CN-121999383-A - Multi-source multi-temporal remote sensing image-based composite water network segmentation method and system

CN121999383ACN 121999383 ACN121999383 ACN 121999383ACN-121999383-A

Abstract

The invention discloses a composite water network segmentation method and a system based on a multi-source multi-temporal remote sensing image, which relate to the technical field of water network identification and comprise the steps of acquiring a remote sensing image of a target area and preprocessing to obtain a reference data set; the method comprises the steps of obtaining multisource multi-temporal remote sensing data of a target area, preprocessing the multisource multi-temporal remote sensing data to obtain a remote sensing image stack, extracting face scale input data and point scale input data which are used as input data based on the remote sensing image stack, inputting the input data to a composite water network segmentation model, obtaining a water network prediction result through sharing a CNN encoder, a time attention module, a Swin transform encoder and a double-branch decoder, reversely transmitting an optimized composite water network segmentation model based on the water network prediction result and a reference data set to obtain an optimized composite water network segmentation model, obtaining a remote sensing image of the area to be predicted, inputting the remote sensing image to the optimized composite water network segmentation model, and obtaining the water network segmentation result. The water network in different areas can be accurately identified.

Inventors

LI MINGYANG
WANG RUI
LIU LILI
GONG XIANGFENG
WANG GUANGHUI
HAO XIAOHUI
YANG DAWEI
GUO XIANHU
HUANG JIWEN
SHI YUZHI
LI FULIN
XU SHANGJIE
CHENG SUZHEN
GUO LEI
XIN HONGJIE
YANG ZHEN

Assignees

山东省水利科学研究院

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (10)

