Large-scale water depth topographic map interpolation method based on deep learning

Abstract

The invention provides a large-scale water depth topographic map interpolation method based on deep learning. The method comprises the steps that high-resolution water depth topographic maps of the same water area at multiple time points are acquired, and a low-resolution water depth map is generated after downsampling and normalization processing. Calculating and normalizing a gradient map of the image to generate a normalized gradient map; and normalizing the plurality of time points to generate a time coding sequence. And combining the low-resolution bathymetric map, the normalized gradient map and the time coding sequence into a three-channel input tensor, and inputting the three-channel input tensor into a deep learning model. The model is trained based on a loss function of physical constraint, and a high-resolution water depth prediction model is generated. And interpolating a newly collected low-resolution bathymetric map by using the model to generate a large-scale bathymetric topographic map. According to the method, multi-time-point high-resolution data are collected and processed, a physical constraint training model is combined, a high-precision large-scale topographic map can be generated by utilizing low-cost large-scale data, and the water depth measurement efficiency and precision are effectively improved.

Description

Technical Field

[0001] This application belongs to the field of bathymetric data interpolation, and particularly relates to a large-scale bathymetric map interpolation method based on deep learning. Background Art

[0002] In the field of bathymetric survey, especially in complex waters such as port areas, due to intensive ship activities, variable terrain, and frequent sediment changes, it is necessary to regularly conduct bathymetric monitoring to ensure channel safety. Traditional bathymetric survey methods are costly, inefficient, and particularly difficult for large-scale monitoring of large areas of water. With the development of deep learning technology, it has become possible to use neural networks for interpolation processing of bathymetric data. However, existing high-resolution methods for bathymetric data based on neural networks mainly rely on image interpolation, suffering from problems such as time mismatch, sampling inconsistency, and lack of targeted training in key terrain areas, resulting in insufficient interpolation accuracy and unable to meet the requirements of high-precision bathymetric survey. Summary of the Invention

[0003] The purpose of this application is to overcome the above-mentioned defects in the prior art and provide a large-scale bathymetric map interpolation method based on deep learning.

[0004] This application provides a large-scale bathymetric map interpolation method based on deep learning, including: Obtain high-resolution bathymetric maps of the same water area at multiple time points; Downsample and normalize the high-resolution bathymetric map to generate a low-resolution bathymetric map; Calculate the gradient map based on the low-resolution bathymetric map and perform normalization processing to generate a normalized gradient map; Normalize the multiple time points to generate a time encoding sequence; Merge the low-resolution bathymetric map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; Input the three-channel input tensor into a deep learning model and train it based on a loss function with physical constraints to generate a high-resolution bathymetric prediction model; Interpolate the newly acquired low-resolution bathymetric map based on the trained high-resolution bathymetric prediction model to generate a large-scale bathymetric map.

[0005] Optionally, downsampling and normalizing the high-resolution bathymetric map to generate a low-resolution bathymetric map includes: Delete the odd rows and odd columns in the high-resolution bathymetric map, and retain the even rows and even columns to generate the low-resolution bathymetric map.

[0006] Optionally, the deep learning model is an ESRGAN generator; The ESRGAN generator includes: A residual dense block module for extracting spatio-temporal features of the three-channel input tensor; An upsampling module for increasing the resolution of the feature map to the target scale.

[0007] Optionally, the loss function of the physical constraint includes: An L1 loss term for calculating the mean absolute error between the predicted water depth value and the true value; A smoothness constraint term for penalizing the second-order gradient change in non-edge regions; A gradient continuity constraint term for ensuring that the predicted gradient is consistent with the Figure 1 normalized gradient.

[0008] Optionally, inputting the three-channel input tensor into a deep learning model and training it based on the loss function of the physical constraint to generate a high-resolution water depth prediction model, including: Dividing the data at the multiple time points into a training set, a validation set, and a test set in chronological order.

