identifying coral reef distribution based on remote sensing imagery, characterized in that, include: S1: Acquire the original remote sensing images and preprocess them to obtain preprocessed data. Construct a coral reef sample dataset based on the preprocessed data. S2: Construct a coral reef distribution identification model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-Scale AttentionUNet model; S3: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; S4: Acquire the image to be tested and identify the coral reef distribution based on the trained coral reef distribution recognition model.
2. The method for identifying coral reef distribution based on remote sensing imagery according to claim 1, characterized in that, In step S1, the preprocessing specifically involves: The original remote sensing images were subjected to atmospheric correction, flare removal, and water body radiation correction.
3. The method for identifying coral reef distribution based on remote sensing imagery according to claim 2, characterized in that, The specific steps for atmospheric correction, flare removal, and water body radiation correction of the original remote sensing images are as follows: Atmospheric correction: Using the FLAASH model, the radiance of the original remote sensing image is converted to reflectance above the sea surface; thus obtaining the pre-processed original remote sensing image. Solar flare removal: The land, clouds, and obvious solar flare areas in the pre-processed original remote sensing image are masked out, and the focus is on the ocean area; thus, a further pre-processed original remote sensing image is obtained. Water body radiation correction: Based on the original remote sensing images that have undergone further preprocessing, the multi-band reflectance images are processed using the Lyzenga model; a depth-invariant index image is generated or the seabed reflectance image is directly derived; the depth-invariant index image or the directly derived seabed reflectance image is used as the preprocessed data.
4. The method for identifying coral reef distribution based on remote sensing imagery according to claim 3, characterized in that, The original remote sensing images are Sentinel-2 high-resolution remote sensing images or WorldView high-resolution remote sensing images; The land, clouds, and obvious solar flare areas in the initial preprocessed original remote sensing image are masked out, and the focus is on the ocean area; the original remote sensing image for further preprocessing is obtained by: using the strong absorption characteristics of the near-infrared band in deep water areas, a linear regression model is established to achieve flare removal.
5. The method for identifying coral reef distribution based on remote sensing imagery according to claim 3, characterized in that, In step S2, the Multi-Scale Attention UNet model includes: Encoder: ResNet-50 is used as the backbone network; it contains 5 stages. After each stage, the feature map size is reduced to 1 / 2 of the original size and the number of channels is doubled, which is used to extract multi-level features from shallow water color to deep reef texture. Decoder: It includes four upsampling operations; after each upsampling, the feature map is concatenated with the attention-weighted encoder feature map to gradually restore the image resolution to the original image size.
6. The method for identifying coral reef distribution based on remote sensing imagery according to claim 5, characterized in that, In step S2, the void space pyramid pooling module specifically works as follows: This module is located after the deepest output of the encoder; it contains 5 parallel branches: A 1x1 convolutional layer; Three 3x3 hollow convolutional layers with dilation rates set to 6, 12, and 18, respectively; A globally averaged pooling layer: capable of adapting to the varying sizes of coral reef patches and expanding the receptive field.
7. The method for identifying coral reef distribution based on remote sensing imagery according to claim 5, characterized in that, In step S2, the dual attention mechanism is specifically as follows: An attention module is embedded before the encoder features are passed to the decoder. Channel attention: Suppress channel responses representing deep-sea or turbid water bodies by learning channel weights through global pooling and MLP; Spatial attention: Spatial weights are learned through convolutional layers to focus on the complex coral reef area while ignoring the smooth sand background.
8. The method for identifying coral reef distribution based on remote sensing imagery according to claim 7, characterized in that, Step S3 includes: Using a hybrid loss function to address the imbalance between positive and negative samples: in, Used to optimize the overlap between the predicted region and the real region; Used to increase the weight of hard-to-classify samples.
9. A method for identifying coral reef distribution based on remote sensing imagery according to claim 8, characterized in that, In step S4, specifically: The image to be tested is input into the trained coral reef distribution recognition model, which outputs a probability map. The probability map is then segmented by thresholding to obtain a binary mask, and morphological opening is used to remove isolated noise points, resulting in the final coral reef distribution map.
10. A coral reef distribution identification system based on remote sensing imagery, characterized in that, include: Acquisition module: Acquires raw remote sensing images and performs preprocessing to obtain preprocessed data, and constructs a coral reef sample dataset based on the preprocessed data; Model building module: Constructs a coral reef distribution recognition model, which is built by introducing a hollow spatial pyramid pooling module and a dual attention mechanism into the Multi-ScaleAttention UNet model; Training module: Train the coral reef distribution recognition model based on the coral reef sample dataset to obtain the trained coral reef distribution recognition model; Recognition module: Acquires the image to be tested and realizes coral reef distribution recognition based on the trained coral reef distribution recognition model.