A water body and geometric parameter monitoring method based on an AESRGAN remote sensing super-resolution model

Abstract

In view of the problem that resolution and monitoring frequency cannot be considered simultaneously in current water body accurate monitoring, the application provides a water body and geometric parameter monitoring method based on an AESRGAN remote sensing super-resolution model. The application constructs a registered super-resolution data set, and proposes an AESRGAN model, introduces an attention mechanism in the generator to strengthen the spatial texture reconstruction capability, and combines a spectral consistency and structural similarity loss function to maintain the spectral fidelity and geometric boundary integrity of the remote sensing image. On this basis, water body extraction and river geometric parameter monitoring are performed. The AESRGAN model constructed based on the relationship of real remote sensing images can significantly improve the spatial geometric accuracy while maintaining the spectral consistency, and the application provides a generalizable technical path for river form and flow monitoring in regions lacking high-resolution images.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy and hydrology technology, especially the technology of dynamic monitoring of surface water by remote sensing, specifically a method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model. Background Technology

[0002] Climate change is causing unprecedented changes in surface water bodies, and accurate monitoring of their spatiotemporal dynamics is crucial for understanding hydrological processes and assessing climate impacts. Remote sensing technology, due to its ability to enable continuous, wide-area, and non-contact observation of surface water bodies, has become an indispensable tool for monitoring water dynamics.

[0003] Most current research relies on publicly available satellite imagery (such as Landsat and Sentinel-2). While these images offer good temporal and spatial resolution, they can only resolve relatively wide rivers and their significant changes over time. Mixed pixel effects often blur water body boundaries, thus affecting monitoring accuracy. High-resolution commercial satellite imagery (such as GF-2 and WordView) can depict spatial details of small to medium-sized rivers, but this usually comes at the cost of observation frequency and spatial coverage, and the imagery acquisition areas are often concentrated in areas of high economic activity. Many researchers have attempted to balance monitoring frequency and resolution by fusing multi-resolution remote sensing imagery; however, cross-sensor heterogeneity leads to a mismatch between spectral response functions and spatial resolution, easily causing spectral distortion and poor data consistency, thus limiting the monitoring accuracy of water bodies and their geometric parameters.

[0004] Super-resolution (SR) technology offers a new approach to overcoming spatial resolution differences caused by satellite imagery. This technique reconstructs high-resolution (HR) images from low-resolution (LR) images. Traditional SR methods include interpolation-based, transform-domain-based, and reconstruction-based methods. In recent years, deep learning methods have demonstrated superior feature representation and reconstruction capabilities, becoming a research hotspot in the SR field. Some researchers have explored using Single Image Super-Resolution (SISR) methods to directly improve the spatial resolution of images by aligning them with high-resolution commercial images. With the development of Generative Adversarial Networks (GANs), their advantages in feature generation and high-frequency detail reconstruction have made adversarial learning-based SR models (GAN-based SR) one of the core directions of current SISR research. Generally, GAN-based models consist of a generator and a discriminator: the former reconstructs LR images into HR images, and the latter determines whether the reconstructed image is a true HR.

[0005] However, achieving super-resolution for all bands of a multispectral image directly is challenging, for example: (1) the number of bands between S2 images and commercial images is mismatched, such as S2 images having 13 bands, while commercial satellite images mostly have only 4 or at most 8 bands; (2) the original resolution of different bands of multispectral images is often inconsistent (e.g., S2 visible band has a resolution of 10m, while other spectral bands have a resolution of 20m or 60m); (3) there are differences in observation time, orbit orientation and geometric correction among multi-source images, which makes registration and joint modeling difficult; (4) implementing super-resolution for all bands will significantly increase the amount of data and computational burden.

[0006] To address the aforementioned issues, some studies have proposed achieving super-resolution directly at the semantic level. This involves training deep learning models to predict high-resolution classification results on original-resolution images, thereby skipping the image reconstruction process. This "semantic-level super-resolution" method can alleviate the spectral heterogeneity problem to some extent, but its end-to-end nature makes it impossible to intuitively analyze classification features from the perspectives of spectral mechanisms, geometric mechanisms, and other aspects.

[0007] It is of practical significance to study how to improve the accuracy and stability of water body and geometric parameter extraction from a mechanistic perspective by utilizing the spectral and spatial features of freely available satellite imagery (such as Landsat and Sentinel-2). For example, regarding S2 imagery, although a 10-meter resolution S2 image can be used to depict a river approximately 30 meters wide, the uncertainty of the riverbank boundary can reach 20 meters due to the mixed pixel effect, which is unacceptable. Therefore, higher image resolution is needed to eliminate the pixel scale differences that exist when S2 imagery is fused with high-resolution satellite imagery, and to map smaller rivers.

[0008] ESRGAN (Enhanced Super-Resolution Generative Adversarial Network), an advanced super-resolution model based on Generative Adversarial Networks (GANs), effectively improves the detail restoration and visual naturalness of generated images by introducing a residual-in-residual dense block (RRDB) structure, demonstrating excellent performance in super-resolution tasks using natural images. However, remote sensing images differ significantly from natural images in terms of spectral dimension and the complexity of land cover categories. Directly applying the standard ESRGAN model often leads to problems such as spectral distortion or blurred land cover boundaries. Improving this model may provide a generalizable technical approach for river morphology and flow monitoring in areas lacking high-resolution imagery. Summary of the Invention

[0009] Due to the mixed pixel effect, publicly available imagery struggles to accurately capture details of small rivers and shorelines. While high-resolution commercial imagery offers high spatial resolution, it is limited by monitoring frequency and coverage. Current precise water body monitoring faces the challenge of balancing resolution and monitoring frequency. This invention aims to provide a water body geometric parameter monitoring method based on a remote sensing super-resolution model, achieving high-fidelity super-resolution reconstruction of remote sensing images to improve the spatial accuracy and geometric consistency of water body remote sensing monitoring. To achieve this objective, this invention designs and proposes an improved model—AESRGAN—based on the ESRGAN model structure, combined with the spectral consistency requirements of remote sensing images and ground feature information constraints. The monitoring of water body geometric parameters based on this model includes the following steps: S1. Collect satellite images of the study area: The satellite images include low-resolution (LR) images and high-resolution (HR) images; S2, Data Pairing: Construct a dataset consisting of HR and LR image pairings for subsequent training, validation and testing of the AESRGAN remote sensing super-resolution model; S3. Constructing the AESRGAN remote sensing super-resolution model: The overall structure of the generator network of the super-resolution model includes a shallow feature extraction module, a deep feature learning module, an attention enhancement module, and an upsampling reconstruction module; the deep feature learning module consists of multiple attention residual dense blocks ARRDB; the ARRDB introduces the CBAM attention mechanism after the residual of the residual-residual dense block RRDB; the RRDB is composed of multiple residual dense blocks RDB cascaded, each RDB containing 4-5 convolutional layers and ReLU activation functions; the CBAM consists of a channel attention module CAM and a spatial attention module SAM, which act on the channel dimension and spatial dimension of the feature map in turn; S4. Use the dataset from S2 to train, validate, and test the AESRGAN remote sensing super-resolution model; S5. LR images are obtained by processing them using the AESRGAN remote sensing super-resolution model; S6. Water body extraction and water body geometric parameter monitoring based on SR.

