State perception and uncertainty driven satellite sounding element learning method

Abstract

The invention discloses a morphological perception and uncertainty-driven satellite sounding element learning method, which comprises the following steps of: constructing a heterogeneous basic model layer, optimizing sample training by adopting a zero-leakage training strategy and combining a spatial distribution adaptive weight mechanism on the basis of the heterogeneous basic model layer, and generating a basic prediction result and uncertainty agent information; determining an optimal space window and calculating an explicit terrain descriptor through multi-scale orthogonal partition topological optimization; fusing the predicted value of the base model, the displayed topographic topological features and the uncertainty features into element feature vectors; and inputting a meta-learner of the residual network containing the attention mechanism, and training by using a weighted loss function to obtain a final sounding result. According to the method, the problems of poor cross-domain robustness, long-tail distribution deviation and the like are solved, the depth sounding precision and space consistency of heterogeneous water areas are improved, and the satellite depth sounding performance is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing water depth inversion technology, and in particular relates to a satellite-derived bathymetry (SDB) meta-learning framework that integrates explicit geomorphic features with adaptive uncertainty weights. Background Technology

[0002] High-precision shallow water bathymetry (SDB) data is a core foundational data source supporting coastal engineering construction, navigation safety, and benthic ecosystem protection. Its inversion accuracy and cross-domain adaptability directly impact the effectiveness of global coastal resource development and ecological governance. Compared to traditional active bathymetry techniques (such as multibeam echo sounders), which offer high accuracy but suffer from high cost and low spatial coverage efficiency, multispectral satellites (such as Sentinel-2 and the Landsat series) have become the mainstream implementation platform for SDB technology due to their advantages of wide coverage, high revisit frequency, and low cost. Related methods are widely used for bathymetry inversion in heterogeneous waters such as turbid ports, clear coral reefs, and complex estuaries.

[0003] The closest existing technology upon which this invention is based mainly focuses on two major directions: "SDB method that integrates traditional machine learning and semi-empirical models" and "deep learning-driven SDB improvement attempts." These two types of technologies are the direct technical basis for the "morphological perception and uncertainty-driven meta-learning framework (MAAS-Net)" proposed in this application. Their specific development status and limitations are as follows: To address the limitations of traditional physical analysis models (such as the Lyzenga algorithm) which offer strong physical interpretability but struggle to handle substrate heterogeneity, and the weak generalization ability of early empirical models, researchers proposed the "Data-Physics Fusion SDB Technology"—which combines the physical logic of semi-empirical models with the nonlinear fitting capabilities of machine learning to improve inversion accuracy in complex environments. Among them, the multi-band physical analysis model proposed by Lyzenga et al. establishes the correlation between spectrum and water depth based on the radiative transfer equation, providing a basic physical framework for SDB (Lyzenga et al., 2006); the semi-empirical ratio algorithm of Stumpf et al. further simplifies the radiative transfer process and achieves efficient water depth inversion in clear water areas (Stumpf et al., 2003); with the development of machine learning, Belgiu and Dragot applied the random forest model to SDB, improving the adaptability of heterogeneous bottom areas by capturing the nonlinear relationship between spectrum and water depth (Belgiu and Dragot, 2016); the XGBoost model proposed by Chen and Guestrin minimizes prediction bias through gradient boosting strategy, achieving an inversion effect of R²≈0.7-0.8 in moderately complex waters (Chen and Guestrin, 2016); Surrisetty et al. further adopted the ensemble learning approach, integrating the outputs of multiple machine learning models to improve the robustness of the results (Surisetty et al., 2021).

[0004] However, such technologies have three major limitations: First, they rely on pixel-level spectral-depth correlation, failing to consider the continuity of seabed topography, which leads to spatial fragmentation of the inversion results and makes it impossible to reconstruct key geomorphic structures such as channel edges and coral reef slopes; second, they have poor adaptability to long-tailed data, with scarce extreme shallow and deep water samples in the depth dataset, and the standard training strategy overfits the median depth, resulting in significant errors in edge regions (Yang et al., 2021); third, they lack uncertainty quantification mechanisms, making it impossible to assess the reliability of the prediction results and limiting their engineering applications under the IHO hydrographic measurement standard (IHO, 2020).

