A method for inverting underwater topography in a coastal shallow sea

Abstract

The application discloses a kind of underwater topography inversion methods of coastal shallow sea, based on " remote sensing + deep learning " mode, realize high-precision inversion to underwater topography of coastal sea under the support of remote sensing theory and deep learning technology support.The inversion method includes constructing model based on convolutional neural network, through screening common part of multi-source data, after a series of processing, sample is made to build deep learning model, with the basis of coastal measured data, with multi-band remote sensing image as input, train model to achieve expected effect.The application fully considers the physical information contained in remote sensing image, establishes the mapping relationship of N*N sample window image based on measured data, overcomes the limitations that traditional measured depth method is time-consuming and labor-intensive and is easily affected by practical factors, and the limitations that traditional depth inversion algorithm needs sea surface roughness, weather data and other various data as support, improves depth inversion product quality and automation degree;Method is easy to implement, and can be used for dynamic monitoring and improving the quality of depth product.

Description

Technical Field

[0001] This invention belongs to the field of quantitative remote sensing and deep learning, and specifically relates to a method for underwater topography inversion in shallow waters near land. Background Technology

[0002] Coastal areas are greatly affected by both natural and human factors. Human land reclamation alters the marine ecology, tidal action drives the movement of sediment on the seabed, plate tectonics causes changes in underwater topography, and the existence of trenches and ridges causes changes in sea surface morphology. There is an urgent need for a product that can fully take into account these influencing factors, which is of great significance to scientific research.

[0003] Due to the unique characteristics of nearshore areas, basic hydrological and topographic observations have long faced the challenge of "difficulty in ship-based measurements on the beach and difficulty in land-based measurements at sea." Traditional depth measurement methods typically employ vessels equipped with depth sounding instruments, presenting a significant challenge in shallow waters and coastal areas with depths less than 5 meters. However, this method is inefficient and struggles to obtain complete data; areas with depths less than 2 meters are virtually impossible to measure. Such shallow, gently sloping beaches are widely distributed in my country, with typical examples including the underwater delta at the Yellow River estuary and the Subei radial sandbar area in the southern Yellow Sea, extending for tens of kilometers. New technologies are urgently needed to fill the gaps in shallow water underwater topographic data. Therefore, the measurement of nearshore shallow water underwater topography is of great importance, urgency, and challenge. Traditional depth inversion algorithms utilize backpropagation (BP) neural networks, considering the influence of parameters such as sediment, chlorophyll a, suspended sediment concentration, and seabed sediment on depth inversion. Factors with strong correlations are selected as input parameters to train the model and obtain the final depth inversion result. However, this method requires multiple hydrological and water quality parameters, which is extremely inconvenient.

[0004] In response to the shortcomings of existing technologies, this invention proposes a product that combines remote sensing and deep learning, aiming to change the current predicament and better serve the monitoring of the water environment. Summary of the Invention

[0005] This invention addresses the limitations of existing technologies, such as the difficulty of measuring depth from ships on the beach and from land at sea, the time-consuming and labor-intensive nature of traditional water depth measurement methods, and the susceptibility to real-world factors, as well as the requirement of various data sources, including sea surface roughness and meteorological data, for traditional water depth inversion algorithms. It provides a method for underwater topography inversion in shallow coastal waters to improve the quality and automation of water depth inversion products.

[0006] This invention is achieved using the following technical solution: a method for underwater topography inversion in shallow coastal waters, comprising the following steps:

[0007] Step S1: Preprocess the original remote sensing images to highlight the response of seabed topographic changes on the images, and extract the effective dataset based on the measured data.

[0008] Step S2: Create model training and test datasets: Based on the effective dataset obtained in step S1, extract and create training and test datasets with the guarantee of measured water depth data.

[0009] Step S3: Build a deep learning model: Build a deep learning model using the Tensorflow framework and train the model with the support of the training dataset;

[0010] Step S4, Quality Evaluation and Application: Validate the model trained in Step S3 based on the test dataset, extract multiple sets of model training errors, and perform dynamic evaluation: If the model error meets the expected effect and passes the quality evaluation, the model quality is determined to be reasonable; if it exceeds the quality requirement range, return to S3 to adjust the model.

[0011] The optimal model with error matching the expected effect is used to perform water depth inversion on specific images to determine its actual application effect, water depth transition trend, and accuracy at the sea-land interface.