[0010] Further optimizations were made, with the high-resolution imagery HR in S1 originating from the Gaofen-2 GF2 satellite and the low-resolution imagery LR originating from the Sentinel-2 satellite.

[0011] Furthermore, in S2: data pairing includes radiometric calibration, atmospheric correction, geometric correction, coordinate reprojection, resolution resampling matching, and spatial registration.

[0012] Furthermore, in S3: the shallow feature extraction module includes a 3×3 Conv standard convolutional layer; the deep feature learning module consists of at least 20 attention residual dense blocks ARRDB; the RRDB is composed of at least 3 residual dense blocks RDB cascaded together; the attention enhancement module includes channel attention and spatial attention; the upsampling reconstruction module consists of a subpixel convolutional layer PixelShuffle and two 3×3 Conv standard convolutional layers.

[0013] Furthermore, in S3: the generator is designed with loss functions based on spectral consistency and structural similarity, as follows:

[0014] in, L G The total loss of the generator. L pix For pixel loss; L per To perceive loss; L gan To combat loss in generators; L spe For spectral loss; L ssim For structural similarity loss; λ pix , λ per、 λ gan , λ spe , λ ssim These are the weighting coefficients. Spectral loss primarily maintains the fidelity of the image's spectral information by measuring the difference in spectral vectors between the generated and real images, thus preventing color distortion or spectral aberration. Structural similarity loss evaluates the similarity of images in terms of brightness, contrast, and structure, preventing structural blurring or distortion in the generated image. The inclusion of these two metrics ensures both the spectral fidelity of the remote sensing image and the clarity of structural details, thereby improving the overall quality of the generated image.

[0015] Furthermore, S5 includes the following steps: S51. The input LR image is processed through the shallow feature extraction module for feature extraction; the shallow feature extraction formula is as follows: F0 = Conv1(I LR ) Among them: I LR This represents a low-resolution input image, and F0 represents the extracted shallow features. S52. Shallow feature F0 is input into the deep feature extraction module for deep feature extraction. The extraction formula is as follows: F deep =ARDBN (ARDB N−1 (...ARDB1(F0)...)) +F0 Among them: ARDB N F represents the Nth ARRDB. deep Indicates deep features; S53. After deep feature extraction, a 3×3 Conv convolutional layer is passed through to transition to the attention enhancement module for further feature calibration. S54. The feature map after passing through the attention module is then subjected to resolution enhancement and spatial reconstruction by the upsampling reconstruction module, as shown in the following formula: F sr = Conv(PixelShuffle(Conv(F deep ))) Where PixelShuffle represents a subpixel convolutional layer, F sr For SR images; S55. The upsampled feature map is finally adjusted and mapped through two convolutional layers, and the adjusted high-resolution image ASR is output.

[0016] Furthermore, in S3: the CBAM consists of a channel attention module (CAM) and a spatial attention module (SAM), which act sequentially on the channel dimension and spatial dimension of the feature map; specifically: the channel attention module (CAM) generates channel weight vectors through global average pooling and max pooling operations to highlight important spectral information; the spatial attention module (SAM) generates a spatial weight map through convolution operations to emphasize the boundaries of ground features and texture details; The Channel Attention Module (CAM) aims to adaptively assign higher weights to important spectral channels by capturing the interdependencies between them. Mc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) Where F is the input feature map; AvgPoo and MaxPool represent global average pooling and max pooling operations, respectively; MLP represents a multilayer perceptron; and σ represents the sigmoid activation function.

[0017] The spatial attention module is used to model location dependencies in the spatial dimension of the feature map, enhancing the model's response to key feature regions. Ms(F) = σ(f 7×7 ([AvgPool(F);MaxPool(F)])) Among them, f 7×7represents a 7×7 convolution operation; [·;·] represents a concatenation operation along the channel dimension; AvgPool(F) and MaxPool(F) represent average and max pooling operations along the channel dimension, respectively; σ represents the Sigmoid activation function.

[0018] Furthermore, the water extraction method in step S6 is the U-Net++ method.

[0019] The advantages and beneficial effects of this invention are: This invention improves the AESRGAN model by introducing an attention mechanism into the generator to enhance spatial feature learning capabilities. Furthermore, it designs spectral consistency and structural similarity loss functions based on remote sensing image features to maintain the spectral fidelity and geometric boundary integrity of the remote sensing images. Based on this model, the method of this invention can fully utilize the rich spectral and spatial structural information contained in high-resolution images, and enhance the accuracy and stability of water body extraction by improving the spatial resolution of remote sensing images. The AESRGAN model constructed based on real remote sensing image relationships can significantly improve spatial geometric accuracy while maintaining spectral consistency, providing an efficient and feasible technical path for the fusion application of multi-source satellite imagery. This method has significant scientific value and application potential in the accurate identification, dynamic monitoring, and quantitative extraction of geoscientific elements of water bodies.

[0020] This method aims to acquire high-quality river geometry information at a spatial resolution of 3m using a super-resolution model based on Sentinel-2 imagery. This invention provides a generalizable technical approach for river morphology and flow monitoring in areas lacking high-resolution imagery. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is the generator structure for the AESRGAN model.

[0023] Figure 2 Details of the Attention Residual-in-Residual Dense Block.

[0024] Figure 3 The U-Net architecture features spectrum normalization.

