A surface water body extraction system and method based on artificial intelligence algorithm

Abstract

A surface water body extraction system and method based on an artificial intelligence algorithm, the system comprising a backend processing system and a front-end operating system; the method: a user registers and logs in through a user login module of the front-end operating system; the user selects data and corresponding parameters through a data selection module, selects a band combination of a remote sensing image and an image stretching method; a model selection module is used to determine a water body extraction model, the backend processing system adjusts the band combination of the remote sensing data according to the corresponding parameters, calls the corresponding model in a database storage module and processes the data using an algorithm processing module to obtain a water body extraction result; whether to verify the water body extraction result is selected through a verification module, if verification is selected, the accuracy of different results is evaluated using precision, recall, F1 score and average intersection over union ratio indexes, if no verification is selected, the water body extraction result is directly output; the present application has the advantages of high water body extraction accuracy and strong technical robustness.

Description

Technical Field

[0001] This invention belongs to the field of water extraction technology, specifically relating to a surface water extraction system and method based on artificial intelligence algorithms. Background Technology

[0002] Over the past few decades, surface water extraction methods based on remote sensing imagery have been mainly classified into three categories: thresholding methods based on single-band images, identification methods based on spectral indices, and image classification methods. Among these, single-band thresholding is one of the most common surface water extraction methods. However, the spectral reflectance characteristics of water bodies vary with factors such as season, geographical location, and depth. Non-water objects may have similar spectral reflectance characteristics to water bodies in specific bands. Therefore, relying solely on spectral reflectance characteristics to determine the spatial distribution of surface water in a region has certain limitations. Spectral index-driven water extraction methods leverage the different spectral reflectance characteristics of water in different bands, combining multiple bands to highlight water information in remote sensing images. For example, the Normalized Difference Water Index (NDWI) proposed by McFeeters highlights water information in images and suppresses vegetation and soil information based on the prominent reflectance of water in the green band and strong absorption in the infrared band. However, different background information has varying effects on the accuracy of water extraction, and NDWI is not well-suited to all real-world conditions.

[0003] With the development of artificial intelligence technology, some scholars have been inspired to study surface water extraction based on machine learning. Image semantic segmentation models are constantly being improved, leading to increased accuracy in surface water extraction. From the initial Fully Convolutional Networks (FCN) capable of pixel-by-pixel classification, to PSPNET which aggregates contextual information from different regions to enhance global information acquisition, and the recently proposed Segmentation Transformer (SETR) semantic segmentation model based on the ViT architecture, it has been proven that Transformer-based semantic segmentation models can learn more semantic features compared to CNN-based models. However, as accuracy improves, we must also pay attention to the network's structural complexity. SETR's complex structure results in very high parameters and computational costs, which mobile models cannot handle. Furthermore, existing water extraction algorithms are not integrated, making the extraction process cumbersome. Moreover, different algorithms produce varying results and processing times for different images.

[0004] With the continuous development of remote sensing technology, it has evolved from simple environmental monitoring to environmental impact assessment and prediction; from one-time monitoring to continuous dynamic monitoring; from single-target to multi-target, multi-level quantitative analysis; and from single satellite data sources to multi-data source, multi-temporal, multi-resolution remote sensing. Meanwhile, remote sensing monitoring has advantages such as large area coverage, real-time performance, and cost-effectiveness. Extracting information such as water body area, geometric morphology, and aquatic ecological environment from satellite remote sensing imagery has been applied in fields such as water resource surveys, environmental protection, and macro-monitoring of water resources.

[0005] In recent years, the development of artificial intelligence technology has prompted some scholars to study the extraction of surface water bodies based on machine learning. For example, Abid et al. used an unsupervised curriculum learning method based on convolutional neural networks to identify water bodies, which overcomes the challenges faced by remote sensing images. Chen et al. used convolutional neural networks to extract water bodies and demonstrated the effectiveness of deep learning methods in water body extraction by comparing them with traditional methods such as the Normalized Difference Water Index (NDWI).

[0006] Currently, most research on surface water extraction based on remote sensing imagery relies on three main techniques: thresholding based on single-band images, identification methods based on spectral indices, and image classification methods. These methods either depend on manually designed computational models or fail to consider the rich band combination information in remote sensing images, and the efficiency, accuracy, and robustness of these models all need improvement. Therefore, rationally evaluating the optimal remote sensing image data source for surface water extraction tasks is particularly important. Currently, there are many image classification methods based on deep learning, such as FCN, PSPNET, U-NET, and the DeepLabV3 series.

