Pit pond water depth inversion and water reserve estimation method based on remote sensing image super-resolution

Abstract

The invention relates to a pit pond water depth inversion and water reserve estimation method based on remote sensing image super-resolution. The method comprises the following steps: acquiring a low-resolution remote sensing image of an area to be detected; inputting the low-resolution image into a trained pit-pond water depth inversion and water reserve estimation model for super-resolution reconstruction and water depth inversion, and generating a pit-pond water depth distribution diagram of the region; based on the pond water depth distribution diagram, spatial analysis is carried out, and the water area and the water reserve of the pond are estimated. Wherein the pit-pond water depth inversion and water reserve estimation model comprises a super-resolution module based on a Swin Transform and a matrix mean discriminator, a water depth inversion module of a random forest model with a built-in spatial feature layer, and a resolution-precision coupling module used for performing end-to-end optimization on the super-resolution module. The method can solve the problem that the prior art is easy to fall into local optimum, and has the characteristics of high calculation efficiency and strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water resource management, in particular to a pit pond water depth inversion and water storage estimation method based on remote sensing image super resolution. BACKGROUND

[0002] Small water bodies such as pits and ponds are important components of water resources and ecological systems. Accurate monitoring of their water depth and volume is of great significance in agricultural irrigation, environmental protection and water resource management. Traditional field measurement methods are accurate but time-consuming and labor-intensive, making it difficult to meet the needs of large-scale and high-frequency monitoring.

[0003] Remote sensing technology provides an effective solution, but faces inherent challenges. First, the spatial resolution of many remote sensing images is limited. For complex-shaped and small-area pits and ponds, a pixel may cover a mixed water-land area, resulting in blurred water body boundaries and severely affecting the accuracy of subsequent water depth inversion. To solve this problem, the academic community has introduced super-resolution (SR) technology to improve image resolution through algorithms.

[0004] However, existing technical approaches usually treat super-resolution and water depth inversion as two independent and serial tasks. This separate processing procedure has a core defect: the goal of the super-resolution model is to generate the most visually realistic high-resolution image, but the texture and detail features of the reconstructed image may not be optimal for subsequent water depth inversion tasks based on physical or statistical models. This task disconnection often results in poor final water depth inversion accuracy, even if the super-resolution image performs well visually, causing waste of computing resources and performance bottlenecks.

[0005] In addition, traditional water depth inversion models rely heavily on pixel spectral information, ignoring the spatial morphological features of pits and ponds as independent geographical units (such as shape, boundary complexity, etc.), which limits the accuracy of the model when dealing with irregular pits and ponds. At the same time, the setting of hyperparameters in constructing the water depth inversion model (such as the random forest model) often relies on traditional methods such as grid search, which is not only computationally inefficient but also prone to local optima, affecting the final performance and robustness of the model.

[0006] Therefore, there is an urgent need for a new technical solution that can deeply couple and optimize the super-resolution reconstruction of remote sensing images with the water depth inversion task, and fully utilize the spatial features of water bodies to systematically improve the overall accuracy and automation level of pit and pond water depth inversion. SUMMARY

[0007] Therefore, it is necessary to address the problem that existing technologies are prone to getting trapped in local optima and to provide a method for pond water depth inversion and water storage estimation based on remote sensing image super-resolution, which has the characteristics of high computational efficiency and strong robustness.

[0008] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows: A model for inverting pond water depth and estimating water storage based on super-resolution of remote sensing images includes: a super-resolution module for receiving low-resolution remote sensing images and outputting reconstructed high-resolution remote sensing images, including a generator network based on SwinTransformer and a matrix mean discriminator as a discriminant network. The water depth inversion module is used to receive the reconstructed high-resolution image and output the water depth distribution of the pit; the water depth inversion module is a random forest model with a built-in spatial feature layer; The resolution-accuracy coupling module is used to construct a joint objective function based on the output accuracy of the super-resolution module and the water depth inversion module, and to perform end-to-end optimization of the super-resolution module.

[0009] Preferably, the generating network includes a shallow feature extraction unit with multi-scale dilated convolution, a deep feature extraction unit with dense residual SwinTransformer module, and an upsampling module consisting of multiple convolutional layers, activation functions, and neighbor interpolation upsampling layers.