[0005] To address the spatial information deficiency problem in traditional machine learning, researchers have developed "deep learning-driven SDB technology." This technology optimizes the inversion effect through patch-level feature extraction and global dependency modeling, forming the technical foundation of the morphology perception module in this invention. The U-Net architecture proposed by Siddique et al. utilizes convolutional neural networks (CNNs) to extract local spatial features, achieving spatial smoothing for water depth prediction (Siddique et al., 2021). The SegNet model by Badrinalayanan et al. further optimizes the encoder-decoder structure, improving the ability to restore details in shallow water areas (Badrinarayanan et al., 2017). Khan et al. introduced the visual Transformer into SDB, modeling long-distance dependencies of spectral sequences through a self-attention mechanism, overcoming the limitations of CNNs' local receptive field (Khan et al., 2022). The BathyFormer architecture proposed by Lv et al. achieves state-of-the-art performance, reducing the RMSE to within 1.5m in clear water areas (Lv et al., 2025).

[0006] However, such deep learning methods still have core limitations: First, they are based on "implicit spectral mapping," treating spatial patterns only as textures rather than landform structures, lacking explicit terrain constraints. In high turbidity (Caballero and Stumpf, 2023) or deep water low signal-to-noise ratio regions (Chen et al., 2024), they are prone to overfitting spectral noise, resulting in deep water extrapolation errors of up to -2.6m (AlNajar et al., 2023). Second, they do not address the long-tail distribution bias, and the standard loss function still favors median water depth, resulting in insufficient prediction accuracy in extreme depth regions. Third, the models are highly complex and have poor interpretability, lacking a collaborative fusion mechanism of spectrum, landform, and uncertainty, resulting in weak cross-domain generalization ability and significant performance degradation in unseen optical environments (Meyer and Pebesma, 2021).

[0007] In summary, current technologies that are closest to the mark recognize the importance of spatial information and data fitting capabilities for SDB (Self-Depth Boundary Depth), but a complete remote sensing depth inversion technology system encompassing "explicit geomorphic feature extraction - long-tailed distribution adaptive weighting - multi-source information synergistic fusion" has not yet been formed. These technologies either focus solely on nonlinear fitting of spectra and depth, or perform only simple spatial smoothing, failing to simultaneously address the three core issues of poor cross-domain robustness, spatial fragmentation, and extreme depth errors. This results in the inability to measure SDB inversion accuracy (R² is often below 0.8 in complex environments) and engineering applicability standards in heterogeneous waters without on-site calibration data. This current technological status quo provides a clear direction and technical foundation for the MAAS-Net framework proposed in this application. Summary of the Invention

[0008] To address the technical problems of poor cross-domain robustness, long-tail distribution bias, spatial fragmentation, and lack of engineering interpretability in existing satellite bathymetry technologies, this invention proposes a morphology-aware and uncertainty-driven satellite bathymetry meta-learning method. Through explicit topographic feature constraints and an adaptive weighting mechanism, it achieves high-precision and high-consistency water depth inversion in heterogeneous waters, overcoming the performance bottleneck of existing technologies in complex environments. To achieve the above objectives, this invention provides a morphology-aware and uncertainty-driven satellite bathymetry meta-learning method, comprising: constructing a heterogeneous base model layer, and on this basis, employing a zero-leakage training strategy to generate base prediction results and spatially distributed adaptive weights (SDAW); performing multi-scale orthogonal partitioning topology optimization to determine the optimal spatial window and extract explicit terrain descriptors; fusing the base model predictions, explicit terrain descriptors, and uncertainty features into a meta-feature vector; inputting the meta-feature vector into a residual network meta-learner with an attention mechanism, and training it through a weighted loss function to obtain the final bathymetry result; and employing a two-stage verification protocol and engineering accuracy indicators to complete model performance evaluation and verification.

