Optical remote sensing image inversion intertidal zone terrain method based on sequence model

Abstract

The invention discloses an intertidal zone terrain inversion method based on an optical remote sensing image of a sequence model, and relates to the field of intertidal zone terrain inversion, and the method comprises the steps: filtering cloud pixels in an intertidal zone remote sensing image, and obtaining preprocessed remote sensing data; generating a water body mask of the preprocessed remote sensing data based on the normalized difference water body index, and determining an intertidal zone range by using a time sequence mask superposition and morphological optimization processing technology; in the intertidal zone range, performing standardization and equalization processing on the multispectral index calculation result and data of each wave band of the intertidal zone remote sensing image to obtain model training data; and constructing an intertidal zone elevation prediction model, and outputting an intertidal zone terrain inversion result by using the intertidal zone elevation prediction model. The method disclosed by the invention reveals the correlation between the intertidal zone elevation and the time sequence remote sensing image, breaks through the space-time limitation of the traditional intertidal zone topographic measurement method, and realizes the automatic inversion of the intertidal zone topography based on the optical image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intertidal zone terrain inversion, in particular to an optical remote sensing image-based intertidal zone terrain inversion method based on a sequence model. BACKGROUND

[0002] The intertidal zone, as a dynamic transition area between the ocean and the land, presents a regular submergence-exposure alternation feature under the periodic action of tides. The unique bottom structure and geomorphic features of this area form a highly specialized ecosystem, providing breeding grounds and habitats for numerous marine organisms, and playing an important ecological function in mitigating coastal erosion and maintaining fishery resources. Influenced by the accelerated rise of global sea level caused by climate change and high-intensity human development activities, this area is experiencing systematic ecological degradation, and precise terrain data is urgently needed to reveal its evolution mechanism.

[0003] Obtaining high-precision intertidal zone terrain data is a basic work for ecological protection. Traditional ground survey methods such as ground-based laser scanning and RTK measurement can achieve centimeter-level accuracy, but are limited by the twice-a-day tidal submergence cycle, with inherent defects such as short operation window and high labor cost. Airborne laser radar technology can partially improve the survey efficiency, but its airspace coordination difficulty and equipment operation cost still restrict its large-scale application. The existing global digital elevation model generally has data missing or meter-level error in the intertidal zone area, which cannot meet the needs of fine management.

[0004] The breakthrough of satellite remote sensing technology provides a new way for intertidal zone terrain monitoring. The current mainstream methods can be divided into synthetic aperture radar and optical remote sensing two technical routes, among which the method based on optical image has more application potential due to its high standardization of data processing flow. Typical optical inversion methods include waterline method and submergence frequency method: the former relies on high and low tide images of long time series to capture the water-land boundary, but has strict requirements on image quality and continuity; the latter constructs an elevation model by statistically analyzing the submergence frequency of pixels, which reduces the requirements on single image, but still needs to accumulate enough observation data.

[0005] There are two technical bottlenecks in the existing methods: first, the traditional water-land binary processing method cannot fully exploit the nonlinear correlation between multi-band reflectivity and elevation; second, the shielding ability of interference factors such as bottom type difference and change of solar incident angle is insufficient. This suggests that the spectral response mechanism of the intertidal zone needs to be reexamined, i.e. the stable water-containing feature formed by frequent submergence in low-altitude areas presents specific reflection signals in the short-wave infrared band; while the dry process of the bottom in high-altitude areas caused by long-term exposure has a significant correlation with the reflectivity in the visible light band. Such spectral-elevation correlation provides a physical basis for establishing a new inversion model, but how to extract effective features from complex time-series images is still a technical difficulty.

[0006] The prior art has not yet proposed an effective solution to the problems in the related art. SUMMARY

[0007] To overcome the limitations of the existing related art, the present application proposes a method for retrieving intertidal zone terrain from optical remote sensing images based on a sequence model, which overcomes the dependence of the existing intertidal zone terrain reconstruction method on continuous observation data and the uneven distribution of data. Compared with the traditional waterline method and the submergence frequency method, the present application excavates the nonlinear correlation between multi-band spectral characteristics and elevation, combines an adaptive data balancing strategy, and realizes high-precision terrain reconstruction under non-continuous observation conditions. The most prominent cost problem in intertidal zone terrain extraction is solved, and the intertidal zone terrain is automatically extracted based on a sequence regression deep learning model.

