A water depth inversion method fusing multi-temporal information and geographic spatial perception mechanism

Abstract

The present application relates to the technical field of marine remote sensing and water depth inversion, and discloses a water depth inversion method fusing multi-temporal information and geographic spatial perception mechanism, single-temporal remote sensing images of the same water area at different time points are acquired first to construct a multi-temporal remote sensing image sequence, then a robust principal component analysis method RPCA is used to fuse and process the multi-temporal remote sensing image sequence to obtain a fused image; then a plurality of core units are cascaded together to complete step-by-step feature extraction, and the last core unit is connected to a regression head to output continuous water depth values, thereby completing the construction of a water depth inversion network model; finally, the water depth inversion network model is trained by using the fused image combined with the geographic spatial position, and the trained water depth inversion network model is used to perform water depth inversion on the data-processed remote sensing image to be detected.

Description

Technical Field

[0001] This invention belongs to the technical field of marine remote sensing and water depth inversion, specifically relating to a water depth inversion method that integrates multi-temporal information and geospatial perception mechanisms. Background Technology

[0002] Shallow water areas (typically 0–20 m deep, with some nearshore areas reaching 30 m) are the most active zones for land-sea interaction. They are not only core areas for coastal engineering activities such as port construction, channel dredging, and breakwater construction, but also important spaces for fisheries production and marine resource development. Furthermore, nearshore shallow waters are widely home to typical ecosystems such as coral reefs and seagrass beds, making them among the most sensitive and vulnerable areas in the marine ecosystem. Therefore, obtaining high-precision, high-spatial-resolution shallow water depth information is crucial for ensuring navigation safety, optimizing coastal engineering planning, carrying out marine ecological protection, and mitigating coastal disaster risks.

[0003] Although the deep learning-based multispectral remote sensing water depth inversion method based on single-temporal imagery has improved in accuracy, it still has some limitations in the complex and ever-changing nearshore water environment. These problems restrict its stability and practical application capabilities to a certain extent, mainly in the following aspects.

[0004] (1) Insufficient utilization of time-series information, susceptible to instantaneous environmental noise. Current mainstream deep learning-based water depth retrieval methods are mostly based on single-scene imagery for modeling, failing to jointly utilize multi-temporal observation data of the same area. However, the nearshore marine environment exhibits significant temporal dynamics, and satellite imaging is easily affected by various random interference factors, such as ships and wakes, enhanced reflection from localized waves, thin clouds, and cloud shadows. These transient noises often manifest as abnormally high reflectivity or abnormally dark areas in the spectrum, easily misidentified by models as water depth changes or special bottom sediments, thus forming striped or patchy anomalous areas in the retrieval results. Due to the lack of mechanisms for utilizing redundant information in the temporal dimension, existing methods struggle to effectively distinguish between stable topographic signals and transient interference signals.

[0005] (2) Limited use of geospatial information makes it difficult to characterize spatial heterogeneity. Convolutional neural networks (CNNs) can extract neighborhood texture features through local receptive fields, but their structural characteristics typically exhibit translation invariance to the input location, meaning they assume different spatial locations have the same feature-target mapping relationship. In real nearshore environments, the relationship between water depth and spectrum often exhibits significant spatial non-uniformity. Most existing models only use spectral bands as input, without explicitly incorporating geospatial location information. This makes it difficult for the models to learn the water depth distribution trends at the regional scale and the differentiated spectral response patterns of different sub-regions, thus affecting their generalization ability in spatially heterogeneous waters.

[0006] (3) The feature extraction scale is relatively singular, making it difficult to simultaneously represent terrain features at different scales. Nearshore shallow water topography is characterized by the coexistence of multi-scale features, such as large-scale gentle-sloping seabeds alongside fine structures like local reefs and sand ridges. Some existing CNN models primarily use fixed-size convolutional kernels for feature extraction, resulting in a relatively limited receptive field. In this context, large-scale topographic trend information may be insufficiently extracted, while small-scale detailed structures are easily weakened during multiple convolutions and downsampling processes, thus affecting the ability to represent complex topographic changes. Summary of the Invention

