Semi-autonomous remote sensing water area generalization segmentation method fused with multispectral inversion

Abstract

The invention discloses a semi-autonomous remote sensing water area generalization segmentation method fused with multispectral inversion, and relates to the technical field of remote sensing image processing, and the specific implementation scheme is as follows: collecting remote sensing multispectral data; remote sensing multispectral data and coordinate point prompt information serve as input of an end-to-end model, and an optimal water area segmentation mask is generated after multispectral inversion branching, generalization segmentation branching and cross-modal feature fusion processing; and converting the optimal water area segmentation mask into a high-precision remote sensing water area segmentation vector result through morphological filtering and vector conversion. According to the method, multiband inversion and generalization segmentation depth are embedded into the same model structure, cooperation of spectral physical prior and semantic features is realized through three-stage progressive fusion, and the framework is optimized through GPU parallel calculation, so that the labeling precision and the calculation efficiency are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and in particular to a semi-autonomous remote sensing water area generalization segmentation method fusing multispectral inversion. BACKGROUND

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

[0003] The traditional method based on multispectral inversion designs a water index by constructing a water index (such as normalized difference water index NDWI, improved NDWI (MNDWI), etc.) using the spectral physical characteristics of water body, highlights the difference between water body and non-water body through difference or ratio operation between spectral bands, and realizes water area extraction by combining fixed threshold or adaptive threshold segmentation. However, this method has the problem that the threshold setting is highly dependent on experience or local data statistics, and has limitations in generalization ability.

[0004] The semantic segmentation method based on deep learning is represented by the fully convolutional network of U-Net and DeepLab. By learning the deep semantic features of remote sensing images, the water area segmentation accuracy in complex scenes is improved. However, the existing technology has the problems of requiring a large amount of data for neural network model training and low computing efficiency. Moreover, only the pixel information of the remote sensing image is emphasized in the model learning process, and the spectral physical characteristics of the water body are not constrained, resulting in the problem that the output result violates the spectral physical characteristics of the water body. SUMMARY

[0005] Therefore, the present application provides a semi-autonomous remote sensing water area generalization segmentation method fusing multispectral inversion to solve the problems of “neural network model training requiring a large amount of data” and “insufficient use of spectral information” in the prior art. The core innovation of the method is to design a new model structure and embed the water index into the end-to-end semantic segmentation model. Through the prior knowledge contained in the water index, the amount of training data required by the model can be effectively reduced. At the same time, the spectral physical characteristics of the water index are used as a constraint to improve the accuracy and generalization ability of the semantic segmentation result.

[0006] The present application provides a semi-autonomous remote sensing water area generalization segmentation method fusing multispectral inversion, which comprises:

[0007] Collecting remote sensing multispectral data;

[0008] The remote sensing multispectral data and the coordinate point prompt information are taken as inputs of an end-to-end model, the spectral features containing the spectral physical characteristics of the water body are obtained through a multispectral inversion branch, the semantic features containing the pixel semantic information and the prompt features containing the coordinate point prompt information are obtained through a generalization segmentation branch, and finally the optimal water area segmentation mask meeting the spectral physical characteristics of the water body and the pixel semantic information is generated through cross-modal feature fusion processing of the spectral features, the semantic features and the prompt features.

[0009] The water area segmentation mask is converted into a high-precision remote sensing water area segmentation vector result through morphological filtering and vector conversion.

[0010] The remote sensing multispectral data include 11-band multispectral remote sensing images of 3 visible light bands and 8 non-visible light bands.

[0011] The multispectral inversion branch processing includes parallel water index calculation and feature splicing and preprocessing.

[0012] The parallel water index calculation specifically includes,

[0013] An independent channel parallel computing framework is adopted to synchronously calculate n kinds of water indexes, and an independent single-channel feature map is generated for each water index; the calculation process is accelerated based on a CUDA (Compute Unified Device Architecture) parallel computing framework, each band operation of the water index is encapsulated as an independent kernel function, and GPU (Graphics Processing Unit) thread-level parallel characteristics are used to realize efficient calculation.