[0025] Figure 4 A comparison of image reconstruction results using different super-resolution models is presented. Super-resolution results for reservoirs (AR), 330 m wide rivers (BW), 158 m wide rivers (RW), 67 m wide rivers (DW), 32 m wide rivers (EW), and 16 m wide rivers (FW) are shown.

[0026] Figure 5. Statistical chart of spectral information of different images.

[0027] Figure 6 A comparison of extraction results from GF2, S2 and SR images using different water body monitoring methods.

[0028] Figure 7 These are the results of river geometry parameter extraction. Pink represents the GF2 geometry parameter extraction result, red represents the S2 geometry parameter extraction result, and orange represents the SR geometry parameter extraction result. The size of the graph represents the area.

[0029] Figure 8 shows the results of the time series river width extraction. It includes the Daqing River (A), Luan River (B), Majia River (C), Nanyun River (D), Tuhai River (E), Yongding River (F), and Ziya River (G).

[0030] Figure 9 The monitoring results of small and medium-sized rivers are presented in different images.

[0031] Figure 10 shows the monitoring accuracy of water body geometric parameters. Detailed Implementation Example

[0032] A method for monitoring water geometry parameters based on the AESRGAN remote sensing super-resolution model includes the following steps: S1. Collect satellite images of the study area: The satellite images include low-resolution images (LR) and high-resolution images (HR); in this embodiment, the high-resolution images (HR) are sourced from the Gaofen-2 GF2 satellite, and the low-resolution images (LR) are sourced from the Sentinel-2 (S2) satellite.

[0033] This embodiment selects the Haihe River Basin as the study area. The Haihe River Basin is mainly composed of small and medium-sized rivers, characterized by small drainage area, steep gradient, and complex river system distribution. These types of rivers rely more on high-resolution remote sensing imagery for accurate spatial characterization, thus providing an ideal experimental setting for the verification of super-resolution technology.

[0034] The remote sensing imagery used in this embodiment comprises two parts: one part is used to construct the training dataset for the super-resolution model, and the other part is used for the temporal extraction and variation analysis of river geometric parameters.

[0035] To construct high-resolution training (HR) images, 331 scenes of Gaofen-2 (GF2) multispectral images from March 28 to November 29, 2024 (non-glacial period) were downloaded from the China Resources Satellite Application Center. Considering the research focus on surface water monitoring, the selected images mainly cover the major channels of the Haihe River Basin and their surrounding areas, encompassing diverse river morphologies, slopes, flow characteristics, and natural and social environmental conditions. The GF2 satellite was successfully launched on August 19, 2014, carrying a set of optical sensors, including one panchromatic band and four multispectral bands (blue, green, red, and near-infrared, RGBN). The multispectral band has a spatial resolution of 3.2 m, the panchromatic band has a resolution of 0.8 m, the sensor revisit period is approximately 5 days, and the combined imaging width of the two cameras is approximately 45 km.

[0036] During the data preprocessing stage, the GF-2 multispectral imagery underwent radiometric calibration, atmospheric correction, and orthorectification. The panchromatic band, due to its stable radiometric characteristics, did not require atmospheric correction. Subsequently, the preprocessed multispectral and panchromatic band images were fused to generate a high spatial resolution reference image, which was used for geometric registration of the subsequent Sentinel-2 images.

[0037] Compared to GF-2's four multispectral bands, the S2 satellite provides 13 spectral bands, covering a broad spectrum from visible light to shortwave infrared. In this implementation, Sentinel-2A / B (S2A / B) images corresponding to the spatial locations of the GF-2 images were downloaded from the Google Earth Engine (GEE) platform, and four bands—blue, green, red, and near-infrared (RGBN)—were selected to ensure spectral consistency. Sentinel-2A and Sentinel-2B were launched in 2015 and 2017, respectively, with a spatial resolution of 10 m and a joint revisit period of 5 days. To minimize the impact of temporal differences on surface changes, we selected S2 images (some of which may contain clouds) with imaging dates no more than ±5 days from the GF-2 images for pairing.

[0038] In addition to the training data, this implementation method also downloaded and processed temporal remote sensing images of major river locations in the Haihe River Basin from 2020 to 2024, including 135 GF2 images and 1529 S2 images. These were used for subsequent dynamic monitoring of river water geometric parameters. Unlike the training data, the GF2 and S2 images here are not paired one-to-one, but rather selected from all images that meet the criteria based on the spatial range and time window of the study locations. All images underwent uniform radiometric calibration, atmospheric correction, and orthorectification according to the above process to ensure comparability and consistency among multi-temporal and multi-source data.

[0039] S2, Data Pairing: Construct a dataset consisting of HR and LR image pairings for subsequent training, validation and testing of the AESRGAN remote sensing super-resolution model; This implementation method constructs paired samples based on S2 and GF-2 multispectral images to ensure that the model can learn the scale mapping relationship and spatial detail features between multi-source images.

[0040] Because GF-2 and S2 images differ in pixel size, geometric projection, and spatial reference frame, high-precision pairing requires steps such as resolution matching, coordinate reprojection, and spatial registration.

[0041] First, the spatial intersection region of the GF-2 and S2 images was extracted as the pairing range. Measurements revealed that the original downloaded S2 pixel side length was approximately 8.9 m, while the GF-2 pixel side length was approximately 3.2 m. To ensure the integer multiple relationship of pixel scale, bicubic interpolation was used to uniformly resample the S2 image to 9 m resolution, while simultaneously resampling the GF-2 image to 3 m resolution. Subsequently, using the UTM projection of the GF-2 image as a reference, the S2 image projection was transformed to the same coordinate system, and the fused 1 m resolution GF-2 panchromatic image was used as the geometric registration reference to ensure precise consistency in the spatial position of the paired images.

[0042] To maintain spectral consistency across bands, this implementation method constructs a corresponding four-band super-resolution model only for the GF-2 multispectral bands (R, G, B, N). The band ranges of the GF-2 image are as follows: blue band (0.45~0.52μm), green band (0.52~0.59μm), red band (0.63~0.69μm), and near-infrared band (0.77~0.89μm).

[0043] S3. Constructing the AESRGAN remote sensing super-resolution model: like Figure 1 As shown in this embodiment, The overall structure of the generator network in the S31 super-resolution model: It includes a shallow feature extraction module, a deep feature learning module, an attention enhancement module, and an upsampling reconstruction module.

[0044] 1) The shallow feature extraction module includes a 3×3 Conv standard convolutional layer.