[0007] This paper proposes a water depth inversion method that integrates multi-temporal information and geospatial perception mechanism, aiming to solve the problems of insufficient stability, limited spatial adaptability, and insufficient multi-scale terrain characterization capability of existing water depth inversion methods based on single-temporal multispectral remote sensing and deep learning in complex nearshore environments.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A water depth inversion method integrating multi-temporal information and geospatial sensing mechanisms includes the following steps: Step 1: Data Processing Single-temporal remote sensing images of the same water area at different time points are acquired, and these images are stacked according to time series to construct a multi-temporal remote sensing image sequence. Robust principal component analysis (RPCA) is then used to fuse the multi-temporal remote sensing image sequence to obtain a fused image. Step 2: Construct a water depth inversion network model Multiple core units are cascaded together to complete stepwise feature extraction, and the last core unit is connected to the regression head to output continuous water depth values. Each core unit includes a cascaded multi-scale convolutional module (MSB) and a channel attention module (SEL). The output features are residually concatenated with the original input features to obtain the final output features. The multi-scale convolutional module employs a parallel convolutional structure to extract small-scale and large-scale spatial features, respectively, and then fuses these features to obtain multi-scale spatial features. The channel attention module (SEL) generates channel weights and performs a dot product with the multi-scale spatial features to obtain recalibrated channel weights. Step 3: Train the water depth inversion network model using fused imagery and geospatial location, and then use the trained water depth inversion network model to perform water depth inversion on the processed remote sensing imagery to be inspected.

[0009] Furthermore, in step two, the multi-scale convolution module (MSB) includes a parallel first branch and a second branch, which are used to extract small-scale spatial features and large-scale spatial features, respectively. Both branches include depthwise separable convolution (DW) and pointwise convolution (PW). The kernel of the depthwise separable convolution (DW) in the first branch is smaller than the kernel of the depthwise separable convolution (DW) in the second branch. After the small-scale spatial features and large-scale spatial features are concatenated, they are channel-fused via pointwise convolution (PW) to obtain multi-scale spatial features.

[0010] Furthermore, the following formula is used to calculate small-scale spatial features. and large-scale spatial features , in, This indicates a batch normalization operation. Represents the ReLU activation function. This represents a pointwise convolution with a 1x1 kernel, where X represents the input feature. , These represent depthwise separable convolutions with kernels of 3*3 and 5*5, respectively. Calculate multi-scale spatial features using the following formula. , .

[0011] Furthermore, in step two, the channel attention module SEL first performs a global average pooling operation on the multi-scale spatial features to obtain the global statistics of each channel, then generates channel weights through a serial two-level pointwise convolution PW operation, and then performs a channel broadcast multiplication operation on the channel weights and the multi-scale spatial features to obtain the recalibrated channel weights.

[0012] Furthermore, the recalibrated channel weights are calculated using the following formula. , in, Represents the ReLU activation function. This represents the Sigmoid activation function. This indicates element-wise multiplication broadcast via a channel. This indicates the channel weights before recalibration.

[0013] Furthermore, the final output feature Y is calculated using the following formula. in, This represents the residual mapping function, when the channel weights When the number of channels is the same as that of the input feature X, Employ identity mapping; when channel weights When the number of channels is inconsistent with the number of channels of the input feature X, Pointwise convolution PW with a 1*1 kernel is used.

[0014] Furthermore, the remote sensing reflectance of each pixel in the fused image corresponding to the blue, green, red, and near-infrared bands is converted into a one-dimensional feature vector, and the longitude and latitude information of the geospatial coordinates corresponding to each pixel is converted into a one-dimensional feature vector, so as to construct a six-channel feature vector as the input of the water depth inversion network model.

[0015] Furthermore, in step one, when using the robust principal component analysis (RPCA) method for fusion processing, each single-temporal remote sensing image includes pixel information in four bands: blue, green, red, and near-infrared. First, the two-dimensional pixel information of the corresponding band of each single-temporal remote sensing image in the multi-temporal remote sensing image is converted into a one-dimensional pixel vector. Then, the one-dimensional pixel vectors of the same band are arranged in chronological order into a matrix to construct four two-dimensional observation matrices. Then, each two-dimensional observation matrix is ​​decomposed into a low-rank matrix and a sparse matrix. Finally, weighting coefficients are introduced to reconstruct the low-rank matrix and the sparse matrix by weighting, thereby obtaining the reconstructed band image and thus the fused image.

[0016] Furthermore, by solving the following convex optimization problem, the two-dimensional observation matrix is... Decompose into low-rank matrices With sparse matrices , in, The nuclear norm, or the sum of singular values, is used to constrain the low-rank property of a matrix. express Norms are used to constrain outlier components in sparse matrices. This represents the regularization parameter, used to balance the weights of low-rank terms and sparse terms; Using the following formula, for low-rank matrices With sparse matrices Weighted reconstruction is performed to obtain the reconstructed band image. Thus, fused images are obtained. in, This represents the weighting coefficient.