[0012] Furthermore, in step S3, the constructed deep learning model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a regression layer connected in sequence, with each of the convolutional layer, pooling layer, and fully connected layer consisting of two layers.

[0013] 1) Input layer: Starting from the input layer, different neural network structures are used to transform the three-dimensional matrix of the previous layer into the three-dimensional matrix of the next layer, until the fully connected layer;

[0014] 2) Convolutional layer: Extracts feature information from the image. The input of each node in the convolutional layer is a small piece of the previous neural network. Through in-depth analysis of each small piece, features with a higher degree of abstraction are obtained.

[0015] 3) Pooling layer: The pooling layer is used to further reduce the number of nodes in the fully connected layer, so as to reduce the number of parameters in the entire neural network. The pooling kernel size is 2*2.

[0016] 4) Fully connected layer: After Flattening the data and connecting two fully connected layers, the single-dimensional multi-parameter data is reduced in dimensionality and parameters and output to a 1*1 dimension;

[0017] 5) Regression layer: Used for classification problems to perform regression prediction. The output scale of the overall model is controlled at 1*1 dimension pixels.

[0018] Furthermore, in step S1, satellite data and satellite information are analyzed to determine the different band information contained in different satellites, which is used to extract the common bands between different satellites and integrate the extracted common bands into a TIFF image; then, the multi-band data is subjected to corresponding data filtering and reflectivity conversion processing to extract the effective dataset.

[0019] Furthermore, step S1, in the process of filtering the data, specifically includes the following steps:

[0020] (1) Preliminary screening: Based on the USGS official website, multi-source data were obtained. First, the latitude and longitude distribution of the measured data was analyzed to determine the distribution of Landsat satellite strip numbers in a certain area. Satellite data was queried based on the strip numbers. The screening conditions were that the cloud cover was less than 20% and the dates should avoid winter.

[0021] (2) Secondary screening: The preliminary screening results are further refined by using ENVI to read the initial image data and manually checking the cloud coverage and image quality.

[0022] (3) Repair the stripe problem that may occur during imaging due to sensor failure to ensure normal use in the future;

[0023] (4) Since the satellite images based on the dimensionless ground reflection signal intensity, i.e., DN value, received by the sensor, but the DN value cannot highlight key details, it is calculated and converted into remote sensing reflectance, and the missing data is supplemented.

[0024] Furthermore, in step S2, model training requires training and testing datasets as support. Therefore, after converting DN values ​​into remote sensing reflectance and completing the data, data augmentation is performed by rotating the data at different angles to increase model robustness and the amount of data.

[0025] Furthermore, in step S2, training datasets and test datasets are created based on measured data. Under the premise of setting image boundary values, samples of a specific window size (e.g., 9*9) are cropped from the remote sensing image with the latitude and longitude of the measured data as the center, while saving sample information and water depth information.

[0026] Furthermore, in step S4, the model's results are combined with the Normalized Difference Water Index (NDWI). This process, which involves calculating the NDWI again after the model has already removed land, leverages the complementary advantages of the two methods to achieve accurate land-sea boundary identification, a natural land-sea transition, and clean land removal.

[0027]

[0028] Green represents the green light band, and NIR represents the near-infrared band.

[0029] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0030] This scheme creatively combines remote sensing theory with deep learning technology using the "remote sensing + deep learning" approach. It integrates multi-source data to extract common bands, performs data augmentation on multi-source and multi-band data to highlight image details and increase data volume, and builds a high-precision and high-efficiency water depth inversion model with the support of the Tensorflow deep learning framework and convolutional neural network.

[0031] Furthermore, it takes into account the physical mechanisms controlling the spatial pattern of optical characteristics and the ocean radiation transmission process. It not only solves the limitations of traditional measured water depth methods, which are time-consuming, labor-intensive, and easily affected by real-world factors, but also the limitations of traditional water depth inversion algorithms, which require multiple data such as sea surface roughness and meteorological data for support. It uses N*N pixels as the input of a convolutional deep learning network to realize the inversion of water depth in near-shore waters. It breaks through the defects of traditional "image-based but spectrum-based" pure microwave methods and "spectrum-based but image-based" pure optical methods, forming a "image-spectrum combined" water depth inversion theory. This can improve the regional applicability and reliability of the inversion model and improve the quality and automation of water depth inversion products. Attached Figure Description

[0032] Figure 1 This is a flowchart of data source processing, model building, and application according to an embodiment of the present invention;