[0007] Currently, most water body extraction tasks utilize true-color synthetic imagery, neglecting the rich band combinations of multispectral and hyperspectral imagery. Furthermore, existing water body extraction tasks largely employ unencapsulated models, resulting in complex extraction processes and excessively high levels of expertise required.

[0008] Single-band thresholding extraction methods effectively and rapidly extract water body information by extracting the grayscale values ​​of a specific band in remote sensing images. However, the spectral reflectance characteristics of water bodies are affected by many factors, and single-band images are not strong at representing the features of ground objects.

[0009] The spectral index method for extracting water body characteristics uses remote sensing data from different spectral bands to calculate values ​​based on the spectral reflectance characteristics of water bodies in different spectral bands, and is used to characterize water body features. However, this method suffers from problems such as the difficulty in determining the threshold.

[0010] Image classification has proven to be a highly effective method for extracting remotely sensed ground features. Furthermore, due to advancements in artificial intelligence, the number of image classification algorithms is increasing, and their accuracy is continuously improving. However, the integration level of current algorithm models remains low. The algorithms themselves are highly specialized and difficult to understand quickly, and their usage is complex, requiring initial environmental configuration and parameter adjustments based on computing power. This clearly cannot meet the needs of rapid and high-precision water feature extraction tasks.

[0011] Patent application CN115393733A discloses a method and system for automatic water body identification based on deep learning. However, it only has a single extraction model, U-NET, which cannot guarantee the robustness and accuracy of the model on other remote sensing datasets. Furthermore, it does not take into account the rich band information of remote sensing images. A single band combination cannot guarantee the accuracy of water body identification. Moreover, the system does not have a self-checking system and cannot verify the identification results. Summary of the Invention

[0012] To overcome the shortcomings of the existing technology, the present invention aims to provide a surface water extraction system and method based on artificial intelligence algorithms. By filtering the band combinations of remote sensing data, the system can obtain the band combination containing the most water feature information, thereby improving the accuracy of water extraction. By proposing the MFM-Segformer algorithm, multimodal feature fusion can be achieved. By refining the water extraction process of image classification, the present invention proposes a water extraction system that can save time, manpower, and financial costs, and has the advantages of high accuracy and strong technical robustness in water extraction.

[0013] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0014] A surface water extraction system based on artificial intelligence algorithms includes a back-end processing system and a front-end operating system. Each module in the front-end operating system sends data streams and corresponding parameters to the back-end processing system, which processes the data and then sends it back to the front-end operating system. The back-end processing system includes an algorithm processing module and a database storage module. The front-end operating system includes a user login module, a data selection module, a model selection module, and a verification module.

[0015] The user login module verifies the user's identity information. After successful verification, the user enters the main interface of the surface water extraction system. The user inputs the remote sensing data of the water body to be extracted through the data selection module. The database storage module provides the model. The remote sensing data of the water body to be extracted from the data selection module and all models in the database storage module are sent to the model selection module. After the model selection module selects a model, the algorithm processing module performs water extraction according to the selected model and sends the water extraction results to the verification module. The verification module verifies and outputs the water extraction results and accuracy indicators. Alternatively, the user can choose not to perform verification and directly output the water extraction results.

[0016] The algorithm processing module incorporates the water extraction algorithm MFM-SegFormer. The multi-scale feature fusion module MFM first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules for multi-scale feature map extraction. The first scale feature extraction module consists of a Pooling layer, a convolutional layer, a BN layer, a ReLU layer, and an Upsampling layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is located in four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an Overlap Patch. The Merging layer consists of a Mix-FNN layer, which is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly.

[0017] The database storage module is used to store water body extraction model data that can be used by the model selection module of the front-end operating system.

[0018] The user login module is used for user registration and login;

[0019] The data selection module is used to select the band combination and image stretching method of the remote sensing image;

[0020] The model selection module is used to select a water extraction model and its specific parameters; the verification module is used to verify the water extraction results.

[0021] The database storage module includes the MFM-SegFormer model.

[0022] The database storage module also includes: FCN, PSPNET, U-NET, Deeplabv3 series or Segformer.

[0023] This invention also provides a method for extracting surface water bodies based on artificial intelligence algorithms, comprising the following steps:

[0024] Step 1: Users register and log in through the user login module of the front-end operating system;

[0025] Step 2: Input the remote sensing data of the water body to be extracted through the data selection module, open the remote sensing image of the water body to be extracted, select the data and corresponding parameters through the data selection module, and select the band combination and image stretching method of the remote sensing image.