[0017] Furthermore, before performing the fusion process, each of the single-temporal remote sensing images is preprocessed by sequentially applying ACOLITE atmospheric correction, solar flare removal, and land-water separation.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) Enhance noise immunity in complex environments To address the issue that single-temporal remote sensing images are easily affected by random environmental factors such as ship activity, enhanced wave reflection, and local cloud shadows at the moment of imaging, resulting in striped or patchy anomalies in the inversion results, this invention introduces a multi-temporal image fusion method based on robust principal component analysis (RPCA). By utilizing the redundancy of multi-temporal observation data of the same area in the temporal dimension, stable background information and instantaneous anomalous signals are separated, thereby obtaining a spectrally stable fused image with a higher signal-to-noise ratio, providing more reliable input data for subsequent water depth inversion.

[0019] (2) Enhance the model's ability to express spatial heterogeneity To address the problem that existing deep learning-based water depth retrieval models primarily rely on spectral features and fail to explicitly utilize geospatial location information, thus struggling to characterize the spatial differences in optical properties and sediment types of nearshore waters, this invention proposes a spatial perception enhancement method based on normalized geographic coordinates. By adding normalized longitude and latitude channels to the model input, the network can learn the joint distribution relationship between spectral features and geographic location, thereby improving the model's adaptability and retrieval stability under different regional conditions.

[0020] (3) Improve the ability to comprehensively depict multi-scale terrain features. To address the problem that existing convolutional neural networks often employ single-scale convolutional kernels during feature extraction, making it difficult to simultaneously capture large-scale topographic trends and small-scale local structures, this invention constructs a deep learning network architecture, Geo-MSENet, that combines a multi-scale feature extraction module (MSB) and a lightweight channel attention module (SEL). This architecture acquires multi-scale spatial features through parallel convolutional operations with different receptive fields and adaptively enhances key spectral information through a channel attention mechanism, thereby improving the model's ability to represent complex nearshore topographic structures. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the overall structure of the water depth inversion network model of the present invention; Figures 3-5 This is a schematic diagram showing the accuracy of the Geo-MSENet model in the embodiments of the present invention on test sets in three study areas, namely, scatter density maps of clear water, moderately turbid water, and highly turbid water. Figures 6-8The diagrams shown are schematic representations of the RPCA fusion results of multi-temporal remote sensing images of three study areas in an embodiment of the present invention. They are RPCA fusion results of clear water, moderately turbid water, and highly turbid water. Figures 9-11 This is a schematic diagram of the water depth inversion results for three study areas in an embodiment of the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the water depth inversion method of this invention that integrates multi-temporal information and geospatial perception mechanism. It should be noted that the description of these embodiments is for the purpose of helping to understand this invention, but does not constitute a limitation of this invention.

[0023] like Figure 1 As shown, this invention provides a water depth inversion method that integrates multi-temporal information and geospatial perception mechanisms. First, a multi-temporal remote sensing image sequence at different time points within the same water area is obtained. Then, the RPCA algorithm is used to fuse the multi-temporal remote sensing image sequence to obtain a fused image. This fused image, combined with geospatial location, is then used as input to a water depth inversion network model to complete the water depth inversion estimation. This enables the model to learn the joint distribution relationship between spectral features and spatial location, thereby enhancing the model's ability to characterize the spatial heterogeneity of water body optical properties and improving generalization performance across different regions. Simultaneously, the model acquires multi-scale spatial features through parallel convolution operations of different receptive fields and adaptively strengthens key spectral information through a channel attention mechanism, thereby improving the model's ability to represent complex nearshore topographic structures.

[0024] Specifically as follows: S1. Remote Sensing Image Acquisition and Screening This embodiment uses L1C-level products from the Sentinel-2 MultiSpectral Instrument (MSI) satellite, part of the European Space Agency's (ESA) Copernicus program, as the basic remote sensing data source. L1C data consists of top of atmospheric reflectance (TOA) products that have undergone geometrical and orthorectification, preserving complete spectral information and making them suitable for subsequent atmospheric correction of water bodies and shallow water depth retrieval.