[0010] Furthermore, the shallow feature extraction unit is used to enlarge the receptive field and capture shallow feature maps outputting multi-scale information. Specifically:

[0011] in, This is the original low-resolution image. For convolutional layers with different spread ratios, These are the corresponding weighting factors; The deep feature extraction unit of the residual Swin Transformer module is used to extract and fuse global information through residual connections and dense connections, and outputs the first... Deep feature map Specifically:

[0012] in, For the first A Swing Transformer module These are weighting coefficients. For residual connection, Number of modules; The upsampling module is used to upsample low-resolution images to the target high-resolution image. Specifically:

[0013] in, The output of the deep feature extraction module is used as the input of the upsampling module. For upsampling networks, the nearest neighbor algorithm is used to enlarge the feature map.

[0014] Furthermore, after receiving the reconstructed high-resolution image, the water depth inversion module uses a random forest model to predict the water depth for each pixel and aggregates the data to output the water depth distribution of the pits and ponds; among which... Predicted value Obtained by ensemble of multiple decision trees: in, To predict water depth, For the first The predicted values ​​of each decision tree. For its weight, This represents the total number of decision trees.

[0015] A training method for a pit / pond water depth inversion and water storage estimation model based on remote sensing image super-resolution includes the following steps: Acquire low-resolution remote sensing images of the pond to be measured, georeferenced original high-resolution remote sensing images, and measured water depth data of the pond to construct a training dataset. The low-resolution imagery is used as input to the super-resolution module of the pit and pond water depth inversion and water storage estimation model, and the super-resolution module outputs a reconstructed high-resolution imagery; the original high-resolution imagery is used as real samples for adversarial training of the discrimination network. The spectral and spatial features extracted from the reconstructed high-resolution images are paired with measured water depth data to construct an inversion training set; based on the inversion training set, the hyperparameters of the random forest model are optimized using an adaptive Bayesian optimization algorithm. The output accuracy of the super-resolution module and the water depth inversion module is evaluated. A resolution-accuracy coupling module is adopted. Based on the output accuracy of the super-resolution module and the water depth inversion module, a joint objective function is constructed. The performance of the super-resolution module and the water depth inversion module is optimized end-to-end to obtain a trained pit water depth inversion and water storage estimation model.

[0016] Preferably, the resolution-accuracy coupling module employs a joint objective function that includes super-resolution loss and water depth inversion error to perform end-to-end collaborative optimization of the two modules; the joint objective function The definition of is: in, This is the loss function of the super-resolution module itself. The root mean square error of the water depth inversion module. To balance the hyperparameters of the two tasks.

[0017] Furthermore, the discriminant network employs a matrix mean discriminant, which evaluates the local differences between the generated image and the real high-resolution image during training using the following formula: in, and These represent the generated image and the real image at different locations. Pixel value at; The adaptive Bayesian optimization algorithm is used to find the cross-validation error function. Minimal hyperparameter combination : in, It should include at least the number of decision trees and the number of features considered when splitting nodes.

[0018] A method for inverting pond water depth and estimating water storage based on super-resolution remote sensing imagery includes the following steps: acquiring low-resolution remote sensing images of the area to be measured; inputting the low-resolution images into the trained pond water depth inversion and water storage estimation model for super-resolution reconstruction and water depth inversion to generate a pond water depth distribution map of the area; and performing spatial analysis based on the pond water depth distribution map to estimate the water area and water storage of the pond.

[0019] Preferably, the method further includes a step of evaluating the accuracy of the inversion results, using the root mean square error index for evaluation.

[0020] in, To invert water depth, To measure the actual water depth, This represents the total number of verification points.

[0021] Furthermore, before performing water depth inversion, a water area extraction step is included. Water areas in the image are initially identified and delineated using a threshold segmentation method, serving as the target area for subsequent processing. The discrimination rule `water_mask` is defined as follows:

[0022] in, For pixels Image intensity value at that location, The preset water area discrimination threshold is used.