[0008] To this end, the specific technical solutions adopted by the present application are as follows: A method for retrieving intertidal zone terrain from optical remote sensing images based on a sequence model, comprising: Based on the cloud coverage percentage determination result of the intertidal zone remote sensing image, and combined with a cloud detection algorithm, cloud pixels in the intertidal zone remote sensing image are filtered to obtain preprocessed remote sensing data; Based on the normalized difference water index, a water mask of the preprocessed remote sensing data is generated, and time series mask superposition and morphological optimization processing techniques are used to determine the intertidal zone range; In the intertidal zone range, multi-spectral index calculation is performed on the pixels of the preprocessed remote sensing data; the multi-spectral index calculation result and the data of each band of the intertidal zone remote sensing image are standardized and equalized to obtain model training data; Based on the model training data, a double deep learning module is trained based on the principle of ensemble learning, and an intertidal zone elevation prediction model is constructed; the intertidal zone terrain retrieval result is output using the intertidal zone elevation prediction model, and is compared and verified with the measured elevation data.

[0009] Further, based on the cloud coverage percentage determination result of the intertidal zone remote sensing image, and combined with a cloud detection algorithm, cloud pixels in the intertidal zone remote sensing image are filtered to obtain preprocessed remote sensing data, which comprises: Based on the scene-level cloud mask attribute in the metadata of the intertidal zone remote sensing image, the cloud coverage percentage of the remote sensing image is calculated, and the remote sensing image with a cloud coverage ratio above a preset threshold is filtered; The filtered intertidal zone remote sensing image is processed using a cloud detection algorithm to remove cloud pixels in the intertidal zone remote sensing image, and preprocessed remote sensing data is obtained.

[0010] Further, based on the normalized difference water index, a water mask of the preprocessed remote sensing data is generated, and time series mask superposition and morphological optimization processing techniques are used to determine the intertidal zone range, which comprises: generating a water mask according to the normalized difference water index and the difference water index threshold of each single image in the pre-processed remote sensing data; After superimposing the water masks at different time points, morphological optimization processing is performed to obtain the intertidal zone range.

[0011] Further, generating a water mask according to the normalized difference water index and the difference water index threshold of each single image in the pre-processed remote sensing data includes: determining the difference water index threshold; and performing normalization processing on the green band and the near-infrared band in the pre-processed remote sensing data to obtain the normalized difference water index; when the normalized difference water index is greater than or equal to the difference water index threshold, generating a water mask, and using the water mask to divide the water and land in the pre-processed remote sensing data.

[0012] Further, in the intertidal zone range, performing multi-spectral index calculation on the pixels of the pre-processed remote sensing data includes: performing calculation of the normalized vegetation index, the normalized water index, and the improved normalized water index on the pixels using the green band, the red band, the near-infrared band, and the short-wave infrared band in the pre-processed remote sensing data in the intertidal zone range.

[0013] Further, performing standardization and equalization processing on the multi-spectral index calculation results and the band data of the intertidal zone remote sensing image to obtain model training data includes: in the order of the blue light normalized index, the green light normalized index, the red light normalized index, the infrared normalized index, the short-wave infrared normalized index, the normalized vegetation index, the normalized water index, the improved normalized water index, and the elevation, constructing a standardized feature sequence; dividing the elevation interval according to a preset step size, and grouping the standardized feature sequence; using hierarchical sampling to optimize the data distribution rule, and performing equalization processing on the standardized feature sequence samples of each group; dividing the equalization-processed standardized feature sequence into model training data including a training set and a test set according to a preset proportion.

[0014] Further, using hierarchical sampling to optimize the data distribution rule to perform equalization processing on the standardized feature sequence samples of each group includes: when the standardized feature sequence samples of each group are less than a sample lower limit value, discarding the corresponding group; when the standardized feature sequence samples of each group are greater than or equal to the sample lower limit value and less than or equal to a sample upper limit value, retaining all of them; When the normalized feature sequence sample of each group is greater than the sample upper limit value, then a normalized feature sequence sample of the sample upper limit value is randomly collected.

[0015] Further, based on the model training data, the double deep learning module is trained in combination with the ensemble learning principle, and a tidal zone elevation prediction model is constructed, including: Based on the ensemble learning principle, the recurrent neural network module and the long short-term memory network module are combined to construct a time series regression model; the model training data is processed through the time series regression model; The output of the time series regression model is connected using a feature splicing layer, and an elevation prediction value is output using a fully connected layer; Based on the loss function and the optimizer, the time series regression model is trained and optimized to obtain the tidal zone elevation prediction model.

[0016] Further, the tidal zone elevation prediction model is used to output the tidal zone terrain inversion result, and compared with the measured elevation data for verification, including: The preprocessed remote sensing data is used as input, the tidal zone elevation prediction model is used to invert the tidal zone terrain, and the measured elevation data is compared to verify the accuracy of the water depth inversion.

[0017] Further, when the tidal zone terrain is inverted and compared with the measured elevation data, evaluation criteria including root mean square error, mean absolute error and determination coefficient are used to verify the accuracy of the water depth inversion.

