A semi-autonomous remote sensing water area generalization segmentation method fusing multi-spectral inversion

Abstract

The application discloses a kind of fusion multispectral inversion's semi-autonomous remote sensing water area generalization segmentation method, it is related to remote sensing image processing technical field, specific implementation scheme is as follows: acquisition remote sensing multispectral data;Remote sensing multispectral data and coordinate point prompt information are used as the input of end-to-end model, after being processed through multispectral inversion branch, generalization segmentation branch, cross-modal feature fusion, generate optimal water area segmentation mask;After morphological filtering and vector conversion, the optimal water area segmentation mask is converted into high-precision remote sensing water area segmentation vector result.The application embeds multispectral inversion and generalization segmentation depth into the same model structure, realizes the cooperation of spectral physical priori and semantic feature by three-stage progressive fusion, and utilizes GPU parallel computing optimization framework, significantly improves the labeling accuracy and calculation efficiency.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a semi-autonomous remote sensing water area generalization segmentation method that integrates multispectral inversion. Background Technology

[0002] Remote sensing water area labeling is a core technology for water resource management, disaster monitoring, and ecological protection. Currently, it mainly relies on two types of methods: traditional methods based on multispectral inversion and semantic segmentation methods based on deep learning.

[0003] Traditional methods based on multispectral inversion construct water indices (such as Normalized Difference Water Index (NDWI) and Modified NDWI (MNDWI)), design water indices using the spectral physical properties of water bodies, and highlight the differences between water bodies and non-water bodies by calculating the differences or ratios between spectral bands. Combined with fixed or adaptive threshold segmentation, water area extraction is achieved. However, there are problems with threshold setting being highly dependent on experience or local data statistics, and there are limitations in generalization ability.

[0004] Semantic segmentation methods based on deep learning, such as fully convolutional networks like U-Net and DeepLab, improve the accuracy of water segmentation in complex scenes by learning deep semantic features from remote sensing images. However, existing technologies suffer from the problems of requiring large amounts of data for neural network model training and low computational efficiency. Furthermore, the model learning process only emphasizes pixel information from remote sensing images without constraining the spectral physical properties of water bodies, resulting in output results that contradict these properties. Summary of the Invention

[0005] Based on this, this application addresses the problems of "neighborhood model training relying on a large amount of data" and "insufficient utilization of spectral information" in existing technologies by providing a semi-autonomous remote sensing water area generalization segmentation method that integrates multispectral inversion. The core innovation of this method lies in: designing a novel model structure that embeds water indicators into an end-to-end semantic segmentation model; effectively reducing the model's training data requirements through the prior knowledge inherent in the water indicators; and simultaneously improving the accuracy and generalization ability of the semantic segmentation results by using the spectral physical characteristics of the water indicators as constraints.

[0006] This application provides a semi-autonomous remote sensing water area generalization segmentation method that integrates multispectral inversion, including:

[0007] Collect remote sensing multispectral data;

[0008] Remote sensing multispectral data and coordinate point cue information are used as inputs to the end-to-end model. After multispectral inversion branch processing, spectral features containing the spectral physical properties of water bodies are obtained. After generalization segmentation branch processing, semantic features containing pixel semantic information and cue features containing coordinate point cue information are obtained. Finally, the spectral features, semantic features, and cue features are combined and cross-modal feature fusion processing is performed to generate the optimal water body segmentation mask that conforms to the spectral physical properties of water bodies and pixel semantic information.

[0009] After morphological filtering and vector transformation, the water area segmentation mask is converted into a high-precision remote sensing water area segmentation vector result.

[0010] The remote sensing multispectral data includes a total of 11 bands of multispectral remote sensing images, consisting of 3 visible light bands and 8 non-visible light bands.

[0011] The multispectral inversion branching process includes parallel water index calculation and feature stitching and preprocessing;

[0012] Parallel water index calculation specifically includes,

[0013] An independent channel parallel computing framework is adopted to perform synchronous calculations on n water indices, and each water index generates an independent single-channel feature map. The hardware acceleration of the calculation process is based on the CUDA (Compute Unified Device Architecture) parallel computing framework, which encapsulates the band operation of each water index into an independent kernel function and utilizes the thread-level parallel characteristics of GPU (Graphics Processing Unit) to achieve efficient calculation.