[0009] Optionally, the process of constructing a heterogeneous base model layer and adopting a zero-leakage training strategy on this basis includes: selecting four complementary regression models to form a heterogeneous ensemble layer: a log-linear model based on Beer-Lambert's law, a random forest (RF) model, an XGBoost model, and a support vector regression (SVR, Nyström approximation) model; calculating spatially distributed adaptive weights (SDAW): obtaining the scarcity of the sample space through kernel density estimation (KDE), using the prediction standard deviation of the four base models as an uncertainty proxy, and combining the log-damped function and percentile pruning to generate sample weights; using 5-fold cross-validation to generate an independent out-of-fold (OOF) map for meta-learner training, and using the ensemble mean map trained on the full data during inference to avoid data leakage.

[0010] Optionally, the multi-scale orthogonal partitioning topology optimization process includes: automated scale search: dividing the spatial window into microscale (3-5 pixels), mesoscale (7-15 pixels), and macroscale (17-35 pixels), and simultaneously optimizing the scale triples and meta-learner hyperparameters using the Tree-structured Parzen Estimator (TPE) algorithm; introducing orthogonality penalty constraints: adding a Pearson correlation coefficient penalty term (λ=0.25) to the objective function to ensure statistical orthogonality of features at different scales; calculating explicit terrain descriptors: including six types of topological features such as terrain roughness, terrain location index (BPI), texture anisotropy, and terrain gradient.

[0011] Optionally, the meta-feature vector is constructed by sequentially connecting three parts: four heterogeneous base model predictions (X_base), explicit terrain descriptor (X_topo), and uncertainty features (X_unc, including model prediction standard deviation and SDAW surrogate value).

[0012] Optionally, the meta-learner structure of the residual network with attention mechanism includes: employing N residual blocks, each block integrating a Squeeze-and-Excitation (SE) attention mechanism, learning channel weights through global average pooling and fully connected layers; the loss function adopts SDAW-weighted MSE loss, with the following formula: ; in For spatially distributed adaptive weights, For true depth, To predict depth, the Adam optimizer and early stopping strategy are used, and the OOF validation loss is monitored to prevent overfitting.

[0013] Optional, the two-stage validation protocol and engineering accuracy metrics include: Two-stage validation: The source domain adopts hierarchical 5-fold cross-validation, and cross-domain independent validation is performed on heterogeneous target datasets (turbid ports, clear water coral reefs, complex confluence areas); Accuracy metrics: encompassing coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), bias, logarithmic RMSE, 90% linearity error (LE90), and IHO pass rate, comprehensively assessing the model's compliance with engineering application standards.

[0014] Optionally, the spatial distribution adaptive weights (SDAW) are calculated as follows: ; in This is the spatial point density (calculated using KDE). The standard deviations for four heterogeneous baseline models are predicted. Optionally, the optimal spatial window is determined using the Optuna optimization framework, with typical scale triples being (5 pixels, 13 pixels, 35 pixels) and feature correlation absolute values ​​≤ 0.4. Technical effects of this invention: This invention discloses a morphological perception and uncertainty-driven satellite depth sounding learning method. Through the synergistic fusion of heterogeneous integration, explicit topographic constraints, and adaptive weights, it effectively solves the problems of cross-domain adaptation, long-tail bias, and spatial fragmentation. In heterogeneous waters, the source domain RMSE is reduced to 0.792m (a 7.4% improvement over the optimal baseline model), the cross-domain clear water coral reef RMSE is reduced by 33.1%, and the turbid confluence area RMSE is reduced by 24.0%; the deep water underestimation bias is reduced from -2.6m to -1.3m, and the IHO 1a standard pass rate reaches 67.7%; the generated depth map has continuous topography and high geomorphological fidelity, which is closer to the high-standard operational application requirements of coastal engineering and navigation safety, and the technical effect is significantly better than existing deep learning and traditional machine learning models. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a morphological perception and uncertainty-driven satellite depth sounding learning method according to an embodiment of the present invention. Figure 2 This is a visualization diagram of the long-tail distribution and SDAW weights in an embodiment of the present invention; Figure 3 This is a diagram of the multi-scale topology optimization trajectory and feature orthogonality matrix in an embodiment of the present invention. Figure 4 This is a scatter plot comparing measured and predicted water depths in an embodiment of the present invention. Figure 5 This is a diagram showing the spatial residual distribution and depth stratification deviation in an embodiment of the present invention. Figure 6 This is a comparison chart of the spatial fidelity of water depth maps in embodiments of the present invention; Figure 7 This is a cross-domain generalization evaluation method for an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] like Figure 1 As shown, this embodiment provides a satellite depth sounder learning method driven by morphological awareness and uncertainty, including: Data acquisition and preprocessing, outputting standardized image data and aligned measured water depth data; A heterogeneous base model layer is constructed, and a zero-leakage training strategy is adopted on this basis to generate basic prediction results and spatially distributed adaptive weights (SDAW). Perform multi-scale orthogonal partitioning topology optimization to extract explicit terrain descriptors; Multi-source features are fused to construct a meta-feature vector, which is then input into a meta-learner for training to obtain the final depth measurement result. A two-stage verification protocol and engineering accuracy indicators were used to complete the model performance evaluation.