[0026] Step 3: Use the model selection module to determine the model for water body extraction. The back-end processing system adjusts the band combination of the remote sensing data according to the corresponding parameters, calls the corresponding model in the database storage module, and uses the algorithm processing module to process the data to obtain the water body extraction result.

[0027] Step 4: The water extraction results obtained in Step 3 are evaluated by the validation module. If validation is selected, the accuracy of different results is evaluated by precision, recall, F1 score and cross-validation ratio. If no validation is selected, the water extraction results are output directly.

[0028] The model in the database storage module of step 3 includes the MFM-SegFormer model.

[0029] The models in the database storage module of step 3 also include: FCN, PSPNET, U-NET, Deeplabv3 series or Segformer.

[0030] Step 3 uses an algorithm processing module to process the data. The specific steps are as follows:

[0031] The water extraction algorithm MFM-SegFormer is adopted. The multi-scale feature fusion module MFM first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules for multi-scale feature map extraction. The first scale feature extraction module consists of a pooling layer, a convolutional layer, a batch normalization (BN) layer, a ReLU layer, and an overlay layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is located in four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an Overlap Patch. The Mix-FNN layer consists of a Merging layer, where the Mix-FNN layer is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention designs an improved water extraction algorithm MFM-SegFormer based on artificial intelligence algorithms. Considering the different efficiencies and applicable ranges of different algorithms, and adding other mainstream algorithms, this model can extract water with higher accuracy and stronger robustness.

[0034] 2. The data selection module included in the front-end operating system of this invention is not only compatible with all remote sensing data types, but also adds the selection of band combinations; this not only enhances the reliability of the samples, but also improves the accuracy of the model in extracting water bodies.

[0035] 3. The model selection module included in the front-end operating system of this invention takes into account the different efficiencies and applicability of different algorithms for different remote sensing image data. Therefore, the model database includes current mainstream image classification models, such as FCN, PSPNET, U-NET, Deeplab3 series, Segformer, and MFM-SegFormer.

[0036] 4. The front-end operating system of this invention includes a verification module that can evaluate the accuracy of water extraction results based on user-selected input of a specified validation set or manual annotation. Precision, recall, F1 score, and average-intersection-union ratio (AUI) are used to evaluate the accuracy of different results.

[0037] In summary, this invention addresses the challenges of current image classification algorithms, which are highly specialized, difficult to understand quickly, and complex to use, requiring pre-application environment configuration and parameter adjustments based on computing power. By integrating water extraction with a front-end operating system and a back-end processing system, this invention not only saves time, manpower, and financial costs but also achieves high accuracy and robustness in water extraction. Attached Figure Description

[0038] Figure 1 This is a system module diagram of the present invention.

[0039] Figure 2 This is the control flow diagram of the present invention.

[0040] Figure 3 This is a diagram of the water extraction network structure of the present invention.

[0041] Figure 4 This is the main interface of the system of this invention.

[0042] Figure 5 This invention relates to the selection of waveband combinations and corresponding parameter settings for water body extraction in the system of this invention.

[0043] Figure 6 The model selection and corresponding parameter settings for water extraction in the system of this invention.

[0044] Figure 7 This invention describes the water treatment process.

[0045] Figure 8 The results of water extraction in this invention.

[0046] Figure 9 The figure shows the simulation results of this invention. Detailed Implementation

[0047] The technical solution adopted by the present invention will be described in detail below with reference to the accompanying drawings.

[0048] The purpose of this invention is not only to achieve high-precision water body extraction through different band combinations of remote sensing data and improved neural networks, but also to integrate the water body extraction process into one unit, making the work more lightweight.

[0049] In addition, this invention also designs an improved SegFormer algorithm, which uses a multimodal feature fusion module (MFM) to improve network performance; that is, an improved water extraction algorithm MFM-SegFormer, which takes into account the different efficiencies and applicable ranges of different algorithms, and adds other mainstream algorithms.

[0050] like Figure 1As shown, this invention provides a surface water extraction system based on artificial intelligence algorithms. The system mainly includes a back-end processing system and a front-end operating system. Each module in the front-end operating system sends data streams and corresponding parameters to the back-end processing system, which processes the data and then sends it back to the front-end operating system. The back-end processing system includes an algorithm processing module and a database storage module. The front-end operating system includes a user login module, a data selection module, a model selection module, and a verification module.