[0023] The beneficial effects of this invention are as follows: This invention presents a pit and pond water depth inversion and water storage estimation model based on remote sensing image super-resolution. The model comprises a super-resolution module based on Swing Transformer and matrix mean discriminator, a water depth inversion module using a random forest model with a built-in spatial feature layer, and a resolution-accuracy coupling module for end-to-end optimization of the super-resolution module. This invention combines the powerful global information extraction capabilities of Swing Transformer with the advantages of Generative Adversarial Networks (GANs) in detail restoration. By incorporating a spatial feature layer into the water depth inversion module, this invention no longer relies solely on the spectral values ​​of a single pixel but also considers the spatial morphological information of the pit and pond, such as its shape and boundaries. This allows the model to better understand the geographical characteristics of the water body itself, exhibiting higher accuracy and robustness when dealing with irregular, small, and micro-sized pits and ponds. Furthermore, this invention uses the final water depth inversion error as a feedback signal to guide and optimize the upstream super-resolution reconstruction process, ensuring that every effort in image reconstruction directly serves the ultimate goal of improving water depth inversion accuracy, thus solving the problem of existing technologies easily getting trapped in local optima. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a model for inverting pond water depth and estimating water storage based on super-resolution of remote sensing images. Figure 2 This is a flowchart illustrating a training method for a pit and pond water depth inversion and water storage estimation model based on remote sensing image super-resolution. Figure 3 This is a flowchart illustrating a method for inverting pond water depth and estimating water storage based on super-resolution remote sensing images. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] Example 1 like Figure 1 As shown, a model for pond water depth inversion and water storage estimation based on remote sensing image super-resolution includes: a super-resolution module for receiving low-resolution remote sensing images and outputting reconstructed high-resolution remote sensing images, including a generator network based on Swin Transformer and a matrix mean discriminator as a discriminant network. The water depth inversion module is used to receive the reconstructed high-resolution image and output the water depth distribution of the pit; the water depth inversion module is a random forest model with a built-in spatial feature layer; The resolution-accuracy coupling module is used to construct a joint objective function based on the output accuracy of the super-resolution module and the water depth inversion module, and to perform end-to-end optimization of the super-resolution module.

[0027] In one specific embodiment, the generative network includes a shallow feature extraction unit with multi-scale dilated convolution, a deep feature extraction unit with dense residual Swin Transformer module, and an upsampling module consisting of multiple convolutional layers, activation functions, and neighbor interpolation upsampling layers.

[0028] In one specific embodiment, the shallow feature extraction unit is used to enlarge the receptive field and capture shallow feature maps output by multi-scale information. Specifically:

[0029] in, This is the original low-resolution image. For convolutional layers with different spread ratios, These are the corresponding weighting factors; The deep feature extraction unit of the residual Swin Transformer module is used to extract and fuse global information through residual connections and dense connections, and outputs the first... Deep feature map Specifically:

[0030] in, For the first A Swing Transformer module These are weighting coefficients. For residual connection, Number of modules; The upsampling module is used to upsample low-resolution images to the target high-resolution image. Specifically:

[0031] in, The output of the deep feature extraction module is used as the input of the upsampling module. For upsampling networks, the nearest neighbor algorithm is used to enlarge the feature map.

[0032] In one specific embodiment, after receiving the reconstructed high-resolution image, the water depth inversion module uses a random forest model to predict the water depth for each pixel and outputs the water depth distribution of the pit / pond by aggregating the data from each pixel; where the points are... Predicted value Obtained by ensemble of multiple decision trees: in, To predict water depth, For the first The predicted values ​​of each decision tree. For its weight, This represents the total number of decision trees.

[0033] Example 2 like Figure 2 As shown, a training method for a pit / pond water depth inversion and water storage estimation model based on remote sensing image super-resolution includes the following steps: Acquire low-resolution remote sensing images of the pond to be measured, georeferenced original high-resolution remote sensing images, and measured water depth data of the pond to construct a training dataset. The low-resolution imagery is used as input to the super-resolution module of the pit and pond water depth inversion and water storage estimation model, and the super-resolution module outputs a reconstructed high-resolution imagery; the original high-resolution imagery is used as real samples for adversarial training of the discrimination network. The spectral and spatial features extracted from the reconstructed high-resolution images are paired with measured water depth data to construct an inversion training set; based on the inversion training set, the hyperparameters of the random forest model are optimized using an adaptive Bayesian optimization algorithm. The output accuracy of the super-resolution module and the water depth inversion module is evaluated. A resolution-accuracy coupling module is adopted. Based on the output accuracy of the super-resolution module and the water depth inversion module, a joint objective function is constructed. The performance of the super-resolution module and the water depth inversion module is optimized end-to-end to obtain a trained pit water depth inversion and water storage estimation model.