[0014] The feature splicing and preprocessing specifically includes,

[0015] The n single-channel feature maps are spliced into an HxWxn multispectral prior feature tensor according to the channel dimension, and dynamic scale adjustment is performed; the feature value range is adjusted through a learnable scaling factor and an offset; and then a convolution-normalization-activation stacked model is used to extract spectral features through multiple layers of convolution, wherein H represents the image height of the multispectral image, and W represents the image width of the multispectral image.

[0016] The generalization segmentation branch processing includes,

[0017] The generalization segmentation branch is composed of an image encoder and a coordinate point prompt encoder, and is used to extract semantic features and prompt features of an image, and to combine the coordinate point prompt information to semi-autonomously correct the output result of the image encoder to optimize the semantic features.

[0018] The visible light spectrum image is processed by the image encoder to obtain the semantic features.

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

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

[0021] The position coding uses a random Gaussian matrix and a trigonometric function to encode the coordinates to scale the two-dimensional coordinates [x, y] to the range of [0, 1], map the two-dimensional coordinates to 128 dimensions by means of the Gaussian matrix [N, 128], and finally process the mapping results with the sin function and the cos function respectively, and then splice them into a 256-dimensional vector, wherein N represents the number of points provided by the user, x represents the column index of the point provided by the user on the multispectral image, and y represents the row index of the point provided by the user on the multispectral image; and then the 256-dimensional vector is processed by using a weight table to obtain a prompt feature.

[0022] The operation of the cross-modal feature fusion processing includes,

[0023] The aligned semantic features and spectral features are cascaded 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; and a bidirectional attention converter is used to realize the deep bidirectional interaction between the prompt feature and the fusion feature tensor; and 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.

[0024] 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,

[0025] The morphological filtering operation is performed on the optimal water area mask to eliminate small noise points and holes in the mask through inflation and corrosion processing, and to enhance the smoothness and integrity of the water body boundary; then the vector conversion algorithm is used to convert the mask into vector polygon data, and each water area object is given geographic coordinates and attribute information to form a high-precision remote sensing water area segmentation vector result.

[0026] Beneficial effects: The application embeds the multi-band inversion and the generalized segmentation into the same model structure, realizes the cooperation of the spectral physical priori and the semantic features through three-level progressive fusion, and uses GPU parallel computing optimization framework to significantly improve the labeling accuracy and the calculation efficiency.

[0027] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the application, nor is it used to limit the scope of the application. Other features of the application will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings are used to better understand the present scheme and do not constitute a limitation on the present application. Among them:

[0029] Figure 1 is a schematic diagram of the overall structure of the model provided according to the present application;

[0030] Figure 2 is a schematic diagram of the multi-spectral inversion branch structure provided according to the present application;

[0031] Figure 3 is a schematic diagram of the generalization segmentation branch structure provided according to the present application;

[0032] Figure 4 is a schematic diagram of the cross-modal feature fusion structure provided according to the present application;

[0033] Figure 5 is a multi-spectral data time distribution diagram provided according to the present application;

[0034] Figure 6 is a model training loss function curve diagram provided according to the present application;

[0035] Figure 7 is a main river scene water area original picture provided according to the present application;

[0036] Figure 8 is a main river scene water area true value picture provided according to the present application;

[0037] Figure 9 is a main river scene water area SAM semi-autonomous segmentation effect picture provided according to the present application;

[0038] Figure 10 is a main river scene water area semi-autonomous segmentation effect picture provided according to the present application;

[0039] Figure 11 is a reservoir scene water area original picture provided according to the present application;

[0040] Figure 12 is a reservoir scene water area true value picture provided according to the present application;

[0041] Figure 13 is a reservoir scene water area SAM semi-autonomous segmentation effect picture provided according to the present application;

[0042] Figure 14 is a reservoir scene water area semi-autonomous segmentation effect picture provided according to the present application;

[0043] Figure 15 is a small river scene water area original picture provided according to the present application;

[0044] Figure 16It is a fine river scene water area true value picture provided according to the present application.

[0045] Figure 17 It is a fine river scene water area SAM semi-autonomous segmentation effect picture provided according to the present application.

[0046] Figure 18 It is a fine river scene water area semi-autonomous segmentation effect picture provided according to the present application. DETAILED DESCRIPTION

[0047] Exemplary embodiments of the present application are described herein below with reference to the accompanying drawings, which include various details of the embodiments of the present application to assist in the understanding thereof. They should, however, be taken as merely exemplary and therefore as not limiting the scope of the application. As such, those having ordinary skill in the art will appreciate that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Likewise, the description set forth herein describes particular embodiments, but those with ordinary skill in the art will appreciate that the scope of the application is not limited to these particular embodiments.