[0025] To ensure the temporal stability of the model training data and the representativeness of the water optical conditions, this embodiment conducted multi-temporal image screening for three study areas (clear water, moderately turbid water, and highly turbid water). For each study area, 25 Sentinel-2 images with cloud cover below 30% were selected, spanning from 2022 to 2024, to enhance the adaptability of the samples under different seasons, water suspended solids concentrations, and lighting conditions, thereby improving the model's generalization ability. In terms of band selection, the visible and near-infrared bands with a spatial resolution of 10 m were selected as the basic spectral input: B2 (Blue, center wavelength approximately 490 nm), B3 (Green, center wavelength approximately 560 nm), B4 (Red, center wavelength approximately 665 nm), and B8 (Near Infrared, center wavelength approximately 842 nm). The aforementioned bands possess strong penetrating power in shallow water bodies and are sensitive to differences in reflectance from suspended matter and sediment, making them a commonly used and stable source of fundamental spectral information in shallow water depth inversion. This embodiment constructs a basic spectral feature stack using these four bands, providing standardized input data for subsequent spectral index construction, multi-scale feature extraction, and machine learning / deep learning model training.

[0026] S2. Remote Sensing Image Standardization Preprocessing To reduce the impact of atmospheric influences, sea surface reflection interference, and land-mixed pixels on the accuracy of water depth inversion, this embodiment sequentially performs atmospheric correction, solar flare removal, and land-water separation processing on each acquired Sentinel-2 L1C remote sensing image to obtain stable and reliable water surface reflectance data. The Sentinel-2 L1C data is processed using ACOLITE (Atmospheric Correction for OLI 'Lite'), a software specifically developed by the Royal Institute of Natural Sciences in Belgium for water remote sensing. This software is based on the dark pixel assumption and a shortwave infrared aerosol inversion strategy, making it suitable for nearshore turbid water environments. To eliminate the influence of specular reflection caused by sea waves on visible light reflectance, this embodiment employs a linear regression flare correction method. This method utilizes the physical characteristics of strong absorption and near-zero reflectance in the near-infrared band in deep water regions, treating the NIR band as an indicator of flare intensity, thereby establishing a linear relationship between it and the visible light band. For each image, sample pixels are selected in the deep water area, and the linear regression slope between the NIR band and each visible light band is calculated. Based on this, the visible light band is corrected, and the calculation formula is as follows: in, To remove the flare Band reflectivity, Original reflectivity The regression slope, For near-infrared reflectivity, This represents the minimum value in the near-infrared band among deep-water pixels in the study area. This processing effectively suppresses the influence of bright flare regions on the water depth inversion model, improving the usability of the true optical signal of the water body.

[0027] To avoid interference from land pixels and non-water areas on model training and prediction results, this embodiment constructs a land-water separation mask to extract the effective water area. Specifically, the Normalized Differential Water Index (NDWI) is used for water body discrimination, and its calculation formula is as follows: in, For green light band reflectivity, This refers to the near-infrared reflectance. Water reflects relatively high light in the green band but low light in the near-infrared band; therefore, the NDWI value of water bodies is typically greater than zero. This embodiment sets a threshold. Water pixels are extracted, and morphological filtering methods are used to remove scattered noise areas, ultimately obtaining a continuous and complete effective water area mask for subsequent water depth inversion feature extraction and model training.

[0028] S3. Multi-temporal remote sensing image fusion based on RPCA To suppress transient noise interference caused by moving vessels, wakes, local splashes, and random environmental disturbances in single-temporal remote sensing images, this embodiment uses a multi-temporal image fusion method based on Robust Principal Component Analysis (RPCA) to perform temporal joint denoising and information enhancement processing on multiple images of the same area, thereby constructing a stable water reflectance image with a high signal-to-noise ratio.

[0029] For N preprocessed multispectral images (i.e., single-temporal remote sensing images) acquired from the same research water area, spatial registration and uniform cropping are first performed on each band to ensure that the spatial position of each pixel in the cropped N images is the same. Then, they are stacked according to the time dimension to obtain multi-temporal remote sensing images.

[0030] Let the spatial size of a single-temporal remote sensing image be... Then, the observations of each band in the time series can be represented as a three-dimensional data volume. To facilitate matrix factorization operations, the two-dimensional spatial dimension corresponding to the observations of each band is expanded into a one-dimensional pixel vector, and then arranged in chronological order into a matrix form to construct four two-dimensional observation matrices X: Each column represents a one-dimensional pixel vector converted from the two-dimensional pixel reflectance vector of the corresponding band of a single-temporal remote sensing image. They can be arranged into a one-dimensional pixel vector in row / column order. Each row represents the sequence of remote sensing reflectance changes at the same spatial location at different times.