[0033] Figure 2 This is a distribution diagram of measured data from an embodiment of the present invention;

[0034] Figure 3 This is a model structure diagram of an embodiment of the present invention;

[0035] Figure 4 These are training results for different sample sizes according to embodiments of the present invention;

[0036] Figure 5 These are diagrams showing the actual application results of different sample sizes according to the embodiments of the present invention. Detailed Implementation

[0037] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may be practiced in other ways than those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0038] Remote sensing images contain a wealth of information. In visible light remote sensing, the near-infrared band exhibits strong absorption of water bodies, while in microwave remote sensing, the response of seabed topography to the image can be directly observed. Considering the crucial controlling role of underwater topography in the spatial distribution pattern of water body optical properties in near-shore waters, and the varying degrees of reflection of incident light by the seabed sediment, this invention, taking into account both the physical mechanisms controlling the spatial pattern of optical properties and ocean radiative transfer processes, combines microwave theory with visible light remote sensing, employing new technologies and ideas to propose a method for inverting underwater topography in near-shore shallow waters. Figure 1 As shown, it includes the following steps:

[0039] Step S1: Preprocess the original remote sensing images to highlight the response of seabed topographic changes on the images, and extract the effective dataset based on the measured data.

[0040] By analyzing satellite data and information, we can determine the different spectral bands contained in different satellites and extract the common spectral bands between different satellites. We can also perform data filtering, reflectivity conversion and other processing on the multi-band data according to conditions such as cloud cover to extract the effective dataset.

[0041] Step S2: Create model training and test datasets: Based on the extraction of common bands between different satellites and the effective dataset obtained after data filtering and reflectivity conversion, and with the guarantee of measured water depth data, training and test datasets are extracted and created according to the latitude and longitude of the measured water depth data in a 9*9 window size.

[0042] Step S3: Construct a deep learning model: Based on building a convolutional neural network (CNN) using Tensorflow, construct a deep learning model containing two convolutional layers, two pooling layers, and two fully connected layers according to actual needs, and train the model with the support of the training dataset.

[0043] Step S4, Quality Evaluation and Application: Use the test dataset to validate the trained model, extract multiple sets of model training errors, and conduct dynamic evaluation by comparison. If the model error passes the quality evaluation (MRE≤20%), the model quality is deemed reasonable. If it exceeds the quality requirement range, return to S3 to adjust the model.

[0044] The optimal model with error matching the expected effect is used to perform water depth inversion on specific images to determine its actual application effect, water depth transition trend, and accuracy at the sea-land interface.

[0045] In fact, the quality of the initial images directly affects the final model training results. When selecting initial images, factors such as imaging time, cloud cover, and region need to be considered to avoid problematic data and areas with poor image quality. A second, more refined screening is then performed on the downloaded, initially screened data to remove problematic data from the details, ensuring the overall data quality. In step S1, this example uses multi-source data Landsat-4 / 5, Landsat-7, and Landsat-8 as examples. Through satellite data and information analysis, the different band information contained is determined, and common bands between different satellites are extracted. Landsat5, Landsat7, and Landsat8 share a total of 6 common bands. Based on the actual situation, multi-band data with less cloud and ice and better imaging conditions are selected. After appropriate reflectivity conversion and other processing, the effective dataset is extracted. Specifically, the data screening includes:

[0046] (1) Preliminary screening stage: For multi-source data acquired from the USGS official website, the first step is to analyze the latitude and longitude distribution of the measured data, such as... Figure 2 Figures (a) and (b) show the distribution of measured data used in this model (water depth data is highly sensitive and classified, and is generally not publicly disclosed). The model determines the distribution of Landsat satellite stripe numbers in a given area. Satellite data is queried based on the stripe numbers, with filtering conditions set to cloud cover less than 20% and dates avoiding winter as much as possible. Suitable imagery is obtained under these filtering conditions.

[0047] (2) Secondary screening stage: The preliminary screening results are further refined. ENVI is used to read the initial image data and the cloud coverage and image quality are manually checked. The images that pass this stage will be used for subsequent processing.

[0048] (3) Since its successful launch in 1972, the Landsat series of satellites has accumulated a large amount of satellite data, which is very helpful for studying long-term series variation problems. However, Landsat-7 may have striping problems during imaging due to sensor failure. Therefore, striping needs to be repaired before using Landsat-7 data to ensure normal use in the future.