[0018] The beneficial effects of the present application are: (1) Break through the continuous observation limit: through morphological optimization and time series mask superposition technology, the dynamic area between the highest and lowest tide levels is effectively extracted, solving the problem of dependence on continuous high and low tide images in traditional methods.

[0019] (2) Enhance data representation ability: innovatively integrate NDWI, MNDWI and other multispectral indexes to construct a 9-dimensional time series feature vector, which improves the feature discrimination degree by 40% compared with single band method.

[0020] (3) Improve model performance: adopt elevation interval layered sampling strategy to ensure the stable prediction ability of the model in different terrain areas.

[0021] (4) The present application realizes three major innovative breakthroughs of the tidal zone terrain reconstruction method: firstly, a deep correlation model of multi-band reflectivity-elevation is established, and a RNN-LSTM hybrid architecture of spatio-temporal feature fusion is proposed. It provides a new technical paradigm for global coastal zone terrain monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0023] Figure 1 is a general flowchart of the present application example; Figure 2 is a general flowchart of the present application example; Figure 3 is a highest-lowest tide level range graph extracted; Figure 4 is a structure diagram of the intertidal zone elevation prediction model; Figure 5 is a prediction result graph of the intertidal zone elevation; Figure 6 is a measured data graph of the intertidal zone elevation; Figure 7 is a comparison graph of the model predicted elevation and the measured elevation; Figure 8 is a display graph of each step according to the embodiment of the present application; Figure 9 is a flowchart of an optical remote sensing image inversion intertidal zone terrain method based on a sequence model according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] To further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can be used to explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible embodiments and advantages of the present application by referring to these contents. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0025] According to the embodiment of the present application, an optical remote sensing image inversion intertidal zone terrain method based on a sequence model is provided, in particular a method which can realize automatic extraction of remote sensing image features and automatic extraction of intertidal zone terrain based on a deep learning sequence regression model.

[0026] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figure 9 The optical remote sensing image inversion intertidal zone terrain method based on a sequence model according to the embodiment of the present application includes: S1, based on the cloud coverage percentage judgment result of the intertidal zone remote sensing image, and combined with a cloud detection algorithm, filtering the cloud pixels in the intertidal zone remote sensing image to obtain the preprocessed remote sensing data.

[0027] S2. A water mask is generated based on the normalized differential water index after preprocessing the remote sensing data, and the intertidal zone range is determined by time series mask overlay and morphological optimization processing techniques.

[0028] S3. Within the intertidal zone, multispectral indexes are calculated for the pixels of the preprocessed remote sensing data. The multispectral index calculation results and the data of each band of the intertidal remote sensing image are standardized and equalized to obtain model training data.

[0029] S4. Based on the model training data and combined with the principle of ensemble learning, train dual deep learning modules and construct an intertidal elevation prediction model; use the intertidal elevation prediction model to output the intertidal topography inversion results and compare and verify them with the measured elevation data.

[0030] In one embodiment, based on the cloud cover percentage determination result of intertidal remote sensing imagery and combined with a cloud detection algorithm, cloud pixels in the intertidal remote sensing imagery are filtered to obtain preprocessed remote sensing data, including: Based on the scene-level cloud mask attribute in the metadata of intertidal remote sensing images, the cloud coverage percentage of the remote sensing images is calculated, and remote sensing images with a cloud coverage percentage above a preset threshold are filtered out. The cloud detection algorithm is used to perform cloud masking on the filtered intertidal remote sensing images to remove cloud pixels from the intertidal remote sensing images, thus obtaining preprocessed remote sensing data.

[0031] In one embodiment, a water mask is generated from the preprocessed remote sensing data based on the normalized differential water index, and the intertidal zone extent is determined using time-series mask overlay and morphological optimization techniques, including: Based on the normalized differential water index and differential water index threshold of each single image in the preprocessed remote sensing data, a water mask is generated; after overlaying the water masks at different time points, morphological optimization is performed to obtain the intertidal zone range.

[0032] In one embodiment, generating a water mask based on the normalized differential water index and differential water index threshold for each individual image in the preprocessed remote sensing data includes: Determine the threshold for the differential water body index; normalize the green band and near-infrared band in the preprocessed remote sensing data to obtain the normalized differential water body index; when the normalized differential water body index is greater than or equal to the threshold, generate a water body mask and use the water body mask to separate water bodies and land in the preprocessed remote sensing data.

[0033] In one embodiment, calculating the multispectral index of pixels in preprocessed remote sensing data within the intertidal zone includes: Using the green, red, near-infrared, and shortwave infrared bands from preprocessed remote sensing data of the intertidal zone, the normalized vegetation index, normalized water index, and improved normalized water index of each pixel were calculated.