[0048] As shown in Figure 1 the present application provides a semi-autonomous remote sensing water area generalization segmentation method fusing multi-spectral inversion, comprising:

[0049] S1: Collect remote sensing multi-spectral data. It should be noted that:

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

[0051] S2: The remote sensing multi-spectral data and the coordinate point prompt information are taken as the input of the end-to-end model, the spectral features containing water body spectral physical characteristics are obtained through the multi-spectral inversion branch processing, the semantic features containing pixel semantic information and the prompt features containing coordinate point prompt information are obtained through the generalization segmentation branch processing, and finally the optimal water area segmentation mask conforming to the water body spectral physical characteristics and the pixel semantic information is generated through the cross-modal feature fusion processing combining the spectral features, the semantic features and the prompt features. It should be noted that:

[0052] The multi-spectral inversion branch as the spectral prior generation core unit of the model adopts parallel multi-channel water index calculation and feature fusion architecture. Its role is to convert the spectral prior knowledge of water body into structured features that can be efficiently utilized by deep learning model through multi-water index parallel calculation and feature fusion. When the labeled data is limited, the strong spectral prior provided by multi-dimensional water index features can replace part of the artificial labeling information, reducing the dependence of the model on large-scale training data.

[0053] The multi-spectral inversion branch processing includes parallel water index calculation and feature splicing and preprocessing;

[0054] The parallel water index calculation specifically includes,

[0055] Adopting an independent channel parallel computing framework, n water indices are calculated synchronously, 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 parallel characteristics of GPU to achieve efficient calculation.

[0056] like Figure 2 As shown in the figure, [H, W, 11] indicates that the image height of the multispectral image is H, the image height is W, the number of spectral bands is 11, and the 11 spectral bands are composed of 3 visible light bands and 8 non-visible light bands, corresponding to the sentinel data source in the embodiment.

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

[0058] NDWI (Normalized Difference Water Index): Enhances water body identification in vegetated areas and suppresses the influence of chlorophyll reflectance. It is suitable for river and lake areas with dense vegetation.

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

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

[0061] EWI (Enhanced Water Index): Emphasizes weak water signals through multi-band combination, suitable for detecting water bodies on ice and under light cloud cover;

[0062] AWEI (Automated Water Extraction Index): Suppresses soil background noise and is more sensitive to exposed surface water (such as seasonal waterlogging in dry rivers and irrigated farmland).

[0063] 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 GPU thread-level parallel characteristics to achieve efficient calculation.

[0064] Feature splicing and preprocessing specifically include:

[0065] The n single-channel feature maps are spliced into a multi-spectral prior feature tensor of HxWxn according to the channel dimension, and dynamic scale adjustment is performed, the range of feature values is adjusted through a learnable scaling factor and an offset, and then a convolution-normalization-activation stacked model is used to extract spectral features through multiple layers of convolution, wherein H represents the image height of the multi-spectral image, and W represents the image width of the multi-spectral image.

[0066] In the present application, first, five single-channel water index feature maps are spliced into a multi-spectral prior feature tensor of HxWx5 according to the channel dimension, the differentiated response of different water indexes to the same area is retained, a complementary spectral feature representation is formed, the scene limitation of a single water index is avoided, and the water body representation capability in a complex scene is improved through multi-dimensional feature combination.

[0067] Subsequently, dynamic scale adjustment is performed, the range of feature values is adjusted through a learnable scaling factor and an offset, so as to avoid the gradient vanishing or gradient explosion problem in subsequent continuous multi-layer convolution feature extraction.

[0068] Finally, a convolution-normalization-activation stacked module is used to gradually extract features through multiple layers of convolution, which facilitates subsequent fusion with the features of the image encoding branch. The structure is as follows:

[0069] First group: 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 reduces the spatial dimension, adapts the calculation amount of the subsequent module, and extracts basic spatial features using a convolution kernel.

[0071] Layer normalization: normalizes the channel dimension of each sample, stabilizes the training process, alleviates gradient vanishing / explosion, and makes the feature distribution of different channels more consistent.