[0009] This application also provides a large-scale water depth topographic map interpolation device based on deep learning, including: An acquisition module for acquiring high-resolution water depth topographic maps of the same water area at multiple time points; A sampling module for downsampling and normalizing the high-resolution water depth topographic map to generate a low-resolution water depth map; A gradient module for calculating a gradient map based on the low-resolution water depth map and normalizing it to generate a normalized gradient map; A sequence module for normalizing the multiple time points to generate a time encoding sequence; A tensor module for combining the low-resolution water depth map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; A training module for inputting the three-channel input tensor into a deep learning model and training it based on the loss function of the physical constraint to generate a high-resolution water depth prediction model; A generation module for interpolating a newly acquired low-resolution water depth map based on the trained high-resolution water depth prediction model to generate a large-scale water depth topographic map.

[0010] Optionally, the sampling module downsamples and normalizes the high-resolution water depth topographic map to generate a low-resolution water depth map, including: Deleting the odd rows and odd columns in the high-resolution water depth topographic map and retaining the even rows and even columns to generate the low-resolution water depth map.

[0011] Optionally, the deep learning model is an ESRGAN generator; The ESRGAN generator includes: A residual dense block module for extracting spatio-temporal features of the three-channel input tensor; An upsampling module for increasing the resolution of the feature map to the target scale.

[0012] Optionally, the loss function of the physical constraint includes: An L1 loss term for calculating the mean absolute error between the predicted water depth value and the true value; A smoothness constraint term for penalizing the second-order gradient change in non-edge regions; A gradient continuity constraint term for ensuring that the predicted gradient is consistent with the Figure 1 normalized gradient.

[0013] Optionally, the training module inputs the three-channel input tensor into the deep learning model, and trains it based on the loss function of the physical constraint to generate a high-resolution water depth prediction model, including: Dividing the data at the multiple time points into a training set, a validation set, and a test set in chronological order.

[0014] The beneficial effects of this application are: This application provides a large-scale water depth topographic map interpolation method based on deep learning, including: obtaining high-resolution water depth topographic maps of the same water area at multiple time points; performing downsampling and normalization processing on the high-resolution water depth topographic maps to generate low-resolution water depth maps; calculating a gradient map based on the low-resolution water depth maps and performing normalization processing to generate a normalized gradient map; performing normalization processing on the multiple time points to generate a time encoding sequence; combining the low-resolution water depth maps, the normalized gradient maps, and the time encoding sequences into a three-channel input tensor; inputting the three-channel input tensor into a deep learning model, and training it based on the loss function of the physical constraint to generate a high-resolution water depth prediction model; interpolating a newly collected low-resolution water depth map based on the trained high-resolution water depth prediction model to generate a large-scale water depth topographic map. This application collects high-resolution water depth topographic maps of the same water area at multiple time points, and performs downsampling and normalization processing. At the same time, it calculates the gradient map and the time encoding sequence as the three-channel input tensor of the deep learning model. By introducing the loss function of the physical constraint for model training, a high-resolution water depth prediction model is generated. It can utilize low-cost large-scale water depth monitoring data to generate high-precision large-scale water depth topographic maps through interpolation, significantly improving the efficiency and accuracy of water depth measurement. Description of the Drawings

[0015] Figure 1It is a schematic flow diagram of the large-scale bathymetric map interpolation method based on deep learning in this application; Figure 2 It is a schematic diagram of the device for the large-scale bathymetric map interpolation method based on deep learning in this application. Detailed implementation manners

[0016] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it can be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, the embodiments are provided so that this disclosure can be more thoroughly understood and the scope of this disclosure can be fully conveyed to those skilled in the art.

[0017] Please refer to Figure 1 As shown, this application provides a large-scale bathymetric map interpolation method based on deep learning, including: S101. Obtain high-resolution bathymetric maps of the same water area at multiple time points; Collect high-resolution bathymetric maps of the same water area / port area at different times, and the high-resolution bathymetric maps record the water depth values of each coordinate at different time points.

[0018] Ensure that the data covers enough time points to capture the changes in the terrain.

[0019] S102. Downsample and normalize the high-resolution bathymetric maps to generate low-resolution bathymetric maps; Perform unified processing on all high-resolution bathymetric maps so that they all cover the same sea area and are converted to the same scale in the same coordinate system.

[0020] At this time, each water depth data can be converted into the format:

[0021] Among them, corresponds to the horizontal and vertical coordinate sequences, and t is the time point. Each time the water depth terrain data has rows columns, and there are data of different time points in total.