1. A composite water network segmentation method based on a multi-source multi-temporal remote sensing image is characterized by comprising the following steps: acquiring a remote sensing image of a target area and preprocessing the remote sensing image to obtain a reference data set; Acquiring multi-source multi-temporal remote sensing data of the target area and preprocessing the data to obtain a remote sensing image stack; extracting face scale input data and point scale input data serving as input data based on the remote sensing image stack; inputting the input data to a composite water network segmentation model, and processing by a shared CNN encoder, a time attention module, a Swin transform encoder and a double-branch decoder to obtain a water network prediction result; The composite water network segmentation model is optimized based on the water network prediction result and the reference data set in a counter-propagation mode, and an optimized composite water network segmentation model is obtained; and acquiring a remote sensing image of the area to be predicted, and inputting the remote sensing image into the optimized composite water network segmentation model to obtain a water network segmentation result.
2. The method for segmenting the composite water network based on the multi-source multi-temporal remote sensing image according to claim 1, wherein the method for acquiring the reference data set is as follows: Classifying the target area based on the water network type to obtain a classified investigation area; carrying out water network distribution identification on the remote sensing images of the classified investigation regions correspondingly, and adjusting region boundaries in combination with field investigation to obtain field investigation data; Integrating the on-site survey data with the existing data set to obtain a preliminary integrated vector data set; Performing cross-validation and spatial adjustment based on the preliminary integrated vector dataset and the existing dataset to obtain a standardized vector reference dataset; And carrying out a rasterization algorithm and binary coding processing on the basis of the standardized vector reference data set to obtain the reference data set.
3. The method for segmenting the composite water network based on the multi-source multi-temporal remote sensing image according to claim 1, wherein the method for acquiring the remote sensing image stack is as follows: Acquiring multispectral images and synthetic aperture radar data of the target area, and correspondingly acquiring multispectral image sets and SAR image sets through space-time filtering screening; Performing geometric correction, polarization extraction and noise suppression on the basis of the SAR image set to obtain an SAR quarter average image stack; removing the atmospheric interference based on the multispectral image set and reserving to obtain multispectral wave bands; quarterly calculating to obtain quarterly derivative spectrum indexes based on the multispectral wave bands; forming a multispectral quarter image stack based on the multispectral wave bands and quarter derivative spectral indexes; and carrying out fusion registration based on the SAR quarter average image stack and the multispectral quarter image stack to obtain the remote sensing image stack.
4. The method for segmenting the composite water network based on the multi-source multi-temporal remote sensing image according to claim 3, wherein the method for acquiring the face scale input data and the point scale input data is as follows: extracting the face scale input data based on the remote sensing image stacks in the coverage range of all the classified investigation regions in the target region; Selecting a plurality of interest points which are determined to be water networks or non-water networks in each classification investigation region, and extracting the point data of a remote sensing image stack from the remote sensing image stack; Constructing matrix data according to the format of each interest point of each behavior and each column as each remote sensing image band based on the point data; and taking the interest points representing the water network in the last column of the matrix data as the point scale input data.
5. The method for segmenting the composite water network based on the multi-source multi-temporal remote sensing image according to claim 4, wherein the method for obtaining the water network prediction result is as follows: inputting the point scale input data and the surface scale input data to the shared CNN encoder in parallel for feature extraction to obtain comprehensive extraction features; inputting the comprehensive extraction features to the time attention module, applying a dynamic attention mechanism to different time steps, and distributing weights for seasonal changes to obtain time attention enhancement features; Inputting the time attention enhancement features to the Swin transducer encoder to obtain advanced space-time features; and inputting the advanced space-time characteristics to the double-branch decoder, and correspondingly obtaining a point scale prediction result and a face scale prediction result to be used as the water network prediction result together.
6. The method for segmenting a composite water network based on a multi-source multi-temporal remote sensing image according to claim 5, wherein the shared CNN encoder comprises two feature extraction branches with shared weights; the two feature extraction branch structures are the same, and each feature extraction branch structure comprises a first convolution layer, a first activation layer, a second convolution layer, a second activation layer, a third convolution layer, a third activation layer and a pooling layer; The point scale input data and the surface scale input data are respectively and sequentially input into the first convolution layer, the first activation layer, the second convolution layer, the second activation layer, the third convolution layer, the third activation layer and the pooling layer, so as to correspondingly obtain point extraction features and surface extraction features; The comprehensive extraction feature is composed based on the point extraction feature and the face extraction feature together.
7. The method for segmenting the composite water network based on the multi-source and multi-temporal remote sensing image according to claim 6, wherein the time attention module comprises a global average pooling layer, a multi-layer perceptron and a Softmax function; based on the point extraction features and the surface extraction features, respectively and sequentially inputting the point extraction features and the surface extraction features into the global average pooling layer for space compression, and carrying out time modeling and the Softmax function by the multi-layer perceptron to correspondingly obtain a first dynamic weight and a second dynamic weight; based on the first dynamic weight and the point extraction feature fusion, a first enhancement feature is obtained; fusing the second dynamic weight with the face extraction feature to obtain a second enhancement feature; The first enhancement feature and the second enhancement feature together comprise the temporal attention enhancement feature.
8. The method for segmenting the composite water network based on the multi-source and multi-temporal remote sensing image according to claim 7, wherein the Swin transducer encoder comprises a patch embedding layer, a first Swin transducer block, a second Swin transducer block, a patch merging layer and a third Swin transducer block; The first enhancement feature and the second enhancement feature are respectively input into the patch embedding layer, the features are divided into non-overlapping patches and subjected to linear projection, and a first block feature and a second block feature are obtained; Based on the first block feature and the second block feature, the first block feature and the second block feature are sequentially input into the first Swin transform block and the second Swin transform block respectively, the local-global dependence is captured and then input into the patch merging layer to be connected with adjacent patches, and the second block feature and the third block feature are input into the third Swin transform block, so that a first space-time feature and a second space-time feature are correspondingly obtained; The advanced spatiotemporal features are composed together based on the first spatiotemporal feature and the second spatiotemporal feature.
9. The method for segmenting the composite water network based on the multi-source and multi-temporal remote sensing image according to claim 8, wherein the dual-branch decoder comprises an up-sampling unit, a point scale segmentation head and a face scale segmentation head; The first space-time feature and the second space-time feature are processed by the up-sampling unit and then are correspondingly input into the point scale segmentation head and the surface scale segmentation head to obtain the point scale prediction result and the surface scale prediction result; the method for obtaining the optimized composite water network segmentation model comprises the following steps: Constructing a point segmentation loss and a surface segmentation loss based on the reference data set and the point scale prediction result and the surface scale prediction result respectively; Constructing a consistency loss based on the point scale prediction result and the face scale prediction result; Constructing and obtaining total loss based on the point segmentation loss, the surface segmentation loss and the consistency loss; And optimizing the composite water network segmentation model based on the total loss back propagation to obtain the optimized composite water network segmentation model.
10. A multi-source multi-temporal remote sensing image-based composite water network segmentation system for executing a multi-source multi-temporal remote sensing image-based composite water network segmentation method according to any one of claims 1 to 9, comprising a reference data acquisition module, a remote sensing data processing module, an input data acquisition module, an initial prediction output module, a model optimization module and a final result output module; the reference data acquisition module is used for acquiring a remote sensing image of the target area and preprocessing the remote sensing image to obtain a reference data set; the remote sensing data processing module is used for acquiring and preprocessing the multi-source multi-temporal remote sensing data of the target area to obtain a remote sensing image stack; the input data acquisition module is used for extracting face scale input data and point scale input data based on the remote sensing image stack to serve as input data; The initial prediction output module is used for inputting the input data into a composite water network segmentation model, and obtaining a water network prediction result through processing of a shared CNN encoder, a time attention module, a Swin transform encoder and a double-branch decoder; The model optimization module is used for optimizing the composite water network segmentation model based on the water network prediction result and the reference data set in a counter-propagation mode to obtain an optimized composite water network segmentation model; And the final result output module is used for acquiring a remote sensing image of the area to be predicted and inputting the remote sensing image into the optimized composite water network segmentation model to obtain a water network segmentation result.