[0031] For each two-dimensional observation matrix X, based on the RPCA theory, stable background information in time-series imagery (such as seabed topography and average optical properties of water bodies) has strong correlation in the time dimension and can be represented as a low-rank structure; while moving targets, wave flashes, and local anomalies have spatiotemporal sparsity and can be represented as sparse perturbations. Therefore, the two-dimensional observation matrix can be... Decompose into low-rank matrices With sparse matrices : in, This represents a low-rank matrix constructed from stable background information. This represents a sparse matrix constructed from instantaneous noise or anomalous disturbances. This decomposition is achieved by solving the following convex optimization problem: In the formula, The nuclear norm (i.e., the sum of singular values) is used to constrain the low-rank property of a matrix; for Norms are used to constrain outlier components in sparse matrices. This is a regularization parameter used to balance the weights of low-rank terms and sparse terms.

[0032] After completing the matrix decomposition, use the low-rank matrix As the primary source of information, and combined with physically meaningful weak variation information in the sparse matrix, weighting coefficients are introduced. The image is weighted and reconstructed to obtain the reconstructed band image. : in, For weighting coefficients. When When a smaller value is selected, transient disturbances such as moving vessels, wakes, and random wave reflections can be effectively suppressed; when When the image size is moderately increased, some true temporal variation information can be preserved. This involves combining the four band images. By combining pixel information from the same spatial location, a fused image with the same structure as a single-temporal remote sensing image can be formed.

[0033] This step significantly improves the signal-to-noise ratio and temporal consistency of the images, generating high-quality synthetic water reflectance images and providing stable input data for subsequent water depth inversion feature construction and model training.

[0034] S4. Construct the Geo-MSENet deep learning water depth inversion model This embodiment constructs a water depth inversion network model, Geo-MSENet, which integrates geospatial perception capabilities and multi-scale feature extraction capabilities, for water depth inversion modeling in shallow water areas. Figure 2 As shown, the network model uses a multi-scale convolutional module (MSB) + a lightweight channel attention module (SEL) + residual connections as core units. Feature extraction is completed through the serial stacking of multiple core units, such as three levels, and finally the regression head outputs continuous water depth values.

[0035] (1) Input feature construction and tensor organization The input to this water depth inversion network model is a six-channel feature tensor. ,Include: Four spectral channels: the remote sensing reflectance Rrs of each pixel in the fused image in the blue, green, red, and near-infrared bands; Two geospatial channels: Longitude (Lon) and Latitude (Lat) corresponding to each pixel in the fused image.

[0036] To avoid training instability caused by differences in geographic coordinate units, both longitude (Lon) and latitude (Lat) need to be normalized. The normalization method is as follows: , Will[ , , , The six-channel feature tensor X is obtained by concatenating Lon' and Lat' in the channel dimension and used as the model input. By explicitly introducing geospatial coordinate information, the network can learn the joint distribution of spectral response and spatial location, thereby enhancing its ability to express the spatial heterogeneity of water body optical properties.

[0037] (2) Multi-Scale Block (MSB) Let the input of the MSB module be... The output is This module employs a parallel convolutional structure and simultaneously sets... and Two convolutional branches, namely the first branch and the second branch, are used to extract spatial structure information focusing on local fine textures and spatial structure information under a larger receptive field, respectively, thereby enhancing the model's ability to express complex underwater terrain changes. To reduce the number of parameters and computational cost, this embodiment uses depthwise convolution (DW) + pointwise convolution n (PW) in each branch. The kernel of the depthwise convolution DW in the first branch is 3*3, and the kernel of the depthwise convolution DW in the second branch is 5*5.

[0038] Depthwise Convolution (DW): Spatial convolution is performed independently on each input channel, extracting only neighborhood structural information. Pointwise Convolution (PW): This linearly combines all input channels at each pixel location, achieving channel blending and dimensionality transformation. PW weights all input channels at each pixel location to generate new channels; assigning n weights generates n channels. The single-branch operation form is: Calculate small-scale spatial features using the following formula. and large-scale spatial features : in, This indicates a batch normalization operation. Represents the ReLU activation function. This indicates a pointwise convolution with a 1*1 kernel, and X represents the input feature.

[0039] To maintain a consistent number of output channels, the output channels of the first and second branches are divided proportionally: Let First branch output Channel, second branch output The channels are then spliced ​​together in the channel dimensions: Then, cross-channel fusion is performed using 1×1 pointwise convolution PW, along with batch normalization (BN) and ReLU activation function to obtain multi-scale spatial features. : This step enables the recombination and information exchange of features at different scales, avoiding channel redundancy caused by simple splicing.