[0014] Feature splicing and preprocessing specifically include,

[0015] The n single-channel feature maps are concatenated into a multispectral prior feature tensor of H×W×n according to the channel dimension, and the scale is dynamically adjusted by using a learnable scaling factor and offset to adjust the feature value range. Then, a stacked model of convolution-normalization-activation is used to extract spectral features through multiple convolutions, where H represents the image height of the multispectral image and W represents the image width of the multispectral image.

[0016] The generalized splitting branching process includes...

[0017] The generalized segmentation branch consists of an image encoder and a coordinate point cue encoder, which are used to extract semantic features and cue features of the image, and combine coordinate point cue information to semi-autonomously correct the output of the image encoder to optimize the semantic features.

[0018] Visible light spectral images are processed by an image encoder to obtain semantic features.

[0019] The coordinate point prompt encoder includes two steps: position encoding and weight superposition.

[0020] Input the two-dimensional pixel coordinates [x, y] of the point provided by the user on the multispectral image;

[0021] Location encoding utilizes a random Gaussian matrix and trigonometric functions to encode the coordinates, scaling the two-dimensional coordinates [x, y] to the range [0, 1]. Then, a Gaussian matrix [N, 128] is used to map the two-dimensional coordinates to 128 dimensions. Finally, the mapping results are processed using sine and cosine functions respectively, and then concatenated into a 256-dimensional vector. Here, N represents the number of points provided by the user, x represents the column index of the user-provided point in the multispectral image, and y represents the row index of the user-provided point in the multispectral image. A weighted table is then used to process the 256-dimensional vector to obtain the prompt features.

[0022] The operations of cross-modal feature fusion processing include,

[0023] The aligned semantic and spectral features are channel-concatenated to form a fused feature tensor, which is then compressed to 256 channels using a 1×1 convolutional layer with a stride of 1. A bidirectional attention converter is used to achieve deep bidirectional interaction between the cue features and the fused feature tensor. A confidence-optimized mask generator is then used to generate four candidate water area masks. The confidence of the four candidate masks is calculated, and the four candidate masks are predicted through three multi-sensor layers. The confidence scores of the four masks are output, and the optimal water area mask is selected through a weighted voting mechanism.

[0024] The specific operations for converting the optimal water segmentation mask into high-precision remote sensing water segmentation vector results through morphological filtering and vector transformation include:

[0025] Morphological filtering is applied to the optimal water area mask to eliminate small noise and holes in the mask through dilation and erosion processing, thereby enhancing the smoothness and integrity of the water body boundary. Then, a vector transformation algorithm is used to convert the mask into vector polygon data, and each water area object is assigned geographic coordinates and attribute information to form a high-precision remote sensing water area segmentation vector result.

[0026] Beneficial effects: This application deeply embeds multi-band inversion and generalized segmentation into the same model structure, achieves synergy between spectral physical priors and semantic features through three-level progressive fusion, and utilizes GPU parallel computing to optimize the framework, significantly improving annotation accuracy and computational efficiency.

[0027] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0028] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein:

[0029] Figure 1 This is a schematic diagram of the overall structure of the model provided in this application;

[0030] Figure 2 This is a schematic diagram of the multispectral inversion branch structure provided in this application;

[0031] Figure 3 This is a schematic diagram of the generalized splitting branch structure provided in this application;

[0032] Figure 4 This is a schematic diagram of the cross-modal feature fusion structure provided in this application;

[0033] Figure 5 It is a time distribution map of multispectral data provided in this application;

[0034] Figure 6 This is a graph of the training loss function based on the model provided in this application;

[0035] Figure 7 It is based on the original water area of ​​the main river scene provided in this application;

[0036] Figure 8 It is based on the actual water area images of the main river scene provided in this application;

[0037] Figure 9 The images are based on the SAM semi-autonomous segmentation effect images of the main river scene provided in this application;

[0038] Figure 10 It is a semi-autonomous segmentation effect image of the main river scene water area provided in this application;