[0051] The user login module verifies the user's identity information. After successful verification, the user enters the main interface of the surface water extraction system. The user inputs the remote sensing data of the water body to be extracted through the data selection module. The database storage module provides the model. The remote sensing data of the water body to be extracted from the data selection module and all models in the database storage module are sent to the model selection module. After the model selection module selects a model, the algorithm processing module performs water extraction according to the selected model and sends the water extraction results to the verification module. The verification module verifies and outputs the water extraction results and accuracy indicators. Alternatively, the user can choose not to perform verification and directly output the water extraction results.

[0052] like Figure 3 As shown, the algorithm processing module incorporates the water extraction algorithm MFM-SegFormer, and a multimodal feature fusion module (MFM) is designed and used to improve network performance. Specifically, the MFM module first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules to extract multi-scale feature maps. The first scale feature extraction module consists of a Pooling layer, a convolutional layer, a BN layer, a ReLU layer, and an Upsampling layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is located within four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an Overlap Patch. The network consists of a Merging layer, where the Mix-FNN layer is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly. Based on this network, a water extraction model with high accuracy and robustness is proposed.

[0053] The database storage module is used to store water body extraction model data that can be used by the model selection module of the front-end operating system.

[0054] The user login module is used for user registration and login;

[0055] The data selection module is used to select the band combination and image stretching method of the remote sensing image; the system is not limited to a single type of remote sensing data, but performs water body extraction on all remote sensing data.

[0056] The model selection module is used to select the water extraction model and the specific parameters of the model;

[0057] The verification module is used to verify the water extraction results.

[0058] The database storage module includes: FCN, PSPNET, U-NET, Deeplabv3 series, Segformer or MFM-SegFormer model.

[0059] By considering factors such as the diversity of remote sensing data and the applicability of algorithms, appropriate models should be selected to ensure the accuracy of water body extraction. This enables rapid and high-precision water body extraction, playing a crucial role in the socio-economic and sustainable development management of water resources.

[0060] FCN, proposed by Long et al. in their 2015 paper "Fully Convolutional Networks for Semantic Segmentation," is a framework for image semantic segmentation. In traditional CNN networks, several fully connected layers follow the final convolutional layer, mapping the feature maps generated by the convolutional layers into a fixed-length feature vector. This structure is suitable for image-level classification and regression tasks, ultimately yielding the classification probability of the input image.

[0061] PSPNET performs feature fusion at four different coarse and fine scales using a pyramid pooling module. The coarsest scale performs global average pooling on the feature map, producing a single-cell output; the finer scales divide the feature map into different sub-regions, producing multi-cell outputs. The outputs at different scales correspond to feature maps of different sizes, and then the low-dimensional feature maps are upsampled through bilinear interpolation to obtain features of the same size.

[0062] UNet combines shallow and deep features through a U-shaped network structure with jumpers for generating the final semantic segmentation map. Unlike FCN, which combines shallow and deep features by concatenation, UNet combines them by addition. It is both lightweight and high-performance, and is therefore often used as a baseline test model for semantic segmentation tasks.

[0063] In 2015, Liang et al. proposed Deeplabv1, which combined deep convolutional neural networks (DCNNs) and probabilistic graphical models (DenseCRFs) to overcome the localization characteristics of deep networks. Deeplabv3, after removing DenseCRF, proposed cascaded modules that progressively increase the dilation rate, while the hollow spatial pyramid pooling module detects features with image-level characteristics by using multiple sampling rates and an effective field of view.

[0064] SegFormer is a simple, efficient, yet powerful semantic segmentation framework that unifies Transformers with a lightweight multilayer perceptron (MLP) decoder. SegFormer includes a novel hierarchical Transformers encoder that outputs multi-scale features. SegFormer avoids complex decoders. The proposed MLP decoder aggregates information from different layers, combining local and global attention to produce powerful representations.

[0065] like Figure 2 As shown, the present invention also provides a method for extracting surface water bodies based on artificial intelligence algorithms, comprising the following steps:

[0066] Step 1: Users register and log in through the user login module of the front-end operating system;

[0067] Step 2, as follows Figure 4 As shown, the remote sensing data of the water body to be extracted is input through the data selection module, and the remote sensing image of the water body to be extracted is opened. The user selects the data and corresponding parameters through the data selection module (such as...). Figure 5 (As shown), select the band combination and image stretching method of the remote sensing image;

[0068] Step 3: Use the model selection module to determine the model for water extraction (e.g., ...). Figure 6 As shown), and select the backbone network and whether pre-trained weights are needed; the backend processing system adjusts the band combination of the remote sensing data according to the corresponding parameters, calls the corresponding model in the database storage module, and uses the algorithm processing module to process the data (such as...). Figure 7 As shown), the water extraction results were obtained (e.g. Figure 8 (as shown);

[0069] Step 4: The user selects whether to validate the water extraction results obtained in Step 3 through the validation module. The user can specify the validation dataset or manually label it according to their needs. If validation is selected, the accuracy of different results is evaluated by indicators such as precision, recall, F1 score and intersection-union ratio. If no validation is selected, the water extraction results are directly output.