[0022] S103. Calculate the gradient map based on the low-resolution bathymetric map and perform normalization processing to generate a normalized gradient map; Generate a low-resolution topographic map and the corresponding normalized results using the above high-resolution water depth terrain data, and the scale is n times that of the original data, where n is a positive integer:

[0023] For example, when the scale is twice the original data, the corresponding result is:

[0024] At this time, in the specific operation: The high-resolution topographic map data can be regarded as a three-dimensional matrix (with dimensions i, j, and t).

[0025] Delete the odd rows and odd columns of the data at each observation time point, and retain the even rows and columns.

[0026] For example, if the high-resolution map is a 40x40 matrix and the scale is 1:1000, after deleting the odd rows and columns, the low-resolution map is a 20x20 matrix, taking the indices [1, 3, 5…17, 19] rows and [1, 3, 5…17, 19] columns from the original matrix. At this time, the scale is twice the original data, 1:2000.

[0027] Since the possible range of water depth values is relatively large (such as from 0 to hundreds of meters), it is necessary to normalize the data and map the values to the range [0, 1] to improve the network convergence speed.

[0028] The normalization formula is: ,

[0029] where, , .

[0030] In order to better capture the terrain edges in the model, such as the steep slopes and canyons on the seabed, a gradient map is generated to highlight the areas with terrain changes, so as to reflect the speed and direction of water depth changes.

[0031] Generate a gradient map using the low-resolution topographic map, and the formula is:

[0032] where, dx and dy are the depth change rates in the x and y directions respectively. This step ensures that the gradient map reflects the actual physical depth changes:

[0033] Then calculate the gradient map using the same normalization method:

[0034] where, , .

[0035] At this time, each gradient map is normalized to the range [0, 1], ensuring that the gradient map is consistent with the input scale of the depth map and can reflect the actual terrain change scale at the same time.

[0036] S104. Normalize the multiple time points to generate a time encoding sequence; Normalize the time points t of each data for all the topographic maps collected. The normalization formula is:

[0037] To ensure the training effect, divide all the segments into a training set (80%), a validation set (10%), and a test set (10%) according to the time series t of different data sets.

[0038] That is, the training set data is the early data, and the validation set and the test set are the recent data.

[0039] S105. Combine the low-resolution bathymetric map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; Select a deep learning architecture suitable for super-resolution: Use the ESRGAN generator (Enhanced Super-Resolution Generative Adversarial Network) as the basic network. The ESRGAN generator includes a shallow feature extraction layer, a deep feature extraction layer (including multiple residual blocks, Residual-in-Residual Dense Blocks, RRDB), an upsampling layer, and an output layer.

[0040] Configure the network input as a three-channel tensor and the output as a topographic map with n times the resolution. Specifically, for the training samples at each time point t, the three-channel inputs are respectively: Channel 1: The normalized low-resolution depth map, that is, the corresponding to time t; Channel 2: The normalized gradient map, that is, the corresponding to time t; Channel 3: An array with the same size as the low-resolution map in Channel 1 and all pixel values are .

[0041] Configure the network output: Output the depth matrix with n times the high resolution corresponding to this time point t result.

[0042] Calculate the loss function by combining all the training set data. The loss function selects the L1 loss (mean absolute error MAE) or the L2 loss (mean square error). The formula is:

[0043] Among them, is the normalization result of the original terrain data, is the predicted value corresponding to time t, and N is the total amount of data. .

[0044] At the same time, add physical constraint terms to the loss function, such as smoothness constraint (penalize the drastic changes in the depth map in non-edge regions, using the L2 norm of the second derivative of the gradient):

[0045] and gradient continuity constraint (ensure that the predicted gradient is Figure 1 consistent with the input gradient, and add gradient loss):

[0046] The loss function can be defined as a combination of the above several:

[0047] where

[0048] S106. Input the three-channel input tensor into the deep learning model and train it based on the loss function with physical constraints to generate a high-resolution water depth prediction model; Use the Adam optimizer with an initial learning rate of 0.0001, and adjust the batch size according to the computing resources.

[0049] During the learning process, adopt the cosine annealing learning rate schedule to reduce the learning rate every 10 epochs (such as from 0.0001 to 0.00001), or use learning rate decay (such as multiplying by 0.1 every 50 epochs), which helps to jump out of local optima and improve the later training effect.