[0072] Gaussian error linear unit: introduces nonlinearity to enable the model to learn more complex feature relationships.

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

[0074] Third group: convolution (kernel width 1, stride 1)

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

[0076] The generalization segmentation branch processing includes,

[0077] The generalization segmentation branch is composed of an image encoder and a coordinate point prompt encoder, which is used to extract semantic features and prompt features of the image, and combine the coordinate point prompt information to semi-autonomously correct the output results of the image encoder to optimize the semantic features.

[0078] The visible light spectral image is processed by the image encoder to obtain semantic features.

[0079] As shown in Figure 3 In the present application, the image encoder adopts a RepVit (reparameterization visual Transformer) structure to extract the generalization semantic features of the visible light spectral image. The structure innovatively fuses the multi-head self-attention mechanism of the Transformer module with the convolution layer through structural reparameterization technology in design. In the initial stage, RepVit uses a lightweight convolution layer to preliminarily process the input visible light spectral image data, quickly extracting local spatial features and spectral details in the image. This design can effectively preserve the high-frequency information of the water body boundary and the subtle differences at the water-land junction. The RepVit structure supports multi-scale feature fusion, and the shallow feature map retains rich spatial details, which helps to accurately locate the water body boundary; the deep feature map extracts high-dimensional semantic information, which can effectively distinguish water body and non-water body areas, especially in complex backgrounds, enhancing the model's understanding ability of water body semantics.

[0080] The coordinate point prompt encoder includes two steps of position encoding and weight superposition, and the core is to convert the sparse point coordinates into a 256-dimensional embedding vector compatible with the model.

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

[0082] The input is the two-dimensional pixel coordinates [x, y] of the points provided by the user on the multispectral image;

[0083] The position encoding uses a random Gaussian matrix and a trigonometric function to encode the coordinates to scale the two-dimensional coordinates [x, y] to the range [0, 1], and maps the two-dimensional coordinates to 128 dimensions through a Gaussian matrix [N, 128], and finally processes the mapping results with sin and cos functions respectively, and then splices them into a 256-dimensional vector, where 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. Then use the weight table to process the 256-dimensional vector to obtain the prompt features.

[0084] The operation of cross-modal feature fusion processing includes,

[0085] The aligned semantic features and spectral features are channel cascaded to form a fusion feature tensor, which is compressed to 256 channels through a 1x1 convolution layer with a step of 1; and a bidirectional attention transformer is used to realize deep bidirectional interaction between the prompt features and the fusion feature tensor; and 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.

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

[0087] As shown in Figure 4 The semantic features (generated by RepVit encoder, size [H / 4, W / 4, 256]) output by the aligned image encoding branch and the spectral features (multi-water index fusion and standardization, size [H / 4, W / 4, 256]) output by the multispectral inversion branch are channel cascaded to form a fusion feature tensor of [H / 4, W / 4, 512], and then compressed to 256 channels through a 1x1 convolution layer with a step of 1, which reduces the computational complexity of subsequent calculations while retaining key information, providing efficient input for bidirectional attention interaction.

[0088] An improved bidirectional attention transformer structure is used to realize bidirectional guidance of semantic features and spectral features.

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

[0090] Spectral constraint path: using cross-attention mechanism, the water spectral prior in multi-spectral features is used as a weight to enhance the activation strength of the water area in the semantic features and improve the feature discrimination of mixed pixels.

[0091] S3: 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. It should be noted that:

[0092] The specific operation of converting the water area segmentation mask into a high-precision remote sensing water area segmentation vector result through morphological filtering and vector conversion includes,

[0093] The optimal water area mask is subjected to a morphological filtering operation, 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 using a vector conversion algorithm, each water area object is assigned geographic coordinates and attribute information, and high-precision remote sensing water area segmentation vector results are formed.