[0034] In one embodiment, the multispectral index calculation results and the data of each band of the intertidal remote sensing image are standardized and equalized to obtain the model training data, including: Standardized feature sequences are constructed in the following order: normalized blue light index, normalized green light index, normalized red light index, normalized infrared index, normalized shortwave infrared index, normalized vegetation index, normalized water index, improved normalized water index, and elevation. Elevation intervals are divided according to a preset step size, and the standardized feature sequences are grouped. Layered sampling is used to optimize data distribution rules, and the standardized feature sequence samples in each group are balanced. The balanced standardized feature sequences are then divided into model training data, including training and testing sets, according to a preset ratio.

[0035] In one embodiment, optimizing the data distribution rules using stratified sampling and performing equalization processing on the standardized feature sequence samples of each group includes: If the number of standardized feature sequence samples in each group is less than the lower limit of the sample size, the corresponding group is discarded; if the number of standardized feature sequence samples in each group is greater than or equal to the lower limit of the sample size and less than or equal to the upper limit of the sample size, all samples are retained; if the number of standardized feature sequence samples in each group is greater than the upper limit of the sample size, a number of standardized feature sequence samples equal to the upper limit of the sample size are randomly collected.

[0036] In one embodiment, based on model training data and combined with ensemble learning principles, a dual deep learning module is trained to construct an intertidal elevation prediction model, including: Based on the principle of ensemble learning, and combined with recurrent neural network and long short-term memory network modules, a temporal regression model is constructed; the training data of the model is processed through the temporal regression model; a feature splicing layer is used to connect the output of the temporal regression model, and a fully connected layer is used to output the elevation prediction value; based on the loss function and optimizer, the temporal regression model is trained and optimized to obtain the intertidal elevation prediction model.

[0037] In one embodiment, the intertidal zone elevation prediction model is used to output the intertidal zone topographic inversion results, and the results are compared and verified with measured elevation data, including: Using preprocessed remote sensing data as input, the intertidal zone topography was inverted using an intertidal zone elevation prediction model, and compared with measured elevation data to verify the accuracy of the water depth inversion.

[0038] In one embodiment, when inverting the intertidal topography and comparing it with measured elevation data, an evaluation criterion including root mean square error, mean absolute error, and coefficient of determination is used to verify the accuracy of the water depth inversion.

[0039] To facilitate understanding of the above technical solutions of the present invention, the working principle of the present invention in actual process will be described in detail below.

[0040] Saint Mark National Ecological Reserve (such as) Figure 1 (As shown). This protected area extends 121 kilometers along the Gulf Coast, encompassing a complex intertidal ecosystem of approximately 1,870 square kilometers, including various geomorphic units such as coastal wetlands, tidal flats, and underwater sandbars. Choosing this area as a typical study site is significant, as its proximity to the Mississippi River estuary allows it to experience both river alluvial and tidal effects, making it an ideal experimental site for studying the dynamic evolution of intertidal topography.

[0041] This embodiment uses the study area as an example to illustrate a method for automatically extracting intertidal topography from optical remote sensing images based on a sequence model. Figure 2 The flowchart shows the process: Sentinel-2 time-series imagery - remote sensing image preprocessing - preprocessed Sentinel-2 imagery - extraction of pixels within the highest and lowest tide levels - Sentinel-2 pixels within the highest and lowest tide levels - sequence data construction - reflectance, index, and height sequence (elevation validation data collected by UAV / ICESat-2) - sequence regression model training - model weights - intertidal topography inversion and validation - intertidal topographic map. ICESat-2 is a satellite equipped with a single-photon lidar. The data and reference standards required for this example are shown in Table 1 below.

[0042] Table 1. Data and Reference Standards Required for This Example Specifically, the following steps are included: Step 1: Remote Sensing Image Preprocessing. For the cloud masking of the time-series Sentinel-2 remote sensing images, image quality issues caused by cloudy weather in the intertidal zone require cloud image processing. First, the cloud coverage percentage is calculated based on the scene-level cloud mask attributes in the remote sensing image metadata. Then, the scene-level cloud coverage percentage is filtered out. The above images are then used to generate a refined cloud mask using a cloud detection algorithm, which removes cloud pixels from the images, ultimately obtaining high-quality land cover information.

[0043] Filtering scene-level cloud coverage ratio The above image illustrates this process, which can be represented by the following formula: ; in For satellite image collection, This is a scenario-level cloud coverage ratio attribute. For the threshold, This is a collection of satellite images of areas with few clouds.