[0034] In one specific embodiment, the resolution-accuracy coupling module employs a joint objective function that includes super-resolution loss and water depth inversion error to perform end-to-end collaborative optimization of the two modules; the joint objective function The definition of is: in, This is the loss function of the super-resolution module itself. The root mean square error of the water depth inversion module. To balance the hyperparameters of the two tasks.

[0035] In one specific embodiment, the discriminant network employs a matrix mean discriminant, which evaluates the local differences between the generated image and the real high-resolution image during training using the following formula: in, and These represent the generated image and the real image at different locations. Pixel value at; The adaptive Bayesian optimization algorithm is used to find the cross-validation error function. Minimal hyperparameter combination : in, It should include at least the number of decision trees and the number of features considered when splitting nodes.

[0036] In this embodiment, the present invention employs an adaptive Bayesian optimization algorithm to adjust the key hyperparameters of the random forest model. Compared with traditional methods such as grid search, Bayesian optimization can find a better combination of parameters with fewer iterations, which not only significantly saves computational resources and time, but also effectively reduces the risk of model overfitting, resulting in an inversion model with better performance and stronger generalization ability.

[0037] Example 3 like Figure 3 As shown, a method for pond water depth inversion and water storage estimation based on remote sensing image super-resolution includes the following steps: acquiring low-resolution remote sensing images of the area to be measured; inputting the low-resolution images into the trained pond water depth inversion and water storage estimation model for super-resolution reconstruction and water depth inversion to generate a pond water depth distribution map of the area; visualizing the water depth inversion results through a Geographic Information System (GIS) platform, and estimating the water area by combining the water depth data. Through spatial analysis, the water area change is calculated, and further combined with water depth data, water storage is estimated, providing a basis for water resource management and ecological protection. The models for inverting water depth and estimating water storage in ponds are updated regularly, and new super-resolution remote sensing image data are used to train the models to adapt to changes in water depth in ponds under different seasons or environmental conditions.

[0038] In one specific embodiment, the method further includes a step of evaluating the accuracy of the inversion results, using the root mean square error index for evaluation:

[0039] in, To invert water depth, To measure the actual water depth, This represents the total number of verification points.

[0040] After obtaining the error index, the inversion accuracy is evaluated by comparing it with the validation data, and the water area extraction method and the interference of algae on the water surface are optimized based on the error analysis.

[0041] In one specific embodiment, before performing water depth inversion, a water area extraction step is included. This step uses a threshold segmentation method to initially identify and delineate water areas in the image, which serve as the target range for subsequent processing. The discrimination rule `water_mask` is defined as follows:

[0042] in, For pixels Image intensity value at that location, The preset water area discrimination threshold is used.

[0043] In this embodiment, in order to verify the actual performance of the model and method proposed in this invention and to specifically illustrate its beneficial effects, this part takes the pond water body in a certain agricultural and mining complex area of ​​the province as the research object and conducts a complete numerical simulation and comparative experiment.

[0044] Data preparation and preprocessing Remote sensing imagery: Imagery from the domestically produced Gaofen-1 (GF-1) satellite covering the study area was acquired. Specifically, 8-meter resolution multispectral imagery was selected as low-resolution (LR) input data, while 2-meter resolution panchromatic imagery was fused with the multispectral imagery to serve as the high-resolution (HR) baseline data.

[0045] Measured data: Within the study area, 200 representative ponds were selected. Using handheld GPS and portable depth sounders, sampling points were set up in each pond, collecting a total of 3,000 measured water depth samples.

[0046] Dataset construction: 200 ponds and their data were randomly divided into a training set (160 ponds, 2400 sample points) and a validation set (40 ponds, 600 sample points) at a ratio of 4:1.

[0047] Experimental Design in this Example To fully demonstrate the technical advantages of this invention, two experimental groups were set up for comparison: Group A (Embodiments of the Invention): The technical solution fully adopts the technical solution defined in the claims of this patent.

[0048] Model structure: The complete model described in claim 1 is adopted, namely, the Swin Transformer-GAN super-resolution module + random forest inversion module with built-in spatial feature layer + resolution-accuracy coupling module.