[0039] Figure 11 It is based on the original water area image of the reservoir scene provided in this application;

[0040] Figure 12 It is based on the actual water area images of the reservoir scene provided in this application;

[0041] Figure 13 The image is a semi-autonomous segmentation (SAM) effect image of the water area in the reservoir scene provided in this application;

[0042] Figure 14 It is a semi-autonomous segmentation effect image of the water area in the reservoir scene provided in this application;

[0043] Figure 15 It is based on the original water area image of the small river scene provided in this application;

[0044] Figure 16It is based on the actual water area images of the small river scene provided in this application;

[0045] Figure 17 The image is a semi-autonomous segmentation result image of the water area in the small river scene provided in this application using SAM.

[0046] Figure 18 The image is a semi-autonomous segmentation effect image of a small river scene provided in this application. Detailed Implementation

[0047] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0048] like Figure 1 As shown, this application provides a semi-autonomous remote sensing water area generalization segmentation method that integrates multispectral inversion, including:

[0049] S1: Acquire remote sensing multispectral data. It should be noted that:

[0050] The remote sensing multispectral data includes a total of 11 bands of multispectral remote sensing images, consisting of 3 visible light bands and 8 non-visible light bands.

[0051] S2: Remote sensing multispectral data and coordinate point cue information are used as input to the end-to-end model. After multispectral inversion branch processing, spectral features containing the spectral physical properties of the water body are obtained. After generalization segmentation branch processing, semantic features containing pixel semantic information and cue features containing coordinate point cue information are obtained. Finally, the spectral features, semantic features, and cue features are combined through cross-modal feature fusion processing to generate the optimal water body segmentation mask that conforms to the spectral physical properties of the water body and the pixel semantic information. It should be noted that:

[0052] The multispectral inversion branch, serving as the core unit for generating spectral priors in the model, employs a parallel multi-channel water index calculation and feature fusion architecture. Its role is to transform the spectral prior knowledge of water bodies into structured features that can be efficiently utilized by deep learning models through parallel computation of multiple water indices and feature fusion. When labeled data is limited, the strong spectral priors provided by multi-dimensional water index features can replace some manually labeled information, reducing the model's dependence on large-scale training data.

[0053] The multispectral inversion branching process includes parallel water index calculation and feature stitching and preprocessing;

[0054] Parallel water index calculation specifically includes,

[0055] An independent channel parallel computing framework is adopted to perform synchronous calculations on n water indices, and each water index generates an independent single-channel feature map; the hardware acceleration of the calculation process is based on the CUDA parallel computing framework, which encapsulates the band operation of each water index into an independent kernel function and utilizes the thread-level parallelism of the GPU to achieve efficient calculation.

[0056] like Figure 2 As shown in the figure, [H,W,11] represents a multispectral image with an image height of H and an image height of W, and 11 spectral bands. The 11 spectral bands consist of 3 visible light bands and 8 non-visible light bands, corresponding to the sentinel data source in the embodiment.

[0057] This application employs an independent channel parallel computing architecture to simultaneously calculate five water indices (NDWI, MNDWI, WNDWI, EWI, and AWEI), generating an independent single-channel feature map (size [H, W, 1]) for each water index.

[0058] NDWI (Normalized Difference Water Index): Enhances water body identification in vegetated areas, suppresses the influence of chlorophyll reflection, and is suitable for densely vegetated river and lake areas.

[0059] MNDWI (Improved NDWI): Enhances the depiction of water body boundaries with complex backgrounds such as buildings and soil, reduces shadow interference, and is suitable for urban flooding and port waters;

[0060] WNDWI (Weighted Normalized Difference Water Index): Optimizes sensitivity to water bodies containing suspended solids and improves the identification accuracy of turbid water bodies (such as estuaries);

[0061] EWI (Enhanced Water Index): It highlights weak water body signals through multi-band combination and is suitable for detecting water bodies on ice surfaces and under thin cloud cover;

[0062] AWEI (Automated Water Extraction Index): Suppresses soil background noise and is more sensitive to water bodies on exposed surfaces (such as seasonal waterlogging in dried-up river channels and irrigated farmland).