[0070] The database storage module in step 3 includes: FCN, PSPNET, U-NET, Deeplabv3 series, Segformer or MFM-SegFormer model.

[0071] Step 3 uses an algorithm processing module to process the data. The specific steps are as follows:

[0072] The water extraction algorithm MFM-SegFormer is adopted. The multi-scale feature fusion module MFM first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules for multi-scale feature map extraction. The first scale feature extraction module consists of a pooling layer, a convolutional layer, a batch normalization (BN) layer, a ReLU layer, and an overlay layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is located in four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an Overlap Patch. The Mix-FNN layer consists of a Merging layer, where the Mix-FNN layer is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly.

[0073] Because existing water body extraction algorithms are not integrated, the extraction process is quite cumbersome. Furthermore, different algorithms produce varying results and take different processing times depending on the image. Therefore, this invention proposes a surface water extraction system based on artificial intelligence algorithms. This system is designed for abundant remote sensing data and incorporates processing functions for various remote sensing data, including Resource series, High Resolution series, Landsat series, and UAV aerial photography. It offers selectable band combination generation, allowing users to choose their preferred band combination images. The system includes the optimal network model for all band combinations and current mainstream image segmentation models. It also provides a function to create labeled datasets based on user data, which can be used to evaluate the accuracy of the water body extraction results.

[0074] With the development of artificial intelligence technology, there are increasingly more image classification algorithms, and their accuracy is constantly improving. However, the integration level of current algorithm models is relatively low. Because the algorithms themselves are highly specialized and difficult to understand quickly, and their usage is complex, requiring environment configuration and parameter modification based on computing power in the early stages, this clearly cannot meet the needs of rapid and high-precision water extraction tasks. This invention extracts water using a deep learning image classification method, encapsulating the necessary functions for water extraction into corresponding modules, thus realizing a complete water extraction system. This end-to-end water extraction system not only saves time, manpower, and financial costs, but also has a wide user base, high accuracy in water extraction, and strong technical robustness, facilitating the efficient conduct of water resource research tasks in my country.

[0075] Experimental Analysis

[0076] The research area of ​​this invention is the Wei River Basin, covering an area of ​​103°57'-110°16'E and 33°42'-37°24'N, spanning Shaanxi, Gansu, and Ningxia Hui Autonomous Regions, with a total area of ​​approximately 134,766 square kilometers. The Wei River Basin has a complex topography, with mountains, hills, plateaus, and plains accounting for 15%, 49%, 19%, and 17% of the total basin area, respectively. The Wei River Basin is an important biodiversity conservation area and a vital ecological barrier in my country.

[0077] like Figure 9 As shown, this embodiment of the invention selected several typical scenarios to qualitatively compare and analyze the extraction results under different conditions. Each column in the figure represents a different scenario. This embodiment of the invention selected a total of 7 scenarios. The first row is the image of the real scene, the second row is the labeled image of each scenario, the third to sixth rows are the prediction results of the comparison model for each scenario, and the seventh row is the prediction result of the MFM-Segformer model. Meanwhile, yellow dashed lines indicate water bodies misclassified as other, and red dashed lines indicate other misclassified as water bodies. The results show that each model has a certain degree of error. The MFM-SegFormer model has the best extraction results, with only a very small area being misclassified, and the highest integrity of the extracted water bodies. In these scenarios, FCN, PSPNet, SETR, and DeeplabV3+ performed poorly in extracting small water bodies.

[0078] It can be seen that the water extraction algorithm MFM-Segformer used in this invention has the best water extraction performance. This may be due to SegFormer's self-attention mechanism, which allows it to focus on the features that need attention without losing global information, and the use of the multimodal feature fusion module MFM to improve network performance. Overall, the input data has a significant impact on the model's water extraction, and the water extraction data extracted by the MFM-SegFormer algorithm of this invention in different scenarios is more accurate than that extracted by FCN, PSPNet, SETR, and DeeplabV3+.