[0044] Sentinel-2 is a satellite constellation operated by the European Space Agency (ESA). The mission currently consists of two satellites (Sentinel-2A and Sentinel-2B) designed to provide high-resolution Earth observation data with a global revisit interval of 5 days. Ten Sentinel-2Level-2A surface reflectance maps of the study area, acquired from January to May 2020, were collected using the GEE platform (Geospatial Analysis Platform). Due to the cloudy climate in the intertidal zone, cloud cover images required processing. First, the cloud cover percentage was calculated using the CLOUDY_PIXEL_PERCENTAGE attribute in the image metadata. Cloud cover percentages were then filtered to a threshold value. More than 20% of the images are then processed to generate cloud masks based on s2cloudless (a technology for removing cloud pixels from satellite images), thus removing cloud pixels from the images.

[0045] Step 2: Intertidal Zone Extraction. Tidal cycle analysis is used to determine the extreme ranges of tidal level changes, and pixels exhibiting complete submersion-exposure alternation characteristics within these ranges are selected. Specifically, a water mask is generated from a single image based on the Normalized Difference Water Index (NDWI). The effective intertidal zone region is determined through time-series mask overlay, and the mask image is morphologically optimized to obtain the accurate intertidal zone range.

[0046] 1) Generation of water body masks for single images: Since the intertidal topography method based on sequence model optical remote sensing image inversion provided by this invention follows the basic assumption that tidal action causes a correlation between surface reflectance and elevation, this assumption can only be true for areas that have been submerged and exposed within the time period corresponding to the remote sensing image. Therefore, it is necessary to obtain the area between the lowest and highest tide levels in the time series.

[0047] To determine the location of these areas, a mask is generated based on the Normalized Differential Water Index (NDWI) with a certain threshold (set to 0.1 here) to separate water bodies and land within a single image. This process can be represented by the following formula: ; ; in, The green band of the Sentinel-2 image. It is in the near-infrared band. The normalized water index, A water mask for a single image.

[0048] 2) Time series mask overlay: Then, to reduce misidentification caused by factors such as waves, fog, and ships in a single image, all masked images within the time series are summed and averaged to extract the region with pixel values ​​between (0, N-1 / N). The number of mask images.

[0049] ; ; in, The frequency that is identified as a body of water. The frequency that is identified as a body of water. The number of times a pixel is identified as a body of water. It serves as a mask for the intertidal zone.

[0050] 3) Morphological optimization processing: The resulting region may still contain numerous noise patches and voids. Opening operations and area filtering are used to remove these patches, and closing operations are used to remove voids, yielding the final effective region mask for that time period. This process can be represented by the following formula: ; ; ; in, For filtering operations, To enable operation, For closing operations, For image, For convolution kernel, For corrosion operation, This is an expansion operation. The final result of this step is as follows: Figure 3 As shown.

[0051] Step 3: Sequence Data Construction. Construct the surface reflectance-exponential sequence of pixels in multi-temporal remote sensing images.

[0052] 1) Multispectral index calculation, which involves calculating the NDVI, NDWI, and MNDWI spectral indices for pixels within the effective region: This invention constructs a multi-band surface reflectance-elevation model for the intertidal zone based on a deep learning sequence regression algorithm. It estimates intertidal elevation using Sentinel-2 pixel time series data within the highest and lowest tide levels from January to May 2020. Sequence regression algorithms typically construct time series models of input bands by fitting their variations to predict elevation values. However, due to weather factors in the study area, cloud cover exists in the available remote sensing images, resulting in unequal time series lengths in the training data and limiting the direct application of these time series. Therefore, bands with strong correlation to elevation are selected, a normalized water index is calculated, and the red, green, blue, near-infrared, and shortwave infrared bands are normalized and their mean values ​​are calculated. This process can be represented by the following formula: ; ; in, Normalized Difference Vegetation Index (NDVI) For the modified normalized water index, R It is in the red light band. SWIR It is in the shortwave infrared band.

[0053] 2) Feature sequence standardization: To facilitate the training of the time series regression model and maintain sequence stability, these bands are arranged in the order of blue light, green light, red light, near-infrared, short-wave infrared, NDVI, NDWI, MNDWI, and elevation into a feature sequence with a time step of 9. The blue light, green light, red light, near-infrared, and short-wave infrared bands are normalized and their mean values ​​are calculated.

[0054] The normalization process uses the maximum-minimum normalization method, and the formula is: ; in , These are the maximum and minimum values ​​for the corresponding band. This is the normalized value.