[0050] Monitor the loss function of the validation set during the learning process. If there is no improvement for several consecutive epochs (for example, five), stop training in advance to prevent overfitting.

[0051] Train for a sufficient number of epochs (such as 100 - 200) until the validation loss converges to obtain the trained network model.

[0052] Using the trained network model, input the normalized test set data to obtain the predicted test set results, and then perform denormalization operations on the n-fold high-resolution depth matrix results of all outputs to obtain the depth values.

[0053] Calculate MAE and PSNR to evaluate the performance of the model. The calculation method of MAE is the same as that of the loss function:

[0054] PSNR is calculated according to the following method: First, perform denormalization operations on the results output by the model:

[0055] Among them, max is the maximum value of the original test set data. Evaluate the performance of the model according to the above indicators. An ablation experiment can also be added during the evaluation to test the differences in Channel 2 (gradient map), Channel 3 (time series), and the values of various loss functions.

[0056] S107. Interpolate the newly acquired low-resolution bathymetric map based on the trained high-resolution bathymetry prediction model to generate a large-scale bathymetric topographic map.

[0057] For the same water area / port area within the monitoring range, during subsequent observations, use a low-cost large-scale bathymetric monitoring method to obtain a large-scale topographic map; Perform normalization using the same global minimum and maximum values as the model training set, input the trained model, and predict the large-scale bathymetric topography; Then denormalize to the actual depth value.

[0058] Post-process the predicted high-resolution topographic map to check for unreasonable depth values (such as extremely large or extremely small values). When necessary, apply smoothing filtering (such as Gaussian blur) or clip to a reasonable range, or terrain adaptive smoothing can also be used.

[0059] As Figure 2 shown, the present application also provides a large-scale bathymetric topographic map interpolation device based on deep learning, including: An acquisition module 201 for acquiring high-resolution bathymetric topographic maps of the same water area at multiple time points; A sampling module 202 for downsampling and normalizing the high-resolution bathymetric topographic map to generate a low-resolution bathymetric map; A gradient module 203 for calculating a gradient map based on the low-resolution bathymetric map and performing normalization processing to generate a normalized gradient map; A sequence module 204 for normalizing the multiple time points to generate a time encoding sequence; A tensor module 205 for combining the low-resolution bathymetric map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; A training module 206 for inputting the three-channel input tensor into a deep learning model and training based on a loss function with physical constraints to generate a high-resolution bathymetry prediction model; A generation module 207 for interpolating a newly acquired low-resolution bathymetric map based on the trained high-resolution bathymetry prediction model to generate a large-scale bathymetric topographic map.

[0060] Further, the sampling module downsamples and normalizes the high-resolution bathymetric topographic map to generate a low-resolution bathymetric map, including: Delete the odd rows and odd columns in the high-resolution bathymetric topographic map, and retain the even rows and even columns to generate the low-resolution bathymetric map.

[0061] Further, the deep learning model is an ESRGAN generator; The ESRGAN generator includes: A residual dense block module for extracting spatio-temporal features of the three-channel input tensor; An upsampling module for increasing the resolution of the feature map to the target scale.

[0062] Further, the loss function of the physical constraint includes: An L1 loss term for calculating the mean absolute error between the predicted bathymetric value and the true value; A smoothness constraint term for penalizing the second-order gradient change in non-edge regions; A gradient continuity constraint term for ensuring that the predicted gradient is consistent with the normalized gradient Figure 1 consistent.

[0063] Further, the training module inputs the three-channel input tensor into the deep learning model and trains it based on the loss function of the physical constraint to generate a high-resolution bathymetric prediction model, including: Split the data at the multiple time points into a training set, a validation set, and a test set in chronological order.

[0064] The above description of the embodiments is to facilitate the understanding and application of the present invention by those of ordinary skill in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative labor. Therefore, the present invention is not limited to the above embodiments, and the improvements and modifications made by those skilled in the art to the present invention should be within the protection scope of the present invention.