[0063] The hardware acceleration of the computation process is based on the CUDA parallel computing framework, which encapsulates the band operation of each water index into an independent kernel function and utilizes the thread-level parallelism of the GPU to achieve efficient computation.

[0064] Feature splicing and preprocessing specifically include,

[0065] The n single-channel feature maps are concatenated into a multispectral prior feature tensor of H×W×n according to the channel dimension, and the scale is dynamically adjusted by using a learnable scaling factor and offset to adjust the feature value range. Then, a stacked model of convolution-normalization-activation is used to extract spectral features through multiple convolutions, where H represents the image height of the multispectral image and W represents the image width of the multispectral image.

[0066] In this application, five single-channel water index feature maps are first stitched together according to the channel dimension to form a multispectral prior feature tensor of H×W×5. The differential response of different water indices to the same region is preserved, forming a complementary spectral feature representation. This avoids the scene limitations of a single water index and improves the water body representation capability of complex scenes through multi-dimensional feature combination.

[0067] Then, dynamic scaling is performed, using learnable scaling factors and offsets to adjust the range of feature values, in order to avoid gradient vanishing or gradient explosion problems during subsequent multi-layer convolution feature extraction.

[0068] Finally, a stacked "convolution-normalization-activation" module is used to extract features step by step through multiple convolutions, which also facilitates subsequent fusion with features from the image coding branch. Its structure is as follows:

[0069] Group 1: Convolution (kernel width 2, stride 2) → Layer normalization → Gaussian error linear unit

[0070] Convolution (kernel width 2, stride 2): Reduces the spatial size of the input by half, quickly reducing the spatial dimension to accommodate the computational load of subsequent modules, while using the convolution kernel to extract basic spatial features.

[0071] Layer normalization: Normalize the channel dimension of each sample to stabilize the training process, alleviate gradient vanishing / exploding, and make the feature distribution of different channels more consistent.

[0072] Gaussian error linear unit: Introduces nonlinearity, allowing the model to learn more complex feature relationships.

[0073] The second group: Convolution (kernel width 2, stride 2) → layer normalization → Gaussian error linear unit: Its significance lies in deepening feature extraction, allowing the spatial patterns of data (such as complex contours) to be gradually abstracted by multiple layers of convolution, generating more representative features.

[0074] Group 3: Convolution (kernel width 1, stride 1)

[0075] Convolution (kernel width 1, stride 1): "Compresses" the features extracted from the first two layers into a unified 256-dimensional embedding space, ensuring compatibility with the output dimension of the image encoder and facilitating subsequent fusion.

[0076] The generalized splitting branching process includes...

[0077] The generalized segmentation branch consists of an image encoder and a coordinate point cue encoder, which are used to extract semantic features and cue features of the image, and combine coordinate point cue information to semi-autonomously correct the output of the image encoder to optimize the semantic features.

[0078] Visible light spectral images are processed by an image encoder to obtain semantic features.

[0079] like Figure 3 As shown in this application, the image encoder employs the RepVit (Reparameterized Visual Transformer) structure to extract generalized semantic features from visible light spectral images. This structure innovatively integrates the multi-head self-attention mechanism of the Transformer module with convolutional layers through structural reparameterization technology. In the initial stage, RepVit uses lightweight convolutional layers to perform preliminary processing on the input visible light spectral image data, quickly extracting local spatial features and spectral details from the image. This design effectively preserves high-frequency information of water body boundaries and subtle differences at the water-land interface. The RepVit structure supports multi-scale feature fusion; shallow feature maps retain rich spatial details, aiding in accurate water body boundary localization; deep feature maps extract high-dimensional semantic information, effectively distinguishing between water and non-water regions, especially enhancing the model's understanding of water semantics in complex backgrounds.

[0080] The coordinate point cue encoder consists of two steps: position encoding and weight stacking. The core is to convert sparse point coordinates into 256-dimensional embedding vectors that are compatible with the model.

[0081] The coordinate point prompt encoder includes two steps: position encoding and weight superposition.