[0045] F0 = Conv1(I LR ) Among them: I LR This represents a low-resolution input image, and F0 represents the extracted shallow features. 2) The deep feature learning module consists of 23 Attention Residual-in-Residual Dense Blocks (ARRDBs). Shallow features are input into the deep feature learning module to model the complex spatial and spectral features of the image. To achieve global residual learning, a long-skip residual connection is constructed between the first convolutional layer and the last ARRDB to promote the stable propagation of deep features.

[0046] F deep =ARDB N (ARDB N−1 (...ARDB1(F0)...)) +F0 ARDB N This represents the Nth ARRDB; a total of 23 ARRDBs were used in this embodiment. F deep It represents deep features.

[0047] The ARRDB module is an improved structure that introduces a Convolutional Block Attention Module (CBAM) on top of the Residual-in-Residual Dense Block (RRDB) in ESRGAN. For example... Figure 2 As shown, the RRDB is composed of multiple residual dense blocks (RDBs) cascaded together. Each RDB contains 5 convolutional layers (Conv) and 4 ReLU activation functions, and features are accumulated and reused layer by layer through dense connections to extract rich local features. The residual scaling factor of the RRDB helps to alleviate gradient explosion and stabilize the training process. CBAM is introduced after the residuals of the RRDB, such as... Figure 2 Each DenseBlock incorporates a CBAM (Concentrated Attention Module) to enhance the generator's ability to perceive key spectral and spatial features in remote sensing imagery. The CBAM consists of a Channel Attention Module (CAM) and a Spatial Attention Module (SAM), which operate sequentially on the channel and spatial dimensions of the feature map. Specifically, the Channel Attention Module generates channel weight vectors through global average pooling and max pooling operations to highlight important spectral information; the Spatial Attention Module generates a spatial weight map through convolution operations to emphasize land cover boundaries and texture details. This hierarchical attention enhancement mechanism significantly improves the model's ability to represent complex land cover features in multispectral remote sensing imagery.

[0048] The channel attention module aims to adaptively assign higher weights to important spectral channels by capturing the interdependencies between channels: Mc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) Where F is the input feature map; AvgPoo and MaxPool represent global average pooling and max pooling operations, respectively; MLP represents a multilayer perceptron; and σ represents the sigmoid activation function.

[0049] The spatial attention module is used to model location dependencies in the spatial dimension of the feature map, enhancing the model's response to key feature regions. Ms(F) = σ(f 7×7 ([AvgPool(F);MaxPool(F)])) Among them, f 7×7 represents a 7×7 convolution operation; [·;·] represents a concatenation operation along the channel dimension; AvgPool(F) and MaxPool(F) represent average and max pooling operations along the channel dimension, respectively; σ represents the Sigmoid activation function.

[0050] 3) Figure 1 In the model, after deep feature extraction, the attention enhancement module further calibrates the features. This module includes channel attention and spatial attention. By combining these mechanisms, it captures long-range dependencies within a larger receptive field, thereby enhancing the model's ability to perceive details such as texture and edges.

[0051] 4) The upsampling reconstruction module (upsampler) consists of a sub-pixel convolutional layer (PixelShuffle) and two 3×3 Conv standard convolutional layers. It is used to achieve efficient spatial reconstruction; the feature map undergoes resolution enhancement through the upsampling and reconstruction module.

[0052] F sr = Conv(PixelShuffle(Conv(F deep ))) Where PixelShuffle represents a subpixel convolutional layer, and Fsr represents the SR image.

[0053] S32 Discriminator In this embodiment, the VGG is modified into a U-Net structure with skip connections in the discriminator design. Figure 3 U-Net outputs the ground truth value for each pixel and can provide detailed pixel-by-pixel feedback to the generator. Simultaneously, spectral normalization regularization is used to stabilize training dynamics and mitigate sharp artifacts introduced by GAN training. Through these adjustments, both local and global features can be captured simultaneously.

[0054] S33 Loss Function The loss function plays a crucial role in deep learning models, measuring the difference between predicted and actual results. Its design directly determines the optimization direction and reconstruction effect of the model. To address the requirements of remote sensing imagery in terms of spectral consistency and spatial structure preservation, this embodiment introduces spectral consistency constraints and structure preservation constraints on the original ESRGAN loss framework, thereby enhancing the model's performance in terms of ground feature boundaries and spectral fidelity. The generator loss is as follows:

[0055] in, L G The total loss of the generator. L pix For pixel loss; L per To perceive loss; L gan To combat loss in generators; L spe For spectral loss; L ssim For structural similarity loss; λ pix , λ per , λ gan , λ spe , λ ssim These are the weighting coefficients; In this embodiment, λ pix =1、λ per =1、λ gan =0.1、λ spe =0.2、λ SSIM =0.5.

[0056] The discriminator loss function is as follows:

[0057] in, L D The total loss of the discriminator, L real D This refers to the loss of real data, specifically the error where the discriminator mistakenly identifies a real image as fake. L fake D This is to generate data loss, which is the error of the discriminator misclassifying a fake image as a real image.

[0058] S4. Use the dataset from S2 to train, validate, and test the AESRGAN remote sensing super-resolution model; SR models are typically trained in 8-bit unsigned integer (unit8) format, which shows little difference in visual representation from high-precision data while significantly improving computational efficiency and storage performance. Therefore, this implementation converts GF2 to 8-bit unsigned format to optimize computational resources.

[0059] Considering that directly training a deep learning model on full-size imagery is not only computationally expensive but also prone to class imbalance due to the low proportion of water pixels, this implementation method uses slice generation to generate samples. Specifically, the GF-2 image is cropped into 128×128 pixel slices, and the corresponding S2 image is cropped into 384×384 pixel slices (three times the size of GF2). After manual verification to remove cloud cover and other problematic data, 21,187 slice pairs were finally obtained.

[0060] Regarding dataset partitioning, following typical machine learning training methods, the data was divided into training and validation sets in an 8:2 ratio. 80% of the image pairs were used for model training, and 20% were used for validation and performance evaluation. Considering the temporal balance of the samples, the training and validation data were randomly partitioned according to their time indices.