[0049] Training method: The training method described in claim 5 is adopted, that is, through a joint objective function. End-to-end collaborative optimization is performed, and an adaptive Bayesian optimization algorithm is used to optimize the hyperparameters of the random forest.

[0050] Group B (Comparative Example: Traditional Separate Approach): Simulates the existing mainstream technical path.

[0051] Model Structure: The super-resolution module and the depth inversion module are independent of each other. The training objective of the super-resolution module (also using SwinTransformer-GAN) is to optimize traditional image quality metrics (such as PSNR), rather than the final inversion accuracy. The depth inversion module uses a standard random forest model and does not contain a spatial feature layer.

[0052] Training method: The process is divided into two steps. First, the super-resolution model is trained independently. Second, the images output by the trained super-resolution model are used to train the water depth inversion model. The hyperparameter tuning of the random forest is performed using the traditional grid search method.

[0053] In this embodiment, training and validation were performed on two groups of experiments, A and B, under the same computing environment. The results are shown in the table below: Table 1: Performance Comparison of Embodiments and Comparative Examples of the Invention

[0054] As shown in Table 1, the overall RMSE (0.15 m) and MAE (0.11 m) of the embodiments of the present invention (Group A) are significantly lower than those of the comparative examples (Group B), which are 0.32 m and 0.25 m, respectively, representing an accuracy improvement of more than double. This fully demonstrates the effectiveness of the "resolution-accuracy coupling module" introduced in this invention. By using the final inversion error as a feedback signal to guide the optimization of the super-resolution module, the reconstructed image features generated are made more "friendly" to the water depth inversion task, thereby fundamentally solving the performance bottleneck caused by the task disconnect in traditional methods.

[0055] When individually evaluating 13 irregularly shaped and tortuous ponds in the validation set, the advantages of this invention (Group A) were more pronounced, with its RMSE (0.18 meters) being only 40% of that of the comparative example (Group B, 0.45 meters). This clearly demonstrates that the "spatial feature layer" built into the Group A model played a crucial role. By extracting and utilizing the spatial geometric information such as the shape and boundaries of the ponds, the model can better understand the inherent characteristics of complex water bodies, thereby making more accurate judgments in areas where spectral information is easily confused.

[0056] In this embodiment, Group A, employing adaptive Bayesian optimization, found a high-performance hyperparameter combination in approximately 2.5 hours. In contrast, Group B, using traditional grid search, took approximately 18 hours to traverse all preset parameter combinations, incurring significant computational costs. This demonstrates that the training method employed in this invention not only improves efficiency by approximately 7 times but also finds a better-performing model through a more intelligent search strategy, effectively enhancing the cost-effectiveness of training and the final performance of the model.

[0057] In this embodiment, once the model in group A is trained, it can automate the entire process from inputting low-resolution imagery to outputting the final water depth distribution map and water storage estimation. Users do not need to worry about the complex intermediate processes, demonstrating the enormous application potential of this invention as an integrated, efficient, and systematic solution.

Claims

1. A model for pond water depth inversion and water storage estimation based on remote sensing image super-resolution, characterized in that, include: The super-resolution module is used to receive low-resolution remote sensing images and output reconstructed high-resolution remote sensing images. It includes a generator network based on Swing Transformer and a matrix mean discriminator as a discriminant network. The water depth inversion module is used to receive the reconstructed high-resolution image and output the water depth distribution of the pit; the water depth inversion module is a random forest model with a built-in spatial feature layer; The resolution-accuracy coupling module is used to construct a joint objective function based on the output accuracy of the super-resolution module and the water depth inversion module, and to perform end-to-end optimization of the super-resolution module.

2. The pit / pond water depth inversion and water storage estimation model according to claim 1, characterized in that, The generative network includes a shallow feature extraction unit with multi-scale dilated convolution, a deep feature extraction unit with dense residual Swing Transformer module, and an upsampling module consisting of multiple convolutional layers, activation functions, and neighbor interpolation upsampling layers.