[0082] Input the two-dimensional pixel coordinates [x, y] of the point provided by the user on the multispectral image;

[0083] Location encoding utilizes a random Gaussian matrix and trigonometric functions to encode the coordinates, scaling the two-dimensional coordinates [x, y] to the range [0, 1]. Then, a Gaussian matrix [N, 128] is used to map the two-dimensional coordinates to 128 dimensions. Finally, the mapping results are processed using sine and cosine functions respectively, and then concatenated into a 256-dimensional vector. Here, N represents the number of points provided by the user, x represents the column index of the user-provided point in the multispectral image, and y represents the row index of the user-provided point in the multispectral image. A weighted table is then used to process the 256-dimensional vector to obtain the prompt features.

[0084] The operations of cross-modal feature fusion processing include,

[0085] The aligned semantic and spectral features are channel-concatenated to form a fused feature tensor, which is then compressed to 256 channels using a 1×1 convolutional layer with a stride of 1. A bidirectional attention converter is used to achieve deep bidirectional interaction between the cue features and the fused feature tensor. A confidence-optimized mask generator is then used to generate four candidate water area masks. The confidence of the four candidate masks is calculated, and the four candidate masks are predicted through three multi-sensor layers. The confidence scores of the four masks are output, and the optimal water area mask is selected through a weighted voting mechanism.

[0086] The cross-modal feature fusion module achieves deep fusion of spectral priors and semantic features through a three-level processing flow of "feature interaction - dynamic weighting - mask optimization".

[0087] like Figure 4 As shown, the semantic features output by the aligned image coding branch (generated by the RepVit encoder, size [H / 4, W / 4, 256]) and the spectral features output by the multispectral inversion branch (fused by the multi-water index and then normalized, size [H / 4, W / 4, 256]) are concatenated to form a fused feature tensor of [H / 4, W / 4, 512]. This tensor is then compressed to 256 channels through a 1×1 convolutional layer with a stride of 1. This process preserves key information while reducing subsequent computational complexity, providing efficient input for bidirectional attention interaction.

[0088] An improved bidirectional attention converter structure is employed to achieve bidirectional guidance of semantic and spectral features:

[0089] Semantic guidance path: The semantic information encoded in the image (such as the spatial association between water bodies and vegetation) is injected into the spectral features through the self-attention mechanism to suppress false spectral responses in non-water areas (such as high NDWI values ​​of building shadows).

[0090] Spectral Constraint Path: By utilizing the cross-attention mechanism, the activation intensity of the water body region in the semantic features is enhanced by using the water body spectral prior in the multispectral features as the weight, thereby improving the feature discrimination of the mixed pixel region.

[0091] S3: After morphological filtering and vector transformation, the optimal water segmentation mask is converted into a high-precision remote sensing water segmentation vector result. It should be noted that:

[0092] The specific operations for converting the water area segmentation mask into high-precision remote sensing water area segmentation vector results through morphological filtering and vector transformation include:

[0093] Morphological filtering is applied to the optimal water area mask to eliminate small noise and holes in the mask through dilation and erosion processing, thereby enhancing the smoothness and integrity of the water body boundary. Then, a vector transformation algorithm is used to convert the mask into vector polygon data, and each water area object is assigned geographic coordinates and attribute information to form a high-precision remote sensing water area segmentation vector result.

[0094] The embodiments use Sentinel multispectral remote sensing imagery as the data source. Specific experiments are conducted to compare and verify the accuracy and efficiency of the method described in this application for segmenting remotely sensed water areas. The specific implementation steps and result analysis are as follows:

[0095] Model training:

[0096] This embodiment uses 1106 Sentinel multispectral remote sensing images as training and testing data, with a total data size of 893GB and a coverage area of ​​approximately 9.98 million square kilometers. The data acquisition period spans from October 2022 to March 2025. Due to seasonal variations and cloud cover, the number of images acquired from August to October is relatively small (see details for seasonal distribution). Figure 5 All data meet the following technical specifications: cloud cover less than 3%, spatial resolution up to 10 meters, and includes multispectral bands such as B2 (blue), B3 (green), B4 (red), and B8 (near-infrared), which can truly reflect the spectral characteristics of water bodies under different seasons and geographical environments.