[0049] (4) Satellite imaging is divided into two types: active and passive. Active imaging relies on sensors to receive signals actively transmitted by the satellite reflected from the ground, while passive imaging relies on sensors to receive sunlight signals reflected from the ground. Under normal circumstances, satellite imaging reflects the intensity of the reflected signal in the target area, i.e., the DN value, which needs to be converted into remote sensing reflectance to highlight image details and reflect its physical meaning.

[0050] In step S2, the data processed in S1 needs to be further processed for use in subsequent model training:

[0051] (1) Since the Landsat data resolution is 30m, if we need to consider both the image as input and its underlying physical meaning, the sample window design cannot be too large or too small. Research has shown that a window size of 9*9 is more appropriate. From a physical perspective, it can take into account the influence of a 270m*270m range on the water depth at the center point. From the perspective of deep learning implementation, the sample size for extracting important information from the image should not be smaller.

[0052] (2) ENVI's secondary development language, IDL, is specifically designed for image data processing and offers excellent operability for image processing. Therefore, the sample cropping program was written using IDL. The main idea of ​​the image cropping algorithm is to read the Landsat header file, obtain its size, boundary latitude and longitude, and other information, and convert the image latitude and longitude information and the measured water depth latitude and longitude information into image row and column numbers. Then, the image is traversed to check if the measured data row and column numbers exist, thereby determining whether to perform a cropping operation. During the cropping operation, the samples are named using the image information and water depth information, and recorded using a TXT document. To improve the robustness of the subsequent model, data augmentation processing is required for the cropped samples. Rotating each group of samples by a certain angle can not only improve the model's fault tolerance but also increase the number of samples. The same naming and recording method is used for the sample data after the rotation operation.

[0053] (3) Tfrecords files are a TensorFlow-specific format for creating and reading samples. Before creating sample files, the cropped samples need further processing. All samples need to be divided into training and testing datasets, and both parts need to be shuffled. This step requires reading a previously saved TXT document, obtaining sample information, classifying the samples based on the information, and saving the classified sample information in new TXT files. During this process, it is necessary to calculate the maximum and minimum values ​​of pixels in all samples for use in subsequent model training normalization.

[0054] Step S3 involves building a Convolutional Neural Network (CNN) using TensorFlow. This involves constructing a deep learning model containing two convolutional layers, two pooling layers, and two fully connected layers, based on actual needs. This is the core of this solution.

[0055] Artificial intelligence has developed rapidly in recent years, and its integration with the field of remote sensing has become increasingly close. However, the most commonly used method is currently the backpropagation (BP) neural network, a machine learning approach that uses remote sensing imagery and meteorological data to achieve point-to-point water depth prediction. Traditional methods suffer from low accuracy and slow fitting speed. To address this problem, this invention creatively combines deep learning methods with remote sensing, achieving point-to-area water depth inversion. The convolutional neural network (CNN) structure constructed in this invention is as follows: Figure 3 As shown, its structure can be divided into five parts: input layer, convolutional layer, pooling layer, fully connected layer, and regression layer. Figure 3 ,in:

[0056] 1) Input Layer: The input layer is the input to the entire neural network. In convolutional neural networks that process images, it typically represents the pixel matrix of an image. For example, in... Figure 3 In the diagram, the width and height of the leftmost 3D matrix represent the image size, while the depth of the 3D matrix represents the image's color channels. Starting from the input layer, the convolutional neural network transforms the 3D matrix of the previous layer into the 3D matrix of the next layer through different neural network structures, until the final fully connected layer;

[0057] 2) Convolutional layer: The convolutional layer is the most important part of a convolutional neural network. Unlike the traditional fully connected layer, the input of each node in the convolutional layer is only a small part of the neural network in the previous layer. The convolutional layer attempts to analyze each small part of the neural network more deeply to obtain features with a higher degree of abstraction. Generally speaking, the node matrix processed by the convolutional layer becomes deeper.

[0058] 3) Pooling Layers: Pooling layers in a neural network do not change the depth of the 3D matrix, but they can reduce the size of the matrix. The pooling operation transforms a high-resolution image into a lower-resolution image. Through pooling layers, the number of nodes in the final fully connected layer can be further reduced, thereby reducing the number of parameters in the entire neural network.