[0055] 3) Training data equalization processing, i.e., optimizing data distribution through stratified sampling of elevation intervals to form time series input data: The elevation distribution of Sentinel-2 pixels within the highest and lowest tide range is highly uneven. In this study area, the elevations corresponding to Sentinel-2 pixels within this range are mainly concentrated between -1.6m and -0.4m. This distribution is very unfavorable for model training. Therefore, the original data was processed to make it more uniform. Specifically, the elevation intervals were first divided with a step size of 0.1m. The data was then grouped. If the number of samples in each interval was less than 50, they were discarded; if it was more than 200, 200 samples were randomly selected; and if it was between 50 and 200, all samples were retained. This uniformization process ensures the representativeness of the data and reduces the data size, thus avoiding bias caused by uneven data distribution to some extent. Finally, the constructed data was randomly sampled in a 7:3 ratio to divide the training and test datasets.

[0056] That is, sampling and training-test data splitting are performed according to the following rules: .

[0057] Step 4: Deep Learning Temporal Regression Modeling. A deep learning regression model is established using surface reflectance-exponential time series and elevation data. Based on the principle of ensemble learning, parallel RNN and LSTM modules process the input sequence. A feature concatenation layer connects the two modules to output fused features. Finally, a fully connected layer outputs the predicted elevation value. A task-appropriate MSE loss function and Adam optimizer are used to construct the intertidal elevation prediction model.

[0058] 1) Model Architecture: Since the values ​​in the sequence are not linearly correlated with the elevation, deep learning models need to pay more attention to the changing patterns between different bands in the sequence. Therefore, compared to traditional neural networks that treat spectral band information as discrete individuals, recurrent neural networks are more suitable for the task of this invention. To improve the performance of the classic deep learning time series regression model and make it better suited for the intertidal elevation prediction task, a new time series regression model is constructed based on the principle of ensemble learning, combining two mainstream time series regression networks: RNN (Recurrent Neural Network) and LSTM (Long Short-Term Memory Network). The model design is as follows: Figure 4 As shown, it includes an RNN module, an LSTM module, and a fully connected module. The input sequence distribution is processed by the RNN and LSTM modules, and then concatenated into the fully connected module (FC layer) using a concat method.

[0059] RNN unit computation: ; ; in, For the current moment The hidden state, For the current input, This is the hidden state from the previous time step. This is the output at the current moment. The weight of the hidden state in the previous time step. The weight input at this moment, The output sample weights, and All are tanh activation functions; the RNN module contains 64 hidden units, the fully connected layer contains 3 hidden layers, and the number of neurons is 128, 64, and 32 respectively.

[0060] LSTM cell computation: ; ; ; ; ; ; in, Here is the weight matrix of the input gate. For the bias term of the input gate, For the sigmoid function, For the tanh function, This is the output vector of candidate cell states; , These are the cell state at the previous time step and the cell state at the current time step, respectively. The output value of the input gate. The weight matrix represents the candidate unit states. The bias term for the candidate cell state. Here is the weight matrix for the forget gate. For the bias term of the forget gate, Here is the weight matrix of the output gate. This is the bias term for the output gate. The activation value of the output gate controls the output of the hidden state at the current moment; the LSTM module contains 64 hidden units, and the fully connected layer contains 3 hidden layers with 128, 64, and 32 neurons respectively.

[0061] The concat feature concatenation layer connects the two modules to output fused features, and finally a fully connected layer is used to output the elevation prediction value: ; in, This is a Concat feature splicing layer. and These are the weights and biases of the fully connected layer, respectively.

[0062] 2) Loss Function and Optimization: The method of this invention is implemented based on the PyTorch framework (a deep learning framework). According to the network training requirements, the network is trained for 1000 epochs, where an epoch represents the training process that goes through all samples in the training dataset. The learning rate is set to 1e-3, and the Adam optimizer is used for training with a batch size of 128. Furthermore, a task-appropriate MSE loss function is used, calculated as follows: ; in, It is the actual value. It is the output of the model. n This represents the number of verification points.

[0063] Step 5: Intertidal Topographic Inversion and Validation. The trained sequential regression model is used to invert the intertidal topography and compared with measured elevation data to validate its accuracy. Specifically, preprocessed Sentinel-2 remote sensing imagery and the trained sequential regression model are used to invert the intertidal topography. The measured elevation data is compared with the predicted intertidal topographic map. The accuracy of the water depth inversion is evaluated using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The performance of the model in intertidal elevation inversion is analyzed.

[0064] 1) Model inversion: The trained sequential regression model was used to perform intertidal topography inversion and compared with measured elevation data to verify the accuracy. Figure 5 This demonstrates the extent and elevation of the unvegetated intertidal zone in the example area from January to May 2020, as output by the model. Specifically, using the Sentinel-2 remote sensing imagery processed in step 1 as input, the sequential regression model trained in step 4 is used to perform intertidal topographic inversion, resulting in a predicted intertidal topographic map.