[0061] S5. LR images are obtained by processing them using the AESRGAN remote sensing super-resolution model; for example... Figure 1 As shown, the main steps include the following: S51. The input LR image is processed through the shallow feature extraction module for feature extraction; the shallow feature extraction formula is as follows: F0 = Conv1(I LR ) Among them: I LR This represents a low-resolution input image, and F0 represents the extracted shallow features. S52. Shallow feature F0 is input into the deep feature extraction module for deep feature extraction. The extraction formula is as follows: F deep =ARDB N (ARDB N−1 (...ARDB1(F0)...)) +F0 Among them: ARDB N F represents the Nth ARRDB. deep Indicates deep features; S53. After deep feature extraction, a 3×3 Conv convolutional layer is passed through to transition to the attention enhancement module for further feature calibration. S54. The feature map after passing through the attention module is then subjected to resolution enhancement and spatial reconstruction by the upsampling reconstruction module, as shown in the following formula: F sr = Conv(PixelShuffle(Conv(F deep ))) Where PixelShuffle represents a subpixel convolutional layer, F sr For SR images; S55. The upsampled feature map is finally adjusted and mapped through two convolutional layers, and the adjusted high-resolution image SR is output.

[0062] S6. Water body extraction and monitoring of water geometric parameters based on SR. S61 Water Monitoring Methods Since GF-2 and ASR imagery only have four spectral bands: blue, green, red, and near-infrared, some water body extraction methods that rely on mid-infrared (MIR) or short-wave infrared (SWIR) bands (such as the Modified Normalized Water Index (MNDWI), Automatic Water Extraction Index (AWEI), and methods based on the HSV color space) are not applicable. For the three types of remote sensing data (S2, GF-2, and ASR imagery) involved in this embodiment, to achieve optimal water body identification for the rivers in the study area, this invention compares various water body extraction methods. Specifically, two threshold-based segmentation methods are employed—NDWI combined with the Otsu automatic thresholding method (NDWI+Otsu). [from Gray-Level Histograms. 2EEEE TRANSACTIONS ON SYSTREMS, MAN, ANDCYBERNETICS, VOL. SMC-9, NO. 1, JANUARY 1979], NDWI combined with K-means clustering (NDWI+KM) [Otsu method and K-means. 2009 Ninth International Conference on Hybrid Intelligent Systems]; Four commonly used machine learning algorithms were also selected: Random Forest (RF) [RandomForests. Machine Learning, 45, 5–32, 2001], Support Vector Machine (SVM) [Tutorial on Support Vector Machine (SVM)], Decision Tree (DT) [Decision tree methods: applications for classification and prediction. Shanghai Archives of Psychiatry, 2015, Vol.27, No. 2], and K Nearest Neighbors (KNN) [KNN Model-Based Approach in Classification. CoopIS / DOA / ODBASE] [2003, LNCS 2888, pp. 986–996, 2003.]; In addition, a deep learning model—U-Net++ [Water Body Segmentation for Satellite Images Using U-Net++. G. Rajalaxmi, SE Vimal, and Janani Selvaraj. https: / / doi.org / 10.1007 / 978-981-97-2069-9_7]. In the feature construction of the machine learning model, in addition to the four bands (blue, green, red, and near-infrared) of the original image, NDWI [NDWI A Normalized Difference WaterIndex for Remote Sensing of Vegetation Liquid Water From Space. REMOTE SENS. ENVIRON.58:257-266 (1996)] and the NDVI index were added to improve the model's ability to distinguish between water and non-water bodies.

[0063] The construction of training samples is a crucial step in water body extraction. To ensure the representativeness and diversity of the samples, this invention selected 129 images covering different river sections and time phases of the Haihe River Basin from S2, GF-2, and SR images, and used manual visual interpretation to annotate the pixel-level samples. The pixels of each image were classified into water bodies (1) and non-water bodies (0). Among them, 109 images were used for model training, and 20 images were used for accuracy verification and confusion matrix analysis. To balance storage efficiency and computational performance, the sample data was stored in 8-bit unsigned integer format.

[0064] S62 River Length and Width Monitoring Methods Based on the water body raster data obtained above, the river centerline is obtained after converting it into a vector file to obtain the river length. The main steps include (1) River line extraction: obtaining the polylines on both sides of the water body; (2) Shoreline pairing: filtering out relatively similar widths, roughly consistent directions, and topologically adjacent boundaries as a pair of shorelines through buffer analysis and topological connectivity; (3) Corresponding point matching: finding the corresponding point on the right bank at the vertex of a shoreline using the nearest neighbor method; (4) Centerline generation: taking the geometric midpoint among the corresponding points and connecting multiple midpoints along the river channel direction to form the initial centerline; (5) Geometric optimization: performing smoothing and topological repair operations on the initial centerline to obtain the river centerline. The length of the river centerline is the length of the water body.

[0065] The length of the line segment perpendicular to the center line of the river and intersecting the broken lines on both banks of the river is the width of the river.

[0066] The evaluation metrics in this embodiment include: 1) Evaluation indicators for super-resolution results of remote sensing images The super-resolution reconstruction results of remote sensing images should not only be visually "realistic" but also maintain high fidelity in terms of spectral consistency. Therefore, this embodiment uses three metrics—peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and spectral loss—to evaluate model performance.

[0067] PSNR is used to evaluate image distortion; a larger value indicates less image distortion.

[0068]

[0069] Where C, H, and W represent the channels, height, and width of the SR image.

[0070] SSIM is used to describe the similarity between two images, measuring brightness, contrast, and structure. SSIM values ​​range from 0 to 1; the closer the value is to 1, the higher the structural similarity between the two images.

[0071]

[0072] Where μ and σ represent the mean and variance of the image, and constants C1 = k1L and C2 = k2L, where L represents the dynamic range of pixel values, typically K1 = 0.01 and k2 = 0.03.

[0073] Spectral loss ( This is used to quantify the deviation in spectral information between the super-resolution image and the original high-resolution image. The smaller the spectral loss, the better the reconstruction effect of the model in terms of spectral consistency.

[0074]

[0075] Where C, H, and W represent the number of image channels, height, and width, respectively, and S SR and S HR These are the spectral reflectance values ​​of the super-resolution result and the reference high-resolution image, respectively.

[0076] 2) Monitoring and evaluation indicators of water bodies and their geometric parameters The overall accuracy (OA) and Kappa coefficient of the confusion matrix are used to evaluate the accuracy of water body extraction. OA refers to the proportion of correctly classified samples out of the total number of samples, and the Kappa coefficient is an accuracy index that takes into account random factors. Higher OA and Kappa values ​​indicate that the classification model has better reliability and stability in water body identification tasks.