[0040] (3) Structure and operation of the lightweight channel attention module (Squeeze-Excitation Lite, SEL) To adaptively emphasize feature channels that are more sensitive to water depth inversion, this embodiment connects a SEL module after the MSB module. It obtains channel statistical features through global average pooling and combines them with... Convolution generates channel weights, enabling adaptive recalibration of the importance of different spectral and spatial feature channels, thereby highlighting key features that contribute more to water depth retrieval. The computation process includes: First, global statistics for each channel are obtained through global average pooling (GAP): Then, channel weights are generated using two layers of 1×1 pointwise convolutions (PW). in, It is the ReLU activation function. Using the Sigmoid activation function, the compression ratio is... That is, the number of intermediate channels is .

[0041] Then, channel weights are recalibrated. (4) Residual connection and channel alignment To mitigate the degradation of deep training and maintain feature transfer stability, the MSB output employs a residual structure: in, For residual branches: when Channel alignment is achieved using 1×1 pointwise convolution PW; when The time is an identity mapping. This design can superimpose multi-scale enhanced features while preserving the original information, thereby improving expressive power and convergence stability.

[0042] (5) Main stacking structure The Geo-MSENet model consists of three core units M connected in series: The subscript indicates the number of output channels.

[0043] The first core unit M has 6 input channels. The PW in the multi-scale convolution branch linearly combines the input channels. The first branch outputs 8 channels, and the second branch outputs 8 channels. The two branches are concatenated in the channel dimension to form 16 channels. The residual branch performs channel mapping on the original 6-channel input through 1×1 convolution to generate 16 channels, which are used for element-wise addition with the main branch output.

[0044] The second core unit M: The input has 16 channels. The PW of the two multi-scale branches maps the input features to 16 channels respectively. After concatenation, 32 channels are obtained. The residual branch also maps the input 16-channel features to 32 channels through 1×1 convolution. After the channel dimension alignment is completed, the residuals are added.

[0045] The third core unit M has 32 input and 32 output channels. The two branches PW each generate 16 channels, which are then concatenated to obtain 32 channels. Since the number of input and output channels is the same, the residual branch uses an identity mapping, directly passing the input features to the output and adding them to the main branch result.

[0046] By progressively expanding the feature dimensions and integrating spatial information at different scales, the network can simultaneously represent shallow water topographic details and larger-scale landform changes.

[0047] (6) Return head and water depth output This embodiment uses a regression head to map high-dimensional features into water depth estimates. The output features of the last level... First, adaptive global average pooling is performed to obtain the global representation vector: Then, regression is performed using two fully connected layers to output the predicted water depth. .

[0048] S5. Model Training and Hyperparameter Setting To ensure that model training, parameter tuning, and accuracy evaluation are independent of each other, this embodiment randomly divides the measured water depth samples into training set, validation set, and test set in a ratio of 8:1:1.

[0049] (1) Loss function design The model training uses a custom adaptive combined loss function, which is a weighted sum of Huber loss and mean squared error (MSE): Among them, Huber loss is approximately squared when the error is small and turns into linear growth when the error is large, which can reduce the impact of outliers on model training; MSE loss maintains high sensitivity to the overall error. The combination of the two can improve the robustness of the model while ensuring the fitting accuracy.

[0050] (2) Optimization strategy and training parameters The AdamW optimizer was chosen as the optimization algorithm to achieve adaptive learning rate updates and to suppress overfitting risk through weight decay. The main training hyperparameter settings are as follows: Batch Size: 32; Initial Learning Rate: Training epochs: 100 During training, the model is monitored based on the changes in validation set error, and the model parameters with the best validation performance are selected for subsequent testing and graphing applications.

[0051] S6. Accuracy Evaluation The model inversion results are quantitatively evaluated using measured water depth data from an independent test set. The selected evaluation metrics include the coefficient of determination. The root mean square error (RMSE) and mean absolute error (MAE) are calculated using the following formulas: in, This is the measured water depth value. For model inversion of water depth values, This is the average of the measured water depths. This represents the number of samples.

[0052] S7. Water Depth Mapping After the model training is completed and its accuracy is verified, the trained Geo-MSENet model is applied to the entire image of the research sea area to perform pixel-by-pixel water depth inversion for each water cell, ultimately generating a high spatial resolution water depth distribution product. This product can be used in nearshore topographic mapping, ecological environment monitoring, and marine engineering applications.