Description

Multi-source multi-temporal remote sensing image-based composite water network segmentation method and system Technical Field The invention relates to the technical field of water network identification, in particular to a composite water network segmentation method and system based on multi-source multi-temporal remote sensing images. Background The accurate extraction of the water network from the remote sensing image has important significance for flood and drought monitoring, city planning, agricultural management and other applications. However, the conventional method faces a great challenge due to various water forms, such as irregular natural river boundaries, narrow and winding irrigation channels, complex urban waterway structures, and the like, and the influence of factors such as illumination, shadow, ground object mixing, and the like. Early common approaches were based primarily on water indices such as Normalized Differential Water Index (NDWI) and Modified NDWI (MNDWI). The method is simple in calculation and high in efficiency, but misjudgment is easy to generate in complex scenes with interferences such as shadows, buildings or vegetation. In recent years, the depth learning technology has significantly improved the accuracy of water body segmentation. Particularly, models based on Full Convolutional Networks (FCNs) and encoder-decoder structures are widely used for water network extraction of high-resolution remote sensing images, wherein U-Net has become a mainstream architecture because of its capability of effectively fusing multi-scale features and spatial details by skipping connection mechanisms. In order to further improve the performance, researchers introduce residual connection, attention mechanisms, expansion convolution and ASPP modules to enhance the recognition capability of small water bodies and complex boundaries, and meanwhile, the use of a lightweight backbone network reduces the calculation cost and supports real-time or large-scale application. In order to solve the problem that optical remote sensing is easily interfered by cloud layers and the atmosphere, multi-source data fusion becomes an important direction, for example, all-weather water body monitoring can be realized by combining Synthetic Aperture Radar (SAR) data with optical images, and a Digital Elevation Model (DEM) or laser radar data is beneficial to improving the segmentation precision of complex areas (such as river banks and urban water channels) by utilizing terrain information. In addition, the introduction of multi-phase data is helpful for distinguishing temporary water bodies from permanent water bodies, and the robustness of the model in a dynamic environment is improved. Despite the remarkable progress of the existing method, water network segmentation still faces a plurality of challenges, namely that firstly, water bodies are highly similar to shadows, dark roofs or dense vegetation in spectrum and are easy to cause misclassification, secondly, the spatial heterogeneity of water body types is strong, the difference of the natural water bodies and artificial water channels in form, width and continuity is large, and particularly, accurate characterization is difficult under low resolution. These problems are particularly prominent in complex environments such as cities, and remain the key difficulties in current research. Therefore, how to accurately identify water networks in different areas is a problem that needs to be solved by those skilled in the art. Disclosure of Invention In view of the above problems, the present invention is provided to provide a method and a system for segmenting a composite water network based on multi-source multi-temporal remote sensing image, which overcome or at least partially solve the above problems, and can accurately identify water networks in different areas. In order to achieve the above purpose, the present invention adopts the following technical scheme: in a first aspect, an embodiment of the present invention provides a method for segmenting a composite water network based on a multi-source multi-temporal remote sensing image, including: acquiring a remote sensing image of a target area and preprocessing the remote sensing image to obtain a reference data set; Acquiring multi-source multi-temporal remote sensing data of the target area and preprocessing the data to obtain a remote sensing image stack; extracting face scale input data and point scale input data serving as input data based on the remote sensing image stack; inputting the input data to a composite water network segmentation model, and processing by a shared CNN encoder, a time attention module, a Swin transform encoder and a double-branch decoder to obtain a water network prediction result; The composite water network segmentation model is optimized based on the water network prediction result and the reference data set in a counter-propagation mode, and an optimized composite water n