3. The pit / pond water depth inversion and water storage estimation model according to claim 2, characterized in that, The shallow feature extraction unit is used to enlarge the receptive field and capture shallow feature maps output by multi-scale information. Specifically: in, This is the original low-resolution image. For convolutional layers with different spread ratios, These are the corresponding weighting factors; The deep feature extraction unit of the residual Swin Transformer module is used to extract and fuse global information through residual connections and dense connections, and outputs the first... Deep feature map Specifically: in, For the first A Swing Transformer module These are weighting coefficients. For residual connection, Number of modules; The upsampling module is used to upsample low-resolution images to the target high-resolution image. Specifically: in, The output of the deep feature extraction module is used as the input of the upsampling module. For upsampling networks, the nearest neighbor algorithm is used to enlarge the feature map.

4. The pit / pond water depth inversion and water storage estimation model according to claim 1, characterized in that, After receiving the reconstructed high-resolution image, the water depth inversion module uses a random forest model to predict the water depth for each pixel and outputs the water depth distribution of the pit / pond by aggregating the data from each pixel; among which... Predicted value Obtained by ensemble of multiple decision trees: in, To predict water depth, For the first The predicted values ​​of each decision tree. For its weight, This represents the total number of decision trees.

5. A training method for a pit / pond water depth inversion and water storage estimation model based on remote sensing image super-resolution, characterized in that, Includes the following steps: Acquire low-resolution remote sensing images of the pond to be measured, georeferenced original high-resolution remote sensing images, and measured water depth data of the pond to construct a training dataset. The low-resolution imagery is used as input to the super-resolution module of the pit and pond water depth inversion and water storage estimation model, and the super-resolution module outputs a reconstructed high-resolution imagery; the original high-resolution imagery is used as real samples for adversarial training of the discrimination network. The spectral and spatial features extracted from the reconstructed high-resolution images are paired with measured water depth data to construct an inversion training set; Based on the inversion training set, the hyperparameters of the random forest model are optimized using an adaptive Bayesian optimization algorithm. The output accuracy of the super-resolution module and the water depth inversion module is evaluated. A resolution-accuracy coupling module is adopted. Based on the output accuracy of the super-resolution module and the water depth inversion module, a joint objective function is constructed. The performance of the super-resolution module and the water depth inversion module is optimized end-to-end to obtain a trained pit water depth inversion and water storage estimation model.

6. The training method for the pit / pond water depth inversion and water storage estimation model according to claim 5, characterized in that, The resolution-accuracy coupling module employs a joint objective function that includes super-resolution loss and water depth inversion error to perform end-to-end collaborative optimization of the two modules; the joint objective function The definition of is: in, This is the loss function of the super-resolution module itself. The root mean square error of the water depth inversion module. To balance the hyperparameters of the two tasks.

7. The training method for the pit / pond water depth inversion and water storage estimation model according to claim 5, characterized in that, The discriminant network employs a matrix mean discriminant, which evaluates the local differences between the generated image and the real high-resolution image during training using the following formula: in, and These represent the generated image and the real image at different locations. Pixel value at; The adaptive Bayesian optimization algorithm is used to find the cross-validation error function. Minimal hyperparameter combination : in, It should include at least the number of decision trees and the number of features considered when splitting nodes.

8. A method for inverting pond water depth and estimating water storage based on remote sensing image super-resolution, characterized in that, Includes the following steps: Acquire low-resolution remote sensing images of the area to be measured; input the low-resolution images into the trained pit and pond water depth inversion and water storage estimation model for super-resolution reconstruction and water depth inversion to generate a pit and pond water depth distribution map of the area; based on the pit and pond water depth distribution map, perform spatial analysis to estimate the water area and water storage of the pit and pond.

9. The method for pond water depth inversion and water storage estimation based on remote sensing image super-resolution as described in claim 8, characterized in that, It also includes a step of evaluating the accuracy of the inversion results, using the root mean square error (RMSE) index for evaluation. in, To invert water depth, To measure the actual water depth, This represents the total number of verification points.

10. The method for pond water depth inversion and water storage estimation based on remote sensing image super-resolution as described in claim 8, characterized in that, Before performing water depth inversion, a water area extraction step is included. This step uses a threshold segmentation method to initially identify and delineate water areas in the image, which will serve as the target area for subsequent processing. The discrimination rule `water_mask` is defined as follows: in, For pixels Image intensity value at that location, The preset water area discrimination threshold is used.

Images