[0094] The embodiments take Sentinel multispectral remote sensing images as data sources, and verify the accuracy and efficiency of the method for segmenting remote sensing water areas through specific experiments. The specific implementation steps and result analysis are as follows:

[0095] Model training:

[0096] In this embodiment, 1106 scenes of Sentinel multispectral remote sensing images are used as training and test data, with a total data capacity of 893 GB and a coverage area of about 9.98 million square kilometers. The data collection time span is from October 2022 to March 2025, among which the number of images obtained from August to October is relatively small due to seasonal changes and cloud coverage (for details, see Figure 5 ). All data meet the following technical indicators: cloud coverage rate less than 3%, spatial resolution up to 10 meters, containing B2 (blue), B3 (green), B4 (red), B8 (near-infrared) and other multispectral bands, which can truly reflect the spectral characteristics of water areas in different seasons and geographical environments.

[0097] Figure 6 The model training loss function curve is shown in the figure, a total of 200 training rounds, the loss function value gradually converges to 0 with the increase of training rounds, indicating that the model has learned the data characteristics defined by the training data set.

[0098] Segmentation effect comparison:

[0099] In this embodiment, the segmentation effect of the method provided in the present application and other semi-automatic segmentation algorithms in remote sensing water area segmentation is compared, and the following methods are used to segment the water area of the test remote sensing image:

[0100] Baseline method: semi-automatic segmentation based on SAM model (requires manual input of a small number of prompt points to assist segmentation);

[0101] Semi-automatic segmentation method of the present application: the model provided in the present application fuses multi-band inversion and generalization segmentation (inputs the same prompt points as SAM);

[0102] Figures 7 to 10The trunk river scene water area original drawing, the real value picture, the SAM semi-autonomous segmentation picture involved in the benchmark method and the semi-autonomous segmentation picture proposed in the application are respectively shown. As can be seen from the picture, the result obtained by SAM semi-autonomous segmentation has more missing on the river edge details, and the gap with the real value in details is large; the semi-autonomous segmentation result proposed in the application performs well on the river edge details, has no big missing, only a small amount of rough edge problem, and is more consistent with the real data.

[0103] Figures 11 to 14 The reservoir scene water area original drawing, the real value picture, the SAM semi-autonomous segmentation picture involved in the benchmark method and the semi-autonomous segmentation picture proposed in the application are respectively shown. As shown in the figure, SAM semi-autonomous segmentation completely misses the identification of the reservoir inlet river and outlet river, and the gap with the real situation is too large; the semi-autonomous segmentation result proposed in the application is more consistent with the real value.

[0104] Figures 15 to 18 The fine river scene water area original drawing, the real value picture, the SAM semi-autonomous segmentation picture involved in the benchmark method and the semi-autonomous segmentation picture proposed in the application are respectively shown. As shown in the figure, the river obtained by SAM semi-autonomous segmentation is too smooth, although it covers the area described by the real situation of the river, but the excessive area is too much and does not conform to the real situation; and the semi-autonomous segmentation result proposed in the application is closer to the real value, only a small amount of structure is excessive on the edge details.

[0105] In the trunk river, reservoir, fine river three scenes, the SAM semi-autonomous segmentation effect and the semi-autonomous segmentation effect described in the application are compared. The method provided in the application is more fine in the processing of water boundary, can capture more detailed water details, and is more consistent with the real situation of the image in form.

[0106] Table 1: Comparison of scene segmentation indicators

[0107]

[0108] In addition, in the trunk river, reservoir, fine river three scenes, the operation amount of manual annotation and semi-autonomous segmentation is compared:

[0109] Manual annotation: professional personnel use traditional GIS software to perform water boundary vectorization, and record the number of coordinate points required for manual input in the annotation process;

[0110] Semi-autonomous segmentation method of the application: on the same image, use the semi-autonomous segmentation function of the model described in the application, and perform the model prediction and raster vectorization post-processing operation described in the application, record the number of coordinate points required 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 labeled points and the number of semi-autonomous segmentation prompt points

[0112]

[0113] From Table 2, it can be seen that the average single scene image manually labeled needs to manually input about 100 coordinate points, and the semi-autonomous method only needs 2-3 prompt points to complete high-quality segmentation, and the coordinate point requirement is reduced by more than 97%.