[0059] 4) Fully connected layer: such as Figure 3 As shown, after multiple rounds of convolutional and pooling layers, the final classification result in a convolutional neural network is typically provided by one or two fully connected layers. After several rounds of convolutional and pooling layers, the information in the image can be considered to have been abstracted into features with higher information content. Convolutional and pooling layers are part of the automatic image feature extraction process. After feature extraction is complete, fully connected layers are still needed to complete the classification task.

[0060] 5) Regression layer: The regression layer is mainly used for classification problems. The commonly used layer is the Softmax layer, but this solution requires regression prediction, so the Softmax layer is replaced by two fully connected layers.

[0061] In this embodiment, the model details need to be designed according to the specific application scenario, including the number of convolutional layers, convolutional kernel size, number of convolutional kernels, number of pooling layers, pooling kernel size, number of pooling kernels, activation function, and gradient descent algorithm.

[0062] In practical applications, a Landsat 8 image is approximately 8000*8000 pixels in size. To avoid obvious striped boundaries in the prediction results, the model output cannot be in units of surfaces. If surfaces are used as the output, the prediction results will fluctuate significantly in areas with complex hydrological environments, resulting in obvious segmentation between surfaces and an inability to achieve a smooth transition. Therefore, this invention controls the overall model output scale to 1*1 pixels, which is the regression layer mentioned in 5). The data is flattened using Flatten and then connected to two fully connected layers to reduce the dimensionality and parameters of the single-dimensional multi-parameter data and output it to a 1*1 dimension, which is the fully connected layer mentioned in 4).

[0063] Similarly, during the input process, since Landsat's resolution is 30m*30m, the actual range of a single image is approximately 240km*240km. To fully consider the physical and spectral information contained in the remote sensing image, it needs to be input using an N*N window, and the window size needs to be set reasonably. Specifically, to study the influence of surrounding hydrological conditions on the center point, the window size needs to be cropped to an odd number. In this embodiment, a 9*9 window as the input size yields better training results and can fully consider the influence of hydrological conditions within a 270m*270m range on the water depth value at the center point, which is the input layer mentioned in 1). Here, because Landsat has 6 common bands, the input latitude is 6, and the overall sample size is 9*9*6.

[0064] Because the above steps need to consider practical requirements, the input size is significantly smaller than that required by traditional mature deep learning networks such as VGG (224*224). These large networks often have dozens or even hundreds of hidden layers, which could easily lead to gradient vanishing or overfitting problems, causing model collapse, in the practical application of this invention. Therefore, image loss needs to be minimized in the convolutional and pooling layers of this invention. As described in 2) and 3), convolution is used to extract feature information from the image, and pooling is used to reduce the size and increase computational efficiency. Therefore, after two convolutional and pooling operations, the image size becomes 3*3, reaching its limit.

[0065] The forward propagation process of a neural network is linear, which cannot fit nonlinear data well and therefore cannot be applied to most practical production. By using an activation function, the parameters are excited to be nonlinear during the forward propagation process. The ReLU function is shown below. Adding this activation function can make the model converge faster and avoid linear propagation.

[0066] f(x) = max(x, 0) (1)

[0067] During the final call, the dimensions of the input image are obtained, a temporary cache is created according to these dimensions, and the predicted values ​​are saved line by line. After the inversion is completed, the cache is used to restore the image according to the predetermined dimensions to obtain the final prediction result.

[0068] In step S4, the accuracy of the trained model needs to be evaluated, which is divided into two parts: mean relative error (MRE) verification and actual spatial distribution trend verification. If the model's mean relative error meets the expected performance (MRE ≤ 20%), the next step of actual spatial distribution trend verification is performed; otherwise, the model is optimized until the requirements are met. In this scheme, the mean relative error (MRE) is 16.8%, which meets the expected requirements. Figure 4 As shown.

[0069] In particular, since the model is used to invert water depth, although land samples are added to identify and remove land, the total amount of land samples cannot account for a large proportion, so the identification of land cannot be 100%. Therefore, this scheme combines the model running results with the Normalized Difference Water Index (NDWI), as shown in formula (2). On the basis of the model itself removing land, the normalized difference water index is calculated again to achieve the complementary advantages of accurate land-sea boundary identification, natural land-sea transition and clean land removal effect, as shown in formula (2). Figure 5 The image shown is a diagram of the water depth in a nearshore area obtained using the model.

[0070]

[0071] Green represents the green light band, and NIR represents the near-infrared band.

[0072] In summary, this solution effectively improves the quality, convenience, and automation of underwater topography inversion by selecting common components from multi-source data and creating samples through a series of processing steps to build a deep learning model.