[0077]

[0078] Where CM[i,i] is the number of correct classifications for the i-th class in the confusion matrix. This represents the total number of samples, where n is the number of categories.

[0079]

[0080] Where OA is the total precision. It is the expected accuracy of random classification.

[0081] The relative error is used to evaluate the accuracy of water body geometric parameters. It reflects the degree of deviation between the monitoring results and the true values. The smaller the value, the higher the accuracy of the model in terms of geometric estimation.

[0082]

[0083] Where VM is the monitored value and VT is the true value.

[0084] The test results of this embodiment are as follows: 1) Performance of the super-resolution model: To systematically evaluate the applicability and performance of the super-resolution model (AESRGAN model) constructed in this embodiment on river bodies of different scales, satellite imagery data covering various typical river width types were selected in this embodiment, including: reservoir (AR), 330 m wide river (BW), 158 m wide river (CW), 67 m wide river (DW), 32 m wide river (EW), and 16 m wide river (FW).

[0085] In comparison, in addition to the high-resolution image HR (derived from GF2) and low-resolution image LR (derived from S2) super-resolution model AESRGAN in this embodiment, three other super-resolution methods were also used: bicubic interpolation, traditional super-resolution method (SRGAN, abbreviated as SR) and ESRGAN model for comparison.

[0086] like Figure 4 As shown in the comparison results, AESRGAN performs best in detail restoration, structure preservation, and visual realism. Specifically, images generated by AESRGAN have clearer water-land boundaries, smoother river textures, and more natural detail transitions. In small-scale river scenes (such as FW), AESRGAN is particularly outstanding in suppressing blurred edges and enhancing texture continuity, effectively restoring detailed information of narrow river sections and tributaries, showing significant advantages over traditional methods.

[0087] Furthermore, for remote sensing imagery, spectral information features are crucial for water body extraction and classification. To analyze the impact of super-resolution processing on spectral fidelity, this implementation statistically analyzed the spectral distribution characteristics of four bands (B, G, R, N) (see...). Figure 5 The results show that super-resolution processing can effectively change the spectral distribution of the medium-resolution image (S2), making it highly similar to the spectral characteristics of the high-resolution image (GF2), and at the same time verifying the fidelity of the super-resolution algorithm in the spectral domain.

[0088] 2) Water monitoring methods To systematically evaluate the impact of super-resolution (SR) on water body extraction accuracy and compare the performance of different methods on images of different resolutions, this study selected S2, GF2, and SR images of major river points and applied seven water body extraction methods: NDWI + Otsu, NDWI + K-Means, DT, RF, SVM, KNN, and U-Net++. All seven methods are existing technologies, and references are provided above.

[0089] Figure 6This study compares the extraction results of different methods on typical river water bodies. The results show that the original GF2 image has clear texture, and the extracted water body boundaries are continuous and closely match the actual water body morphology. The S2 image, due to its 10-meter resolution, exhibits significant pixel effects, resulting in coarse water body boundaries and noticeable jagged edges, making it difficult to accurately reflect the true river morphology. The SR image, after AESRGAN super-resolution, shows significantly improved spatial detail and enhanced water body boundary and regional continuity, with its extraction results visually approaching those of the GF2 image.

[0090] A confusion matrix was calculated from the extracted results, and the average value of each river point was taken to obtain the average overall precision (OA) and average Kappa coefficient. The results are shown in Table 1:

[0091] In terms of image type, GF2 has the highest overall water body extraction accuracy, while SR has an accuracy close to that of GF2, and both are superior to S2 imagery.

[0092] For S2 imagery, the deep learning (DL) method performed best, achieving an OA of 0.94; followed by NDWI + Kmeans and NDWI + Otsu, both with an OA of 0.93; the decision tree (DT) and random forest (RF) methods had an OA of 0.92; and the support vector machine (SVM) method had the lowest overall accuracy at 0.89. Considering both OA and Kappa coefficient, the DL method demonstrated superior performance in water body extraction from S2 imagery. For GF2 imagery, NDWI + Otsu performed best, achieving an OA of 0.98; SVM, RF, KNN, and DL methods all had an OA of 0.97; the DT method had an OA of 0.96; and the NDWI + Kmeans method had the lowest OA at only 0.880. Considering both overall accuracy and Kappa coefficient, NDWI + Otsu demonstrated superior performance in water body extraction from GF-2 imagery. For SR imagery, the U-Net++ method performed best, with an OA of 0.97; KNN and NDWI+otsu methods were next, with an OA of 0.96; SVM was 0.95; RF was 0.94; NDWI+KM was 0.93; and the DT method was the lowest, with an OA of 0.92. Considering both OA and Kappa coefficients, DL exhibited the best performance in water body extraction from ASR imagery.

[0093] SR processing significantly improves the water body extraction accuracy of S2 images, making its performance close to that of GF2 images under various methods. At the same time, deep learning methods still have the best performance on SR images.

[0094] 3) Monitoring of water body geometric parameters Using the optimal water body extraction method for each image, the geometric parameters (area, river length, and river width) of rivers at different scales were extracted (Figure 7). The results show that the spatial resolution of the image and super-resolution processing have a significant impact on the quantification accuracy of the geometric features of the water body, and the responses of water bodies at different scales are different.

[0095] High-resolution GF2 imagery demonstrates significant advantages in quantizing the geometric features of all water bodies. For large-scale water bodies, its length, width, and area remain highly consistent; for small-scale water bodies, the quantization error is also relatively small, indicating that high-resolution imagery can accurately capture subtle boundary information of rivers. In contrast, S2 imagery generally suffers from distortion in geometric parameter extraction, especially for small-scale water bodies (such as EW and FW), where the extraction error is relatively large. This is closely related to its 10-meter spatial resolution, which cannot effectively resolve the edges of small water bodies. SR imagery processed by AESRGAN shows improvement in the extraction of geometric features for all water bodies. For large-scale water bodies, its geometric parameters are close to those of GF2, verifying the effectiveness of the super-resolution model in spatial detail recovery; for small-scale water bodies, although the improvement is limited, it is still a significant improvement over the original S2 imagery. Overall, high-resolution imagery remains the optimal data source for extracting water body geometric parameters, while super-resolution processing can significantly improve the accuracy of medium-resolution imagery in water body geometric quantization.