Claims

1. A large-scale bathymetric topographic map interpolation method based on deep learning, characterized in that, Including: Obtain high-resolution bathymetric topographic maps of the same water area at multiple time points; Downsample and normalize the high-resolution bathymetric topographic maps to generate low-resolution bathymetric maps; Calculate a gradient map based on the low-resolution bathymetric map and perform normalization processing to generate a normalized gradient map; Normalize the multiple time points to generate a time encoding sequence; Merge the low-resolution bathymetric map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; Input the three-channel input tensor into a deep learning model and train it based on a loss function with physical constraints to generate a high-resolution bathymetric prediction model; Interpolate a newly acquired low-resolution bathymetric map based on the trained high-resolution bathymetric prediction model to generate a large-scale bathymetric topographic map.

2. The large-scale bathymetric topographic map interpolation method based on deep learning according to claim 1, wherein Downsample and normalize the high-resolution bathymetric topographic maps to generate low-resolution bathymetric maps, including: Delete the odd rows and odd columns in the high-resolution bathymetric topographic maps and retain the even rows and even columns to generate the low-resolution bathymetric maps.

3. A large-scale bathymetric topographic map interpolation method based on deep learning according to claim 1, characterized in that, The deep learning model is an ESRGAN generator; The ESRGAN generator includes: A residual dense block module for extracting spatio-temporal features of the three-channel input tensor; An upsampling module for increasing the resolution of the feature map to the target scale.

4. A large-scale bathymetric topographic map interpolation method based on deep learning according to claim 1, characterized in that, The loss function with physical constraints includes: An L1 loss term for calculating the mean absolute error between the predicted bathymetric values and the true values; A smoothness constraint term for penalizing the second-order gradient changes in non-edge regions; A gradient continuity constraint term for ensuring that the predicted gradient is consistent with the normalized gradient map.

5. A large-scale bathymetric topographic map interpolation method based on deep learning according to claim 1, characterized in that Input the three-channel input tensor into a deep learning model and train it based on a loss function with physical constraints to generate a high-resolution bathymetric prediction model, including: Divide the data of the multiple time points into a training set, a validation set, and a test set in chronological order.

6. An interpolation device for large-scale bathymetric topographic maps based on deep learning, characterized in that, Including: An acquisition module for obtaining high-resolution bathymetric topographic maps of the same water area at multiple time points; A sampling module for downsampling and normalizing the high-resolution bathymetric topographic maps to generate low-resolution bathymetric maps; A gradient module for calculating a gradient map based on the low-resolution bathymetric map and performing normalization processing to generate a normalized gradient map; A sequence module for normalizing the multiple time points to generate a time encoding sequence; A tensor module for merging the low-resolution bathymetric map, the normalized gradient map, and the time encoding sequence into a three-channel input tensor; A training module for inputting the three-channel input tensor into a deep learning model and training it based on a loss function with physical constraints to generate a high-resolution bathymetric prediction model; A generation module for interpolating a newly acquired low-resolution bathymetric map based on the trained high-resolution bathymetric prediction model to generate a large-scale bathymetric topographic map.

7. An interpolation device for large-scale bathymetric topographic maps based on deep learning according to claim 6, characterized in that, The sampling module downsamples and normalizes the high-resolution bathymetric topographic maps to generate low-resolution bathymetric maps, including: Delete the odd rows and odd columns in the high-resolution bathymetric topographic maps and retain the even rows and even columns to generate the low-resolution bathymetric maps.

8. A large-scale bathymetric topographic map interpolation device based on deep learning according to claim 6, characterized in that The deep learning model is an ESRGAN generator; The ESRGAN generator includes: Residual Dense Block module for extracting spatio-temporal features of the three-channel input tensor; Upsampling module for upsampling the feature map resolution to the target scale.

9. The large-scale bathymetric topographic map interpolation device based on deep learning according to claim 6, characterized in that The loss function of the physical constraint includes: L1 loss term for calculating the mean absolute error between the predicted water depth value and the true value; Smoothness constraint term for penalizing the second-order gradient change in non-edge regions; Gradient continuity constraint term for ensuring that the predicted gradient is consistent with the normalized gradient map.

10. The large-scale bathymetric topographic map interpolation device based on deep learning according to claim 6, wherein, The training module inputs the three-channel input tensor into the deep learning model and trains it based on the loss function of the physical constraint to generate a high-resolution water depth prediction model, including: Dividing the data at the multiple time points into a training set, a validation set, and a test set in chronological order.

Images