[0053] To verify the feasibility of the proposed water depth inversion method integrating multi-temporal information and geospatial sensing mechanisms, we conducted validation studies in three representative water bodies with different optical properties along the Chinese coast (clear water, moderately turbid water, and highly turbid water). Experimental results show that this method significantly improves upon existing technologies in terms of inversion accuracy, noise resistance, and spatial adaptability. Specific technical effects are as follows: (1) Improve the accuracy of water depth inversion in complex nearshore waters In test regions with varying optical complexities, the method of this invention achieved stable inversion accuracy. Clear water (Zone A): The root mean square error (RMSE) of the inversion results is approximately 0.60 m, and the mean absolute error (MAE) is 0.31 m; Moderately turbid water (Zone B): In areas affected by seasonal changes in the optical properties of water bodies, the RMSE is approximately 0.81 m, demonstrating good environmental adaptability; Highly turbid waters (Zone C): In estuarine waters with high suspended sediment content, the RMSE is approximately 1.23m.

[0054] Compared with traditional empirical models, random forests (RF), XGBoost, CatBoost, and U-Net, etc. Figure 3 - 5) The method of the present invention shows a lower level of RMSE in all regions, with the error reduced by 22%–81%, 18%–80% and 42%–70% respectively, demonstrating a more stable adaptability to complex optical environments.

[0055] (2) Improve the ability to suppress transient environmental noise. To address the issue that single-temporal images are susceptible to transient interference such as ship wakes, enhanced wave reflections, and local cloud shadows, this invention effectively reduces the impact of such anomalous signals on the inversion results through RPCA multi-temporal fusion processing (see...). Figure 6 - 8). The fused image exhibits a more continuous background and more stable underwater reflection information in space. In the generated depth distribution map, strip-like or patchy anomalies caused by transient environmental noise are significantly reduced, and the underwater topographic structure shows a more continuous spatial variation trend, see... Figures 9-11 These figures represent the water depth inversion results for clear water and moderately turbid water, respectively. Among them, (a) is the water depth inversion result based on RPCA fusion of multi-temporal images; (b) is the water depth inversion result based on single-temporal remote sensing images; (c) is an enlarged view of the area within the red box in (a); (d) is an enlarged view of the area within the red box in (b); (e) is an enlarged view of the area within the yellow box in (a); and (f) is an enlarged view of the area within the yellow box in (b).

[0056] (3) Enhance the model's adaptability to spatially heterogeneous environments. By introducing normalized geographic coordinate channels and a multi-scale feature extraction structure into the Geo-MSENet model, the model can better characterize the influence of water optical properties and sediment differences on spectral response in different regions. Ablation experiments show that in complex water scenarios, the model error is further reduced (RMSE decreases by 37-50%) after introducing geospatial information. Simultaneously, it maintains relatively stable inversion results for large-scale gentle seabeds and locally complex terrains, indicating that this structure helps improve the model's spatial generalization ability.

[0057] (4) Possesses good application and promotion potential This invention utilizes publicly available Sentinel-2 multispectral satellite data for water depth inversion, eliminating the need for large-scale field measurement equipment and enabling the acquisition of shallow water depth information over a wide area at a relatively low cost. Furthermore, for shoals and complex nearshore areas that are difficult to cover with traditional shipborne sonar measurements, this method can provide supplementary spatial distribution information, demonstrating practical value in nearshore topographic surveys, waterway planning, and coastal environmental monitoring.

[0058] It is important to note that the schemes and arrangements of this application shown in the exemplary embodiments are merely exemplary. Although only a few embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible (e.g., variations in various parameter values ​​(temperature, power, humidity, etc.), installation arrangements, names, colors, logical orders, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in this application. Therefore, all such modifications are also included within the scope of the invention, and the order or sequence of any process or method steps may be changed or rearranged according to alternative embodiments. In the claims, any "apparatus plus function" clause is intended to cover the structure described herein for performing the function, and not only structurally equivalent but also equivalent in structure. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of the invention. Therefore, the invention is not limited to the particular embodiments but extends to a variety of modifications that still fall within the scope of the appended claims.

[0059] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0060] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for water depth inversion that integrates multi-temporal information and geospatial sensing mechanisms, characterized in that... Includes the following steps: Step 1: Data Processing Single-temporal remote sensing images of the same water area at different time points are acquired, and these images are stacked according to time series to construct a multi-temporal remote sensing image sequence. Robust principal component analysis (RPCA) is then used to fuse the multi-temporal remote sensing image sequence to obtain a fused image. Step 2: Construct a water depth inversion network model Multiple core units are cascaded together to complete stepwise feature extraction, and the last core unit is connected to the regression head to output continuous water depth values. Each of the core units includes a multi-scale convolutional module (MSB) and a channel attention module (SEL) connected in series. The output features are residually connected with the original input features to obtain the final output features. The multi-scale convolutional module adopts a parallel convolutional structure to extract small-scale spatial features and large-scale spatial features respectively, and then fuses the small-scale spatial features and large-scale spatial features to obtain multi-scale spatial features. The channel attention module (SEL) is used to generate channel weights and then perform dot product fusion with multi-scale spatial features to obtain recalibrated channel weights. Step 3: Train the water depth inversion network model using fused imagery and geospatial location, and then use the trained water depth inversion network model to perform water depth inversion on the processed remote sensing imagery to be inspected.