[0096] Furthermore, we analyzed the impact of time-series variations of the same river on the extraction of geometric parameters. Given the significant spatiotemporal variability in river width, this implementation extracted the time series of river widths for major rivers from 2020 to 2024. Figure 8 The results showed that the river width extracted from the SR image was highly consistent with that of the GF2 image, and its time series curve was stable and had a high degree of fit, which could reflect the trend of river width change over time. However, due to the spatial resolution limitation, the S2 image not only showed a significant deviation in the absolute value of river width, but also had distortion in the trend of change, making it difficult to accurately capture the small fluctuations in the time series.

[0097] Based on the analysis in this implementation: From the perspective of dataset construction: This invention constructs the GF2–S2 remote sensing super-resolution dataset based on high-resolution GF2 imagery and medium-resolution S2 imagery. Through radiometric calibration, atmospheric correction, geometric correction, coordinate reprojection, resolution resampling matching, and spatial registration, HR-LR training samples are established. This dataset exhibits high consistency in both spatial and spectral dimensions, providing authentic and reliable data support for remote sensing image super-resolution research.

[0098] From the perspective of model construction: To systematically evaluate the performance of different super-resolution models in remote sensing image reconstruction tasks, this study calculated the PSNR, SSIM, and Spectral Loss of Cubic, SRGAN, ESRGAN, and the improved AESRGAN model (Table 2). These three metrics reflect the overall performance of the models from three dimensions: geometric detail restoration, structural preservation, and spectral consistency.

[0099]

[0100] The results show that the AESRGAN model achieved the best performance across all three metrics. Specifically, the PSNR improved to 33.83, approximately 1.49 dB higher than ESRGAN, indicating a significant advantage in pixel-level reconstruction accuracy. The SSIM reached 0.882, demonstrating the model's ability to more effectively recover the spatial structure and texture relationships of ground features. Simultaneously, the spectral loss decreased to 0.008, significantly better than other models, indicating that AESRGAN has a stronger ability to maintain spectral consistency. This improvement in spectral fidelity is crucial for remote sensing tasks, avoiding spectral aliasing caused by artifacts and improving the classification stability of targets such as water bodies.

[0101] From the perspective of water body remote sensing monitoring: The key difference between 10 m and 2 m resolution water monitoring models lies in their accuracy in segmenting riverbanks and identifying small-scale rivers. Land-water boundary pixels may only account for a small fraction of the total river pixels; therefore, lower-resolution imagery may still exhibit high accuracy (as shown in Table 1), which is satisfactory, but this difference has significant implications for fine-grained hydrological monitoring.

[0102] Figure 9 This is a comparison of water body monitoring results from FW. It can be seen that GF2 and SR depict riverbanks better, while S2, due to its jagged edges, cannot fully depict the details of the riverbanks. Notably, within the red box in the image, a small-scale river approximately 5m wide, GF2 and SR can both successfully identify it, while S2 cannot detect it. This indicates that when the river scale is close to or smaller than the original spatial resolution of S2 (10m), it has a significant blind spot in monitoring small rivers due to the mixed pixel effect. In contrast, AESRGAN, through detail enhancement and boundary reconstruction, effectively compensates for the lack of spatial detail in medium-resolution images, enabling the identification of small river channels.

[0103] Monitoring of water body geometric parameters: This invention uses the water body geometric parameters extracted from GF2 imagery as the true values, and selects typical river segments (AR, BW, CW, DW, EW, FW) for relative error analysis of area, river length, and river width. Figure 10), quantitative analysis of how AESRGAN improves data error.

[0104] In terms of river area, the relative error of S2 imagery at most scales ranges from -0.88% to 7.64%, while in the FW river segment, the relative area error of S2 imagery reaches as high as 53.59%, showing a significant deviation. In contrast, the relative error of SR imagery is -8.74%, demonstrating a significant improvement in overall accuracy. In other river segments, the absolute value of the relative error of SR imagery is below 0.3%, highly consistent with the results of GF2. Regarding river length extraction, both S2 and SR imagery show relatively small extraction errors, with the absolute value of the relative error below 10% in most river segments, with SR performing slightly better than S2. In the FW scale river segment, the error of SR (2.40%) is lower than that of S2 (5.47%). Overall, the accuracy of SR imagery is slightly better than that of S2 imagery. This is because river length estimation is mainly affected by the continuity of spatial structure and does not involve the spectral mixing effect of mixed pixels. For river width, in the FW river segment, the relative error of river width in S2 imagery is 18.75%, while the error of SR imagery decreases to -6.25%. In the remaining river sections, the error of SR imagery was generally one-third or more lower than that of S2 imagery. Compared to S2, SR imagery improved the accuracy of area, river length, and river width by 8.51%, 0.34%, and 3.28%, respectively. Overall, SR significantly reduced errors compared to S2 at most scales, especially for parameters that are more affected by mixed pixels and are sensitive to resolution. All error values ​​exhibited randomness within the positive and negative ranges, without forming a consistent bias trend.

[0105] The area and geometric parameter errors caused by spatial resolution limitations in S2 imagery are significantly reduced after AESRGAN processing. AESRGAN effectively mitigates morphological distortions and boundary discontinuities caused by scale effects, significantly improving spatial geometric accuracy while maintaining spectral consistency. High-quality super-resolution reconstruction not only improves visual resolution and image detail but also quantitatively enhances the extraction accuracy of key hydrological geometric parameters such as water area and river width. This indicates that spatial augmentation techniques based on deep super-resolution models can provide more reliable data support and theoretical basis for watershed-scale hydrological element inversion and hydrodynamic model parameterization.

[0106] In summary, this implementation method constructs a water body monitoring research framework based on the AESRGAN model, which can accurately classify rivers up to 3 meters in diameter from publicly available S2 multispectral imagery. The performance of super-resolution imagery in water body identification and geometric feature extraction is also systematically evaluated. The main conclusions are as follows: A GF2-S2 super-resolution dataset suitable for remote sensing scenarios was constructed. Based on GF2 (3 m) and S2 (10 m) images, a GF2-S2 super-resolution dataset with real remote sensing features was constructed through registration, reprojection and resolution pairing. This provides a reliable high- and low-resolution comparison data source for subsequent model training and enriches the dataset in the field of remote sensing super-resolution.