[0114] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any change or replacement within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generalized segmentation of semi-autonomous remote sensing waters integrated with multispectral inversion, characterized by: include: Collect remote sensing multispectral data; Remote sensing multispectral data and coordinate point prompt information are used as inputs of the end-to-end model. Spectral features containing the spectral physical characteristics of water bodies are obtained through multispectral inversion branch processing. Semantic features containing pixel semantic information and prompt features containing coordinate point prompt information are obtained through generalized segmentation branch processing. Finally, spectral features, semantic features, and prompt features are combined through cross-modal feature fusion processing to generate the optimal water segmentation mask that conforms to the spectral physical characteristics of water bodies and pixel semantic information. The optimal water segmentation mask is converted into a high-precision remote sensing water segmentation vector result through morphological filtering and vector conversion.

2. The method for generalized segmentation of semi-autonomous remote sensing waters integrated with multispectral inversion according to claim 1, characterized in that: The remote sensing multispectral data includes 3 visible light bands and 8 non-visible light bands, totaling 11 bands of multispectral remote sensing images.

3. The method for generalized segmentation of semi-autonomous remote sensing water areas integrated with multispectral inversion according to claim 2, characterized in that: The multispectral inversion branch processing includes parallel water index calculation and feature splicing and preprocessing; The parallel water index calculation specifically includes: Adopting an independent channel parallel computing framework, n water indices are calculated synchronously, 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 GPU thread-level parallel characteristics to achieve efficient calculation; Feature splicing and preprocessing specifically include: The n single-channel feature maps are concatenated into a H×W×n multispectral prior feature tensor according to the channel dimension, and the scale is dynamically adjusted. The eigenvalue range is adjusted by a learnable scaling factor and offset. The convolution-normalization-activation stacking model is then used to extract spectral features through multi-layer convolution, where H represents the image height of the multispectral image and W represents the image width of the multispectral image.

4. The method for generalized segmentation of semi-autonomous remote sensing water areas integrated with multispectral inversion according to claim 3, characterized in that: The generalized segmentation branch processing includes: The generalized segmentation branch consists of an image encoder and a coordinate point hint encoder, which is used to extract the semantic features and hint features of the image, and semi-autonomously correct the output of the image encoder based on the coordinate point hint information to optimize the semantic features.

5. The method for generalized segmentation of semi-autonomous remote sensing water areas integrated with multispectral inversion according to claim 4, characterized in that: The visible light spectrum image is processed by the image encoder to obtain semantic features.

6. The method for generalized segmentation of semi-autonomous remote sensing waters integrated with multispectral inversion according to claim 5, characterized in that: The coordinate point prompt encoder includes two steps: position encoding and weight superposition; Input the pixel 2D coordinates [x, y] of the point provided by the user on the multispectral image; Position encoding uses a random Gaussian matrix and trigonometric functions to encode coordinates. The two-dimensional coordinates [x, y] are scaled to the range [0, 1]. The two-dimensional coordinates are mapped to 128 dimensions using a Gaussian matrix [N, 128]. The mapping results are processed using the sin and cos functions, respectively, and then concatenated into a 256-dimensional vector, where N represents the number of points provided by the user, x represents the column index of the user-provided point on the multispectral image, and y represents the row index of the user-provided point on the multispectral image. The 256-dimensional vector is then processed using a weight table to obtain the prompt feature.

7. The method for generalized segmentation of semi-autonomous remote sensing waters integrated with multispectral inversion according to claim 6, characterized in that: The operations of cross-modal feature fusion processing include: The aligned semantic features and spectral features are channel-concatenated to form a fused feature tensor, which is compressed to 256 channels through a 1×1 convolution layer with a stride of 1. A bidirectional attention converter is used to achieve deep bidirectional interaction between the hint features and the fused feature tensor. The confidence-optimized mask generator is then used to first generate four candidate water masks, calculate the confidence of the four candidate masks, predict the four candidate masks through three multi-perception layers, output four confidence scores, and select the optimal water mask through a weighted voting mechanism.

8. The method for generalized segmentation of semi-autonomous remote sensing water areas integrated with multispectral inversion according to claim 7, characterized in that: The specific operations of converting the optimal water segmentation mask into a high-precision remote sensing water segmentation vector result through morphological filtering and vector conversion include: Morphological filtering operations are performed on the optimal water area mask, and small noise points and holes in the mask are eliminated through dilation and erosion processing to enhance the smoothness and integrity of the water body boundary; then the mask is converted into vector polygon data using a vector conversion algorithm, and each water area object is given geographic coordinates and attribute information to form a high-precision remote sensing water area segmentation vector result.

Images