[0073] This invention uses the Landsat series satellites as an example to outline its implementation principle, overall process, and model structure. Specifically, a 9*9 window is used as a sample system for model construction. However, this invention can also be implemented in other ways different from those described herein. All research methods and implementation processes mentioned in this invention are applicable to other series of satellites, other sensors, and other window sizes.

[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for underwater topography inversion in shallow coastal waters, characterized in that, Includes the following steps: Step S1: Preprocess the original remote sensing images to highlight the response of seabed topographic changes on the images, and extract the effective dataset based on the measured data. By analyzing satellite data and information, we can identify the different band information contained in different satellites, extract the common bands between different satellites, and integrate the extracted common bands into a TIFF image. Then, we can perform corresponding data filtering and reflectivity conversion on the multi-band data to extract the effective dataset. The specific steps involved in data filtering are as follows: (1) Preliminary screening: Based on the USGS official website, multi-source data is obtained. First, the latitude and longitude distribution of the measured data is analyzed to determine the distribution of Landsat satellite strip numbers in a certain area. Satellite data is then queried based on the strip numbers. (2) Secondary screening: The preliminary screening results are further refined by using ENVI to read the initial image data and manually checking the cloud coverage and image quality. (3) Repair the striping problem that may occur during imaging due to sensor failure to ensure normal use in the future; (4) Image is formed based on the dimensionless ground reflection signal intensity DN value received by the sensor, the DN value is calculated and converted into remote sensing reflectance, and missing data is filled in; Step S2: Create model training and test datasets: Based on the effective dataset obtained in step S1, extract and create training and test datasets with the guarantee of measured water depth data. Step S3: Build a deep learning model: Build a deep learning model using the Tensorflow framework and train the model with the support of the training dataset; The constructed deep learning model includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a regression layer connected in sequence. Each of the convolutional layer, pooling layer, and fully connected layer consists of two layers. 1) Input layer: The image is input in an N*N window and the window size is cropped according to an odd number. Starting from the input layer, the three-dimensional matrix of the previous layer is transformed into the three-dimensional matrix of the next layer through different neural network structures, until the fully connected layer. 2) Convolutional layer: Extracts feature information from the image. The input of each node in the convolutional layer is a small piece of the previous neural network. Through in-depth analysis of each small piece, features with a higher degree of abstraction are obtained. 3) Pooling layer: The pooling layer is used to further reduce the number of nodes in the fully connected layer, thereby reducing the number of parameters in the entire neural network; 4) Fully connected layer: After Flattening the data and connecting two fully connected layers, the single-dimensional multi-parameter data is reduced in dimensionality and parameters and output to a 1*1 dimension; 5) Regression layer: Commonly used in classification problems to perform regression prediction. The output scale of the overall model is controlled at 1*1 dimension pixels. Step S4, Quality Evaluation and Application: Validate the model trained in Step S3 based on the test dataset, extract multiple sets of model training errors, and perform dynamic evaluation: If the model error meets the expected effect and passes the quality evaluation, the model quality is determined to be reasonable; if it exceeds the quality requirement range, return to S3 to adjust the model. The optimal model with error matching the expected effect is used to perform water depth inversion on specific images to determine its actual application effect, water depth transition trend and accuracy at the sea-land junction. By combining the model results with the Normalized Difference Water Index (NDWI), and then performing another round of NDWI calculation on top of the model's own land exclusion process, the complementary advantages of the two methods are achieved, resulting in accurate land-sea boundary identification, a natural land-sea transition, and clean land exclusion. Green represents the green light band, and NIR represents the near-infrared band.

2. The underwater topography inversion method for near-shore shallow waters according to claim 1, characterized in that: In step S1, the initial screening criteria are cloud cover of less than 20% and dates that avoid winter.

3. The underwater topography inversion method for near-shore shallow waters according to claim 2, characterized in that: In step S2, model training requires training and testing datasets as support. Therefore, after converting DN values ​​into remote sensing reflectance and completing the data, data augmentation is performed by rotating the data at different angles to increase model robustness and the amount of data.

4. The underwater topography inversion method for near-shore shallow waters according to claim 1, characterized in that: In step S2, training and testing datasets are created based on measured data. Under the premise of setting image boundary values, samples of a specific window size are cropped from the remote sensing image with the latitude and longitude of the measured data as the center, while saving sample information and water depth information.

Images