[0097] Figure 6 The model was trained for 200 epochs, and the loss function curve was plotted. The loss function value gradually converged to 0 as the number of training epochs increased, indicating that the model had learned the data features defined in the training dataset.

[0098] Segmentation effect comparison:

[0099] This embodiment will compare the segmentation effect of the method provided in this application with other semi-autonomous segmentation algorithms in remote sensing water segmentation, and perform water segmentation on the test remote sensing images in the following ways:

[0100] Baseline method: Semi-autonomous segmentation based on the SAM model (requires manual input of a few prompts to assist segmentation);

[0101] The semi-autonomous segmentation method of this application: The model provided in this application integrates multi-band inversion and generalized segmentation (the input is the same as the prompt points of SAM);

[0102] Figures 7 to 10The images show the original water area of ​​the main river scene, the ground truth image, the SAM semi-autonomous segmentation image used in the benchmark method, and the semi-autonomous segmentation image proposed in this application. As can be seen from the images, the SAM semi-autonomous segmentation result has many omissions in the details of the river edge and a large difference from the ground truth in terms of details. The semi-autonomous segmentation result proposed in this application performs better in terms of the details of the river edge, with no major omissions and only a few edge roughness issues, which is closer to the real data.

[0103] Figures 11 to 14 The images show the original water area of ​​the reservoir scene, the ground truth image, the SAM semi-autonomous segmentation image used in the benchmark method, and the semi-autonomous segmentation image proposed in this application. As shown in the figure, the SAM semi-autonomous segmentation completely misses the identification of the reservoir's inlet and outlet rivers, resulting in a significant discrepancy with the actual situation. The semi-autonomous segmentation result proposed in this application is more consistent with the actual value.

[0104] Figures 15 to 18 The images show the original water area of ​​a small river scene, the ground truth image, the SAM semi-autonomous segmentation image used in the benchmark method, and the semi-autonomous segmentation image proposed in this application. As shown in the figure, the river channel obtained by SAM semi-autonomous segmentation is too smooth. Although it covers the area described by the actual river situation, there are too many extra areas, which does not conform to the actual situation. In contrast, the semi-autonomous segmentation result proposed in this application is closer to the actual value, with only a small amount of extra structure at the edge details.

[0105] In three scenarios—main rivers, reservoirs, and small rivers—the semi-autonomous segmentation effect of SAM was compared with that of the semi-autonomous segmentation effect described in this application. The method provided in this application is more refined in the processing of water area boundaries, can capture more detailed water area details, and is more in line with the real situation of the image in terms of shape.

[0106] Table 1: Comparison of Segmentation Metrics by Scene

[0107]

[0108] Furthermore, the workload of manual annotation and semi-autonomous segmentation methods was compared in three scenarios: main rivers, reservoirs, and small rivers.

[0109] Manual annotation: Professionals use traditional GIS software to vectorize the water boundary and record the number of coordinate points that need to be manually entered during the annotation process;

[0110] The semi-autonomous segmentation method of this application: On the same image, the semi-autonomous segmentation function of the model described in this application is used, and the model prediction and raster vectorization post-processing operation described in this application are performed. The number of coordinate points required is recorded when the water boundary vector result generated by the model is basically consistent with the manual annotation.

[0111] Table 2: Comparison of the number of manually annotated points and the number of semi-autonomous segmentation prompts

[0112]