2. The water depth inversion method according to claim 1, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: In step two, the multi-scale convolution module (MSB) includes a parallel first branch and a second branch, which are used to extract small-scale spatial features and large-scale spatial features, respectively. Both branches include depthwise separable convolution (DW) and pointwise convolution (PW). The kernel of the depthwise separable convolution (DW) in the first branch is smaller than the kernel of the depthwise separable convolution (DW) in the second branch. After the small-scale spatial features and large-scale spatial features are concatenated, channel fusion is performed through pointwise convolution (PW) to obtain multi-scale spatial features.

3. The water depth inversion method according to claim 2, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: Calculate small-scale spatial features using the following formula. and large-scale spatial features , in, This indicates a batch normalization operation. Represents the ReLU activation function. This represents a pointwise convolution with a 1x1 kernel, where X represents the input feature. , These represent depthwise separable convolutions with kernels of 3*3 and 5*5, respectively. Calculate multi-scale spatial features using the following formula. , 。 4. The water depth inversion method according to claim 2, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: In step two, the channel attention module SEL first performs global average pooling on the multi-scale spatial features to obtain the global statistics of each channel, then generates channel weights through a two-stage pointwise convolution PW operation, and then performs channel broadcast multiplication operation on the channel weights and multi-scale spatial features to obtain the recalibrated channel weights.

5. The water depth inversion method according to claim 4, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: Calculate the recalibrated channel weights using the following formula. , in, Represents the ReLU activation function. This represents the Sigmoid activation function. This indicates element-wise multiplication broadcast via a channel. This indicates the channel weights before recalibration.

6. The water depth inversion method according to claim 5, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: The final output feature Y is calculated using the following formula. in, This represents the residual mapping function, when the channel weights When the number of channels is the same as that of the input feature X, Employ identity mapping; when channel weights When the number of channels is inconsistent with the number of channels of the input feature X, Pointwise convolution PW with a 1*1 kernel is used.

7. The water depth inversion method according to claim 1, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: The remote sensing reflectance of each pixel in the fused image in the blue, green, red and near-infrared bands is converted into a one-dimensional feature vector. At the same time, the longitude and latitude information of the geospatial coordinates corresponding to each pixel are converted into a one-dimensional feature vector to construct a six-channel feature vector as the input of the water depth inversion network model.

8. The water depth inversion method according to claim 1, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: In step one, when using the robust principal component analysis (RPCA) method for fusion processing, each single-temporal remote sensing image includes pixel information in four bands: blue, green, red, and near-infrared. First, the two-dimensional pixel information of the corresponding band of each single-temporal remote sensing image in the multi-temporal remote sensing image is converted into a one-dimensional pixel vector. Then, the one-dimensional pixel vectors of the same band are arranged in chronological order into a matrix to construct four two-dimensional observation matrices. Each two-dimensional observation matrix is ​​then decomposed into a low-rank matrix and a sparse matrix. Finally, weighting coefficients are introduced to reconstruct the low-rank matrix and the sparse matrix by weighting, thereby obtaining the reconstructed band image and thus the fused image.

9. The water depth inversion method according to claim 8, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: By solving the following convex optimization problem, the two-dimensional observation matrix is... Decompose into low-rank matrices With sparse matrices , in, The nuclear norm, or the sum of singular values, is used to constrain the low-rank property of a matrix. express Norms are used to constrain outlier components in sparse matrices. This represents the regularization parameter, used to balance the weights of low-rank terms and sparse terms; Using the following formula, for low-rank matrices With sparse matrices Weighted reconstruction is performed to obtain the reconstructed band image. Thus, fused images are obtained. in, This represents the weighting coefficient.

10. The water depth inversion method according to claim 8, which integrates multi-temporal information and geospatial sensing mechanism, is characterized in that: Before the fusion process, each of the single-temporal remote sensing images is preprocessed by sequentially applying ACOLITE atmospheric correction, solar flare removal, and land-water separation.

Images