Claims

1. A surface water extraction system based on artificial intelligence algorithms, comprising a back-end processing system and a front-end operating system, wherein each module in the front-end operating system sends data streams and corresponding parameters to the back-end processing system, which processes the data and then sends it back to the front-end operating system, characterized in that: The backend processing system includes an algorithm processing module and a database storage module; the frontend operating system includes a user login module, a data selection module, a model selection module, and a verification module. The user login module verifies the user's identity information. After successful verification, the user enters the main interface of the surface water extraction system. The user inputs the remote sensing data of the water body to be extracted through the data selection module. The database storage module provides the model. The remote sensing data of the water body to be extracted from the data selection module and all the models in the database storage module are sent to the model selection module. After the model selection module selects a model, the algorithm processing module performs water extraction according to the selected model and sends the water extraction results to the verification module. The verification module verifies and outputs the water extraction results and accuracy indicators. Alternatively, the user can choose not to perform verification and directly output the water extraction results. The algorithm processing module incorporates the water extraction algorithm MFM-SegFormer. The multi-scale feature fusion module MFM first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules for multi-scale feature map extraction. The first scale feature extraction module consists of a Pooling layer, a convolutional layer, a Batch Normalization (BN) layer, a ReLU layer, and an Upsampling layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is integrated into four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an Overlap Patch. The Merging layer consists of a Mix-FNN layer, which is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly. The database storage module is used to store water body extraction model data that can be used by the model selection module of the front-end operating system.

2. The surface water extraction system based on artificial intelligence algorithm according to claim 1, characterized in that: include: The user login module is used for user registration and login; The data selection module is used to select the band combination and image stretching method of the remote sensing image; The model selection module is used to select the water extraction model and the specific parameters of the model; The verification module is used to verify the water extraction results.

3. The surface water extraction system based on artificial intelligence algorithm according to claim 1, characterized in that: The database storage module includes the MFM-SegFormer model.

4. The surface water extraction system based on artificial intelligence algorithm according to claim 1, characterized in that: The database storage module also includes: FCN, PSPNET, U-NET, Deeplabv3 series or Segformer.

5. A method for extracting surface water bodies based on artificial intelligence algorithms, characterized in that: Includes the following steps: Step 1: Users register and log in through the user login module of the front-end operating system; Step 2: Input the remote sensing data of the water body to be extracted through the data selection module, open the remote sensing image of the water body to be extracted, select the data and corresponding parameters through the data selection module, and select the band combination and image stretching method of the remote sensing image. Step 3: Use the model selection module to determine the model for water body extraction. The back-end processing system adjusts the band combination of the remote sensing data according to the corresponding parameters, calls the corresponding model in the database storage module, and uses the algorithm processing module to process the data to obtain the water body extraction result. Step 4: The water extraction results obtained in Step 3 are evaluated by the validation module. If validation is selected, the accuracy of different results is evaluated by precision, recall, F1 score and cross-validation ratio. If no validation is selected, the water extraction results are output directly. Step 3 uses an algorithm processing module to process the data; the specific steps are as follows: The water extraction algorithm MFM-SegFormer is adopted. The multi-scale feature fusion module MFM first feeds the deep feature maps extracted by the Transformer network into five parallel multi-scale feature extraction modules for multi-scale feature map extraction. The first scale feature extraction module consists of a Pooling layer, a convolutional layer, a Batch Normalization (BN) layer, a ReLU layer, and an Upsampling layer. The other four scale feature extraction modules each consist of a convolutional layer, a BN layer, and a ReLU layer with different strides and kernel sizes. After these five parallel feature extractions at different scales, the features are stacked and fed into the final feature integration module for feature map stacking. The final feature integration module consists of a 1*1 convolutional layer, a BN layer, and a ReLU layer. The MFM layer is located in four serially connected Transformer Blocks. Each Transformer Block consists of an Efficient Self-Attention layer, a Mix-FNN layer, and an OverlapPatch. The Mix-FNN layer consists of a Merging layer, where the Mix-FNN layer is composed of an MLP layer, a 3*3 convolutional layer, a GELU layer, and another MLP layer connected linearly.

6. The method for extracting surface water bodies based on artificial intelligence algorithms according to claim 5, characterized in that: The model in the database storage module of step 3 includes the MFM-SegFormer model.

7. The method for extracting surface water bodies based on artificial intelligence algorithms according to claim 5, characterized in that: The models in the database storage module of step 3 also include: FCN, PSPNET, U-NET, Deeplabv3 series or Segformer.

Images