[0019] Furthermore, the data acquisition and preprocessing process includes: Collect Sentinel-2 multispectral satellite imagery, measured water depth data, and tidal data for the target area; Radiometric calibration, atmospheric correction, and flare removal preprocessing were performed on satellite imagery using Acolite software. Geometric registration of images and measured water depth data is achieved, and tidal level correction is performed using tidal data to eliminate the influence of tides; Remove outliers and output standardized multi-band spectral data and measured water depth dataset with precise geographic location.

[0020] Furthermore, the process of constructing a heterogeneous base model layer and then employing a zero-leakage training strategy on top of it includes: Four complementary regression models were selected to form a heterogeneous ensemble layer: log-linear model, random forest (RF) model, XGBoost model, and support vector regression (SVR, Nyström approximation) model. The formula for calculating Spatial Distribution Adaptive Weights (SDAW) is as follows: middle Kernel density estimation (KDE) calculation, The standard deviation of the predictions from the heterogeneous basic model; Five-fold cross-validation is used to generate independent out-of-fold (OOF) maps, and full data training is used to generate an ensemble mean map to avoid data leakage; By combining measured water depth data with spectral features, hyperparameters of each base model were optimized, and the optimization range is shown in Table 1.

[0021] Furthermore, the process of multi-scale orthogonal partitioning topology optimization includes: Automated scale search: Divide the window into three levels: micro (3-5 pixels), meso (7-15 pixels), and macro (17-35 pixels), and optimize the scale triplet using the TPE algorithm; Introducing orthogonality penalty constraints, the objective function is: Six types of explicit terrain descriptors are calculated, including terrain roughness, terrain location index (BPI), texture anisotropy, and terrain gradient.

[0022] Furthermore, the construction of meta-feature vectors and the training process of the meta-learner include: The meta-feature vector consists of three parts: middle Base model predictions Shape descriptor, Standard deviation and SDAW proxy value; The meta-learner employs a residual network with SE attention mechanism, and the loss function is SDAW-weighted MSE: . The Adam optimizer and early stopping strategy monitor OOF to verify loss and prevent overfitting.

[0023] Furthermore, the two-stage verification protocol and engineering accuracy metrics include: Within the domain, a stratified 5-fold cross-validation was used, while cross-domain independent validation was conducted in Nanshan Port, Qilianyu, and the Yellow River confluence area. Accuracy metrics include R², RMSE, MAE, Bias, logarithmic RMSE, LE90, and IHO pass rate, providing a comprehensive evaluation of model performance.

[0024] Specifically, the implementation process of this embodiment includes: Step 1: Data Acquisition and Preprocessing. Acquire Sentinel-2 multispectral satellite imagery, measured depth control point data, and tidal observation data for the target area. The image time series should be consistent with or as close as possible to the measured time; alternatively, other multispectral satellite imagery such as the Landsat series can be used. Image preprocessing includes: first, radiometric calibration to convert digital quantization values ​​into apparent radiance; then, atmospheric correction to remove atmospheric scattering and absorption interference; followed by geometric correction to ensure accurate spatial positioning; and finally, flare removal to eliminate the influence of direct solar reflection from the water surface. All preprocessing operations are performed in Acolite software. After preprocessing, the image is registered with electronic nautical charts, valid depth point data is extracted, and tidal level correction is performed using tidal data to obtain standardized image data and aligned measured depth datasets.