[0113] As shown in Table 2, manual annotation of a single scene image requires an average of about 100 coordinate points to be manually entered, while the semi-autonomous method of this application only requires 2-3 prompt points to complete high-quality segmentation, reducing the coordinate point requirement by more than 97%.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A semi-autonomous remote sensing water body generalized segmentation method fusing multispectral inversion, characterized in that, The method comprises the following steps: Collecting remote sensing multispectral data; Taking the remote sensing multispectral data and the coordinate point prompt information as inputs of an end-to-end model, processing the remote sensing multispectral data through a multispectral inversion branch to obtain spectral features containing spectral physical characteristics of water bodies, processing the remote sensing multispectral data through a generalization segmentation branch to obtain semantic features containing pixel semantic information, and processing the coordinate point prompt information to obtain prompt features, and finally combining the spectral features, the semantic features and the prompt features through cross-modal feature fusion processing to generate an optimal water area segmentation mask conforming to the spectral physical characteristics of the water bodies and the pixel semantic information. The multispectral inversion branch processing comprises parallel water index calculation and feature splicing and preprocessing. The parallel water index calculation specifically comprises the following steps: An independent channel parallel calculation framework is adopted to synchronously calculate n water indexes, and an independent single-channel feature map is generated for each water index. The hardware acceleration in the calculation process is based on a CUDA parallel calculation framework, each water index band operation is encapsulated as an independent kernel function, and GPU thread-level parallel characteristics are used to realize efficient calculation. The cross-modal feature fusion processing comprises the following steps: The aligned semantic features and the spectral features are concatenated at the channel level to form a fusion feature tensor, which is compressed to 256 channels through a 1x1 convolution layer with a step of 1; a bidirectional attention converter is used to realize deep bidirectional interaction between the prompt features and the fusion feature tensor; a confidence optimization mask generator is used to first generate four candidate water area masks, calculate the confidence of the four candidate masks, predict the four candidate masks through three multiple perception layers, output four confidence scores, and select the optimal water area mask through a weighted voting mechanism; The optimal water area segmentation mask is converted into a high-precision remote sensing water area segmentation vector result through morphological filtering and vector conversion.

2. The semi-autonomous remote sensing water body generalized segmentation method of claim 1, wherein: The remote sensing multispectral data comprises 11-band multispectral remote sensing images including 3 visible light bands and 8 non-visible light bands.

3. The semi-autonomous remote sensing water area generalization segmentation method of claim 2, wherein the feature splicing and preprocessing specifically comprises the following steps: The n single-channel feature maps are spliced into a multispectral prior feature tensor of HxWxn according to the channel dimension, and dynamic scale adjustment is performed through a learnable scaling factor and an offset to adjust the feature value range; a convolution-normalization-activation stacked model is used to extract spectral features through multiple convolution layers, wherein H represents the image height of the multispectral image, and W represents the image width of the multispectral image. The generalization segmentation branch processing comprises the following steps:

4. The semi-autonomous remote sensing water body generalized segmentation method of claim 3, wherein: The generalization segmentation branch is composed of an image encoder and a coordinate point prompt encoder, which are used to extract semantic features and prompt features, and combine the coordinate point prompt information to semi-autonomously correct the output results of the image encoder to optimize the semantic features. The visible light spectral image is processed by the image encoder to obtain the semantic features.

5. The semi-autonomous remote sensing water body generalized segmentation method of claim 4, wherein: The coordinate point prompt encoder comprises two steps of position encoding and weight superposition.

6. The semi-autonomous remote sensing water body generalized segmentation method of claim 5, wherein: The two-dimensional pixel coordinates [x, y] of the point provided by the user on the multispectral image are input. ​ The position coding encodes the two-dimensional coordinates [x, y] into the range of [0, 1] by using a random Gaussian matrix and a trigonometric function, maps the two-dimensional coordinates into 128 dimensions by means of a Gaussian matrix [N, 128], and finally processes the mapping result by using a sin function and a cos function respectively, and splices them into a 256-dimensional vector, wherein N represents the number of points provided by the user, x represents the column index of the points provided by the user on the multispectral image, and y represents the row index of the points provided by the user on the multispectral image. The 256-dimensional vector is processed by using a weight table to obtain a prompt feature.

7. The semi-autonomous remote sensing water body generalized segmentation method of claim 1 or 6, wherein: The specific operation of converting the optimal water area segmentation mask into a high-precision remote sensing water area segmentation vector result through morphological filtering and vector conversion includes, The morphological filtering operation is adopted for the optimal water area mask, small noise points and holes in the mask are eliminated through expansion and corrosion processing, and the smoothness and integrity of the water body boundary are enhanced; then the mask is converted into vector polygon data by using a vector conversion algorithm, each water area object is given geographic coordinates and attribute information, and a high-precision remote sensing water area segmentation vector result is formed.

Images