[0065] 2) Accuracy verification: To verify the accuracy and reliability of the method of this invention, the following example will be used for further explanation. This invention uses root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The accuracy of water depth inversion is evaluated using this method, and the intertidal elevation inversion effect under this strategy is analyzed. RMSE and MAE are used to evaluate the error between the inversion results and observed values; the smaller these values, the better the inversion effect. 2 Also known as the fit index, R is commonly used to describe the degree of fit between the inversion model and the observed values.2 The range is [0, 1]. The closer the value is to 1, the more consistent the inversion result is with the true distribution of the observed values. That is, by comparing the measured elevation data with the intertidal zone predicted topographic map, the accuracy of water depth inversion is evaluated by the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²), and the effect of this strategy in intertidal zone elevation inversion is analyzed.

[0066] ; ; .

[0067] The terrain calculated by the model versus the high-precision terrain obtained from UAV LiDAR measurements ( Figure 6 A comparison was made with the actual values ​​(as shown in the figure). Quantitative evaluation results show that the DEM (Digital Elevation Model) constructed by the sequence regression model has an RMSE of 0.202m and a MAE of 0.330m compared to the true values. 2 The value is 0.835, and the result is as follows: Figure 7 As shown, n=1682. Through visual interpretation, the DEM generated by the model has good consistency with the high-precision topographic reference data obtained by LiDAR photogrammetry. These DEMs have a spatial resolution of 10m, which can well display the details of the intertidal topography. In addition, in order to verify the reliability and stability of the sequential regression model in constructing the intertidal DEM, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Recurrent Neural Network (RNN) were selected to extract the intertidal topography from cloudless Sentinel-2 images (January 2020 to May 2020) under the same experimental environment. The RMSE, MAE, and R² of different methods are shown in Table 2 below.

[0068] Table 2. Accuracy Evaluation of Different Methods As shown in Table 2, the sequence regression model based on multi-source information fusion proposed in this invention outperforms traditional RNN, GRU, and LSTM models in terms of accuracy. Compared to the best RNN method, RMSE is reduced by approximately 11.01%, MAE by approximately 3.51%, and R² is increased to 0.835, demonstrating stronger fitting ability and error control. Furthermore, this method effectively enhances the model's ability to identify key features by introducing an attention mechanism, making the generated DEM more closely resemble the real terrain in terms of spatial structure and elevation distribution. It possesses good adaptability and is suitable for automated extraction tasks of complex intertidal terrain.

[0069] This invention provides an innovative solution to overcome existing technological bottlenecks by introducing deep learning technology. Deep learning sequence regression models, which have developed rapidly in recent years, demonstrate significant advantages in handling the complex nonlinear relationship between intertidal remote sensing imagery and elevation modeling through the adaptive feature extraction capabilities of multi-layer neural networks. Compared to the reliance on linear assumptions in traditional regression models, these models can automatically capture spatial patterns in multispectral images through convolutional layers, while maintaining temporal sensitivity to tidal cycle dynamics through recurrent neural network structures. Architectures represented by Long Short-Term Memory (LSTM) networks effectively solve the challenge of modeling long-term dependencies in intertidal data through a triple gating mechanism of forget gates, input gates, and output gates. A method for extracting shoreline topography from remote sensing images is developed, addressing the most prominent cost issue in intertidal topography extraction, and automatically extracting intertidal topography based on a sequence regression deep learning model.

[0070] This invention constructs a deep learning framework integrating RNN-LSTM, and determines the optimal network parameters through end-to-end training, resulting in a mean square error of less than 0.25m between the predicted elevation and the measured data. Specifically, the Adam optimizer is used for model training, with a learning rate of 1e-3 and a batch size of 128, and the MSE loss function is applied for gradient updates.

[0071] In summary, the specific process of each step in this invention is as follows: Figure 8 As shown, Step 1 includes Sentinel-2 time-series imagery and cloud masking. Step 2 includes land-water segmentation, extraction of the highest and lowest tide levels, and cropping. Step 3 includes index calculation of pixels within the highest and lowest tide levels, mean synthesis, band selection, normalization, and mean synthesis, with further merging processing. Step 4 includes the input feature content, elevation training data collected by UAV / ICESat-2, construction of the sequence regression model, model weights, and model training. Step 5 includes outputting the intertidal zone predicted elevation map through the sequence regression model and collecting elevation validation data using UAV / ICESat-2 for accuracy verification.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for retrieving intertidal topography from optical remote sensing images based on a sequence model, characterized in that, include: Based on the cloud coverage percentage judgment result of the intertidal remote sensing image, and combined with the cloud detection algorithm, the cloud pixels in the intertidal remote sensing image are filtered to obtain the preprocessed remote sensing data. Water masks were generated from preprocessed remote sensing data based on the normalized differential water index, and the intertidal zone range was determined by time series mask overlay and morphological optimization techniques. Within the intertidal zone, multispectral indices are calculated for pixels in the preprocessed remote sensing data; the multispectral index calculation results and the data of each band of the intertidal remote sensing image are standardized and equalized to obtain model training data. Based on the model training data, and combined with the principle of ensemble learning, dual deep learning modules are trained, and an intertidal elevation prediction model is constructed. The intertidal topography inversion results are output using the intertidal elevation prediction model and compared with the measured elevation data for verification.

2. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The cloud cover percentage determination result based on the intertidal remote sensing image, combined with the cloud detection algorithm, filters cloud pixels in the intertidal remote sensing image to obtain preprocessed remote sensing data, including: Based on the scene-level cloud mask attribute in the metadata of intertidal remote sensing images, the cloud coverage percentage of the remote sensing images is calculated, and remote sensing images with a cloud coverage percentage above a preset threshold are filtered out. The cloud detection algorithm was used to perform cloud masking on the filtered intertidal remote sensing image to remove cloud pixels from the intertidal remote sensing image, thus obtaining preprocessed remote sensing data.

3. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The water body mask generated from the preprocessed remote sensing data based on the normalized differential water index, and the intertidal zone range determined using time-series mask overlay and morphological optimization techniques, include: A water mask is generated based on the normalized differential water index and differential water index threshold of each single image in the preprocessed remote sensing data. After superimposing the water masks at different time points, morphological optimization was performed to obtain the intertidal zone range.

4. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 3, characterized in that, The step of generating a water mask based on the normalized differential water index and differential water index threshold of each single image in the preprocessed remote sensing data includes: Determine the threshold for the differential water body index; normalize the green band and near-infrared band in the preprocessed remote sensing data to obtain the normalized differential water body index; When the normalized differential water index is greater than or equal to the differential water index threshold, a water mask is generated, and the water mask is used to separate water bodies and land in the preprocessed remote sensing data.

5. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The calculation of multispectral indices for pixels in the preprocessed remote sensing data within the intertidal zone includes: Using the green, red, near-infrared, and shortwave infrared bands from preprocessed remote sensing data of the intertidal zone, the normalized vegetation index, normalized water index, and improved normalized water index of each pixel were calculated.

6. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The standardization and equalization processing of the multispectral index calculation results and the intertidal remote sensing image data of each band yields the following model training data: A standardized feature sequence was constructed in the order of normalized blue light index, normalized green light index, normalized red light index, normalized infrared index, normalized shortwave infrared index, normalized vegetation index, normalized water index, improved normalized water index, and elevation. Elevation intervals are divided according to a preset step size, and standardized feature sequences are grouped. Data distribution rules are optimized using stratified sampling, and standardized feature sequence samples in each group are balanced. The standardized feature sequences after equalization are divided into model training data, including training and testing sets, according to a preset ratio.

7. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 6, characterized in that, The step of optimizing data distribution rules using stratified sampling and performing equalization processing on the standardized feature sequence samples of each group includes: If the number of standardized feature sequence samples in each group is less than the lower limit of the sample size, then the corresponding group is discarded. If the standardized feature sequence samples of each group are greater than or equal to the lower limit of the sample size and less than or equal to the upper limit of the sample size, then all samples are retained. When the number of standardized feature sequence samples in each group is greater than the upper limit of the sample size, then a number of standardized feature sequence samples equal to the upper limit of the sample size are randomly collected.

8. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The process of training dual deep learning modules based on model training data, combining ensemble learning principles, and constructing an intertidal elevation prediction model includes: Based on the principle of ensemble learning, and combined with recurrent neural network modules and long short-term memory network modules, a time-series regression model is constructed; the training data of the model is then processed through the time-series regression model. A feature stitching layer is used to connect the output of the time series regression model, and a fully connected layer is used to output the elevation prediction value. Based on the loss function and optimizer, the time series regression model is trained and optimized to obtain the intertidal zone elevation prediction model.

9. The method for retrieving intertidal topography from optical remote sensing images based on a sequence model according to claim 1, characterized in that, The process of using the intertidal elevation prediction model to output the intertidal topographic inversion results and comparing them with measured elevation data includes: Using preprocessed remote sensing data as input, the intertidal zone topography was inverted using an intertidal zone elevation prediction model, and compared with measured elevation data to verify the accuracy of the water depth inversion.

10. A method for retrieving intertidal topography from optical remote sensing images based on a sequence model, as described in claim 9, is characterized in that... When inverting the intertidal topography and comparing it with measured elevation data, an evaluation standard including root mean square error, mean absolute error, and coefficient of determination is used to verify the accuracy of the water depth inversion.

Images