[0025] Step 2: Heterogeneous Matrix Model Training and Zero-Leakage Strategy Implementation. This step consists of three processes: First, based on the preprocessed spectral features and measured water depth data, four heterogeneous matrix models are trained respectively; second, hyperparameter optimization and accuracy evaluation are performed; finally, OOF prediction maps and integrated mean maps are generated.

[0026] Hyperparameter optimization was achieved using the Optuna framework, employing 5-fold cross-validation. The average of the 5-fold validation accuracy was used as the performance index of the base model. The range of hyperparameter optimization is shown in Table 1.

[0027] Table 1. Optimization range of hyperparameters for the base model

[0028] The core principles and implementation methods of the four base models are as follows: Log-linear model: Based on the Beer-Lambert law, a linear relationship is established between the logarithm of spectral reflectance and water depth, providing physical constraints. The formula is as follows: ; in Because of the water depth, For spectral reflectance, , These are fitting parameters, suitable for deep-water extrapolation scenarios.

[0029] Random Forest (RF): This method constructs multiple decision trees using bootstrap sampling. Each tree splits by randomly selecting a subset of features, enhancing the independence between trees. In regression tasks, minimizing the mean squared error is used as the splitting criterion. The final prediction result is the mean of the predictions from all decision trees, expressed by the formula: ; in For the number of decision trees, This is the prediction function for the k-th tree, which can effectively suppress noise interference.

[0030] XGBoost: A strong learner built iteratively based on a gradient boosting strategy. The objective function for the t-th iteration is expanded using a second-order Taylor series as follows: ; in The first derivative of the loss function. It is the second derivative. This is a regularization term.

[0031] The final prediction result is the sum of the predictions from all prediction trees: ; It can gradually reduce prediction bias and adapt to water depth heterogeneity.

[0032] SVR (Nyström approximation): Maps low-dimensional spectral features to a high-dimensional space using the RBF kernel function, and finds the optimal hyperplane. The optimization objective is: ; The constraints are , in Here, C is the feature mapping function, and C is the regularization parameter. This is the error pipeline threshold. The final prediction formula is: ; It can capture fine spectral responses.

[0033] Simultaneously, spatially distributed adaptive weights (SDAW) are calculated: the point density of the sample space is calculated through kernel density estimation (KDE). The standard deviation of the prediction results using the four-base model As an agent of uncertainty, sample weights are obtained through logarithmic damping and percentile clipping (clipping range [0.05, =0.95]). This forces the model to focus on "difficult samples" in the long-tailed distribution.

[0034] Step 3: Multi-scale orthogonal partitioning topology optimization. First, an automated scale search is performed. Based on the Optuna framework, the TPE algorithm is called to simultaneously optimize the hyperparameters of the micro, meso, and macro scale triples (K1, K2, K3) and the meta-learner. Typical optimal scales are (5 pixels, 13 pixels, 35 pixels). An orthogonality penalty constraint is introduced by adding a Pearson correlation coefficient penalty term to the objective function to ensure that the features extracted at different scales are statistically orthogonal (absolute correlation value ≤ 0.4), maximizing feature information gain.

[0035] The core formula for calculating six types of explicit terrain descriptors based on optimal scaling is as follows: Terrain roughness: ; in The depth measurement value of the pixels within the window. The number of pixels in the window distinguishes between coral reefs and sandy substrates; Topographic Position Index (BPI): ; in For the target pixel depth, The average depth of the macroscopic window is used to identify depressions and ridges; Texture anisotropy: ; in , The magnitude of the gradient smoothing in the horizontal and vertical directions. =1e-6, quantifying the directionality of sand ripples; Terrain gradient: ; Depicting the slope and fold characteristics of the landform; Topographic curvature: Calculated by fitting a quadratic surface, reflecting the degree of curvature of the terrain; Slope: Calculated based on the maximum slope method, quantifying the terrain tilt angle.