[0107] An improved AESRGAN model is proposed. Based on the ESRGAN framework, an attention mechanism is introduced into the generator to construct an ARRDB module to enhance spatial texture representation. Spectral consistency and structural similarity terms are added to the loss function. The discriminator is modified to a U-Net network with spectral normalization. Experimental results show that the AESRGAN model achieves state-of-the-art performance on the GF2–S2 dataset: PSNR = 33.83 dB, SSIM = 0.882, and Spectral Loss = 0.008, significantly outperforming benchmark models such as Cubic, SRGAN, and ESRGAN. This indicates that it achieves synergistic optimization in pixel reconstruction, structure preservation, and spectral fidelity.

[0108] Super-resolution imaging significantly improves the accuracy of water body extraction. Seven water body extraction methods (NDWI+Otsu, NDWI+KMeans, DT, RF, SVM, KNN, and UNet++) were compared and analyzed on three types of images: S2, GF2, and SR (AESRGAN results). The results showed that the overall classification accuracy (OA) and Kappa coefficient of SR imagery were significantly better than the original S2 imagery, and very close to those of the high-resolution GF2 imagery. The optimal water body extraction method for both GF2 and SR images was U-Net++, while the NDWI+Otsu method performed best for S2 imagery. Visual comparison showed that the water body boundaries in SR imagery were smoother and more continuous, and the spatial morphology was closer to that of the high-resolution imagery. For narrow rivers (approximately 5 m wide) with a resolution less than 10 m smaller than the original S2 imagery, SR imagery was able to identify water bodies that S2 imagery could not capture.

[0109] Super-resolution reconstruction effectively improves the monitoring accuracy of water body geometric parameters. Using the GF2 extraction results as the true values, relative error analysis of water body geometric parameters shows that the SR imagery improves the accuracy of area, river length, and river width by 8.51%, 0.34%, and 3.28%, respectively, compared to S2. Overall, SR significantly reduces errors compared to S2 at most scales, especially for parameters that are more affected by mixed pixels and are sensitive to resolution. All error values ​​exhibit randomness within the positive and negative ranges, without forming a consistent bias trend.

[0110] Finally, it should be noted that the above description is only used to illustrate the technical solutions of the present invention and is not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A method for monitoring water body and geometric parameters based on an AESRGAN remote sensing super-resolution model, characterized in that: Includes the following steps: S1. Collect satellite images of the study area: The satellite images include low-resolution (LR) images and high-resolution (HR) images; S2, Data Pairing: Construct a dataset consisting of HR and LR image pairings for subsequent training, validation and testing of the AESRGAN remote sensing super-resolution model; S3. Constructing the AESRGAN remote sensing super-resolution model: The overall structure of the generator network of the super-resolution model includes a shallow feature extraction module, a deep feature learning module, an attention enhancement module, and an upsampling reconstruction module; the deep feature learning module consists of multiple attention residual dense blocks (ARRDBs); the ARRDBs introduce the CBAM attention mechanism into the residual-residual dense block (RRDB); the RRDBs are composed of multiple cascaded residual dense blocks (RDBs), each RDB containing 4-5 convolutional layers and ReLU activation functions; the CBAMs consist of a channel attention module (CAM) and a spatial attention module (SAM), which act sequentially on the channel dimension and spatial dimension of the feature map; The generator network employs spectral consistency and structural similarity loss functions, as follows: ； in, L G The total loss of the generator network, L pix For pixel loss; L per To perceive loss; L gan To mitigate the loss in the generator network; L spe For spectral loss; L ssim For structural similarity loss; λ pix , λ per、 λ gan , λ spe , λ ssim These are the weighting coefficients; S4. Use the dataset from S2 to train, validate, and test the AESRGAN remote sensing super-resolution model; S5. Obtain SR imagery by processing LR images based on the AESRGAN remote sensing super-resolution model; including the following steps: S51. The input LR image is processed through the shallow feature extraction module for feature extraction; the shallow feature extraction formula is as follows: F0 = Conv1(I LR ) Among them: I LR This represents a low-resolution input image, and F0 represents the extracted shallow features. S52. Shallow feature F0 is input into the deep feature extraction module for deep feature extraction. The extraction formula is as follows: F deep =ARDB N (ARDB N−1 (...ARDB1(F0)...)) +F0 Among them: ARDB N F represents the Nth ARRDB. deep Indicates deep features; S53. After deep feature extraction, a 3×3 Conv convolutional layer is passed through to transition to the attention enhancement module for further feature calibration. S54. The feature map after passing through the attention module is then subjected to resolution enhancement and spatial reconstruction by the upsampling reconstruction module, as shown in the following formula: F sr = Conv(PixelShuffle(Conv(F deep ))) Where PixelShuffle represents a subpixel convolutional layer, F sr For SR images; S55. Perform final feature adjustment and mapping on the upsampled feature map through two convolutional layers, and output the adjusted high-resolution image SR. S6. Water body extraction and water body geometric parameter monitoring based on SR.

2. The method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model according to claim 1, characterized in that: The high-resolution imagery HR mentioned in S1 is from the Gaofen-2 GF2 satellite, and the low-resolution imagery LR is from the Sentinel-2 satellite.

3. The method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model according to claim 1, characterized in that: In S2: Data pairing includes radiometric calibration, atmospheric correction, geometric correction, coordinate reprojection, resolution resampling matching, and spatial registration.

4. The method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model according to claim 1, characterized in that: In S3: the shallow feature extraction module includes a 3×3 Conv standard convolutional layer; the deep feature learning module consists of at least 20 attention residual dense blocks ARRDB; the RRDB is composed of at least 3 residual dense blocks RDB cascaded together; the attention enhancement module includes channel attention and spatial attention; the upsampling reconstruction module consists of a subpixel convolutional layer PixelShuffle and two 3×3 Conv standard convolutional layers.

5. The method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model according to claim 1, characterized in that: In S3: The CBAM consists of a channel attention module (CAM) and a spatial attention module (SAM), which act sequentially on the channel dimension and spatial dimension of the feature map; specifically: the channel attention module (CAM) generates channel weight vectors through global average pooling and max pooling operations; the spatial attention module (SAM) generates a spatial weight map through convolution operations.

6. The method for monitoring water bodies and geometric parameters based on the AESRGAN remote sensing super-resolution model according to claim 1, characterized in that: The water extraction method in step S6 is the U-Net++ method.

Images