[0036] Step 4: Meta-learning Fusion and Training. Construct a meta-feature vector of "base model prediction - terrain features - uncertainty", and then... The sequential connection results in a total dimension of 4 (base model predictions) + 6 (terrain descriptors) + 2 (uncertainty features) = 12 dimensions.

[0037] The meta-learner employs a residual network with SE attention mechanism. Each residual block contains two stages: Squeeze and Excitation. The Squeeze stage aggregates spatial information through global average pooling to obtain channel statistical features. The excitation phase learns channel weights through a two-layer fully connected network. ,in It is the ReLU activation function. The Sigmoid function is used; feature recalibration is achieved through dot product. The importance of features is dynamically adjusted.

[0038] The SDAW-weighted MSE loss function was used, and the Adam optimizer was selected (initial learning rate 0.001, weight decay 1e-5). An early stopping strategy was implemented, stopping training if the loss did not decrease after 20 consecutive OOF validations to avoid overfitting. After training, the meta-feature vector corresponding to the integrated mean map was input to generate the final water depth prediction map.

[0039] Step 5: Accuracy Evaluation and Effect Verification. Three water depth checkpoints were selected in the heterogeneous study area. The inversion results were compared with the measured water depths, and a multi-dimensional accuracy index was used for evaluation. The core formula is as follows: Coefficient of determination: ; Root mean square error: ; 90% linearity error: ; IHO approval rate: ; in ( , ,).

[0040] This invention specifically addresses four major shortcomings of existing technologies. Its effectiveness has been verified through multi-regional trials, and the usage data is verifiable and traceable. Significant improvements in cross-domain accuracy: Nanshan Port (turbid water, 0.5-13m) RMSE=0.792m, R²=0.852, an improvement of 7.4% compared to the best baseline model; Qilianyu (clear water coral reef, 0.5-47.25m) RMSE=1.756m, a reduction of 33.1% compared to the best baseline model; Yellow River confluence area (turbid-clear water interaction, 0.5-9m) RMSE=0.471m, the best among all comparison models; The long-tail distribution exhibits strong adaptability: the underestimation bias in deep water (10-15m) is reduced from -2.6m to -1.3m, and the prediction accuracy in shallow water (<2m) is improved by more than 30%, with LE90 reduced to 1.114m, representing a 19.5% improvement over the optimal base model. High spatial fidelity: effectively suppresses checkerboard artifacts and pixel noise, maintains structural integrity in key areas such as channel edges, coral reef slopes, and river connectivity, and exhibits random residual distribution with no systematic bias; Engineering applicability meets standards: The pass rate for IHO Level 1a standards reached 67.7%, far exceeding that of traditional base models (48.3%), and is closer to the operational application needs of coastal engineering, navigation safety and other fields.

[0041] Figure 4The comparison of measured and predicted water depths shows that MAAS-Net has the highest good fit between the predicted and measured values ​​(R²=0.852), which is significantly better than RF, ResMLP and simple average ensemble models. Figure 5 The residual distribution and depth stratification bias of the display space were shown. The MAAS-Net residuals were randomly distributed, and the deep water bias was significantly reduced. The IHO had the highest pass rate. Figure 6 Compared to the spatial fidelity of depth maps, MAAS-Net can maintain terrain continuity in various heterogeneous environments, while traditional base models suffer from problems such as patchy noise and artificial discontinuities. Figure 7 Cross-domain generalization evaluation shows that MAAS-Net has the best RMSE stability and the lowest LE90 risk in unseen domains, verifying the universality of the method.

[0042] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A satellite depth sounding learning method driven by morphological perception and uncertainty, characterized in that, include: A heterogeneous base model layer is constructed, and a base model prediction water depth map is generated on this basis. A strict zero-leakage cross-training strategy is adopted. Based on the prediction results generated by four complementary base models, the training weights of the samples are optimized by combining the Spatial Distribution Adaptive Weights (SDAW) mechanism to generate independent Out-Of-Fold (OOF) maps and ensemble mean maps in a five-fold cross-training manner. Multi-scale orthogonal partitioning topology optimization is performed. The optimal multi-scale window is determined through automated scale search. Orthogonality penalty constraints are introduced to reduce feature redundancy. Explicit terrain topology features are calculated based on the optimal scale. A meta-learner with morphological awareness attention is constructed, fusing base model predictions, explicit terrain descriptors, and uncertainty-related features into an input vector. This vector is trained using a residual network with a Squeeze-and-Excitation (SE) attention mechanism, and the model is optimized using the SDAW-weighted MSE loss function. The meta-feature vector is then input into the meta-learner model to obtain the final inverted water depth value for each pixel. Cross-domain independent verification and accuracy assessment were implemented. A two-stage verification protocol combining intra-domain hierarchical 5-fold cross-validation and cross-domain independent verification was adopted. Statistical indicators and IHO hydrographic engineering accuracy indicators were introduced to comprehensively evaluate the model performance.

2. The morphology-aware and uncertainty-driven satellite depth sounding learning method as described in claim 1, characterized in that, The four complementary regression models in the steps include: a log-linear model based on the Beer-Lambert law, a random forest (RF) model using a Bagging strategy, an XGBoost model with gradient boosting, and a support vector regression (SVR) model based on the Nyström approximation.

3. The morphology-aware and uncertainty-driven satellite depth sounder learning method as described in claim 1, characterized in that, The Spatial Distribution Adaptive Weight (SDAW) is calculated as follows: ； in, Spatial point density is calculated using kernel density estimation (KDE). The standard deviation is predicted for four basic models. Log damping function suppresses extreme weights, and percentile pruning ensures numerical stability.

4. The satellite depth sounder learning method driven by morphological perception and uncertainty as described in claim 1, characterized in that, The multi-scale window in the above steps includes microscale (3-5 pixels), mesoscale (7-15 pixels), and macroscale (17-35 pixels). The scale triples and meta-learner hyperparameters are simultaneously optimized using the Tree-structured Parzen Estimator (TPE) algorithm.

5. The morphology-aware and uncertainty-driven satellite depth sounder learning method as described in claim 1, characterized in that, The objective function of the orthogonality penalty constraint is: ； To verify the error, The penalty coefficient is set to 0.

25. Pearson correlation coefficients are extracted between feature maps at different scales.

6. The morphology-aware and uncertainty-driven satellite depth sounder learning method as described in claim 1, characterized in that, The explicit terrain descriptor includes terrain roughness, terrain location index (BPI), texture anisotropy, and terrain gradient, which are calculated as follows: Terrain roughness: ； in For pixels The depth measurement value is as follows. This represents the number of pixels within the window. Topographic Position Index (BPI): ； in For the target pixel depth, The average depth within the macroscopic window; Texture anisotropy: ； in and These represent the spatial smoothing magnitudes of the gradients in the horizontal and vertical directions, respectively. To prevent small constants from being divided by zero; Terrain gradient: 。 7. The morphology-aware and uncertainty-driven satellite depth sounding learning method as described in claim 1, characterized in that, The input vector in the above steps is constructed as follows: ； in These are the predicted values ​​from the four basic models. For six topological description features, Include With agent SDAW value.

8. The morphology-aware and uncertainty-driven satellite depth sounder learning method as described in claim 1, characterized in that, The SDAW-weighted MSE loss function is: ； in For batch size, This is the actual depth value. To predict depth values, the Adam optimizer and an early stopping strategy are used to train the model.

9. The morphology-aware and uncertainty-driven satellite depth sounder learning method as described in claim 1, characterized in that, The accuracy evaluation indicators include the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), bias, logarithmic RMSE, 90% linearity error (LE90), and IHO pass rate, calculated as follows: ； ； ； ； Log-RMSE (LogRMSE) is used to balance the difference between shallow and deep water: ； 90% linearity error (LE90) refers to the 90th quantile of the absolute error distribution. ； IHO pass rate: The percentage of depth points that meet TVU requirements out of the total number of depth points. ； 。

Images