CN-120125928-B - Lake water surface extraction method for multi-source data fusion and deep feature screening

CN120125928BCN 120125928 BCN120125928 BCN 120125928BCN-120125928-B

Abstract

The invention relates to the field of lake water surface extraction, in particular to a lake water surface extraction method for multi-source data fusion and deep feature screening. The method comprises the steps of preprocessing multi-source remote sensing data, eliminating differences among data sources, designing a 1D-CNN and HRnet-AM feature extraction module, extracting spectral features, texture features and shape features from images of different data sources, constructing a self-adaptive weight quantization module A-WCM, carrying out fusion analysis and quantization on the multi-element features, designing a redundancy rejection module REM, rejecting redundant or irrelevant features, carrying out layer-by-layer fusion and recombination on the features extracted by all branches by adopting a deep semantic fusion technology, constructing a decision tree optimal growth model OGMT, and effectively reducing the complexity of the model through an optimal path selection and feature screening mechanism. The invention improves the recognition efficiency and accuracy of the lake water surface, can more reliably capture and analyze the lake water surface change, and provides powerful support for the management and maintenance of the lake ecological system.

Inventors

ZHANG YISHU
JIANG WEIXIA
TANG YUANYUAN
CHEN FANGFANG
WANG KAIHUA
HUANG LEI
HUANG WANYI
Long Sijia

Assignees

中国地质调查局长沙自然资源综合调查中心

Dates

Publication Date: 20260508
Application Date: 20250107

Claims (10)

1. A lake water surface extraction method for multi-source data fusion and deep feature screening is characterized by comprising the following steps: S1, acquiring multisource remote sensing image data, including Landsat image data, sentinel image data, high score and resource satellite image data, preprocessing the image data, and constructing a training set; S2, constructing an MD-DFE neural network model, wherein the model comprises a differential parallel 1D-CNN and HRnet-AM feature extraction module, a self-adaptive weight quantization module, a redundant feature eliminating module and a multi-source feature deep fusion module; The multi-source feature extraction method comprises the steps of extracting spectral features from preprocessed Landsat image data through 1D-CNN, extracting texture features from preprocessed Sentinel image data through 1D-CNN, generating image blocks from preprocessed high-resolution and resource satellite image data after multi-scale segmentation, extracting semantic features of the image blocks through HRnet-AM, and carrying out multi-source feature fusion through Concat algorithm, wherein an adaptive weight quantization module is used for quantizing weights of the multi-source features; S3, training the MD-DFE neural network model by using a training set, extracting the fused features, inputting the fused features into the optimal growth model OGMT of the decision tree, acquiring nodes and threshold values by an optimal path selection and feature screening mechanism, constructing a decision tree classification model, and realizing automatic extraction of the lake water surface by using the decision tree classification model.
2. The method for extracting water surface of lake by multi-source data fusion and deep feature screening of claim 1, wherein, And downloading multi-source remote sensing image data according to the area where the lake is located, the characteristics of the weather and the field investigation time, wherein the Sentinel image data are VV+VH dual-polarized data so as to enhance the accuracy of classification of the ground features.
3. The method for extracting water from a lake of claim 1 wherein S1, the preprocessing is, Aiming at optical image data, firstly converting DN values of the images into earth surface reflectivity for eliminating the influence of atmospheric scattering and absorption on the imaging of the images in the imaging process, then resampling the images to uniform spatial resolution to ensure the consistency of multi-source data on spatial scale, thereby facilitating subsequent fusion and analysis, and finally realizing geometric correction through GCPs, wherein the optical image data are Landsat-8 multispectral remote sensing images and high-resolution and resource satellite data images; For the Sentinel image data, firstly, radiation correction is carried out, then resampling is carried out to the resolution consistent with the optical image, the latest track file is downloaded and applied, the track correction is automatically carried out on the Sentinel image, and finally, the removal of thermal infrared noise and the filtering of speckle noise are carried out on the Sentinel image data, so that the image quality is improved, and clearer texture characteristics are obtained.
4. The method for extracting the lake water surface by utilizing the multi-source data fusion and deep feature screening as claimed in claim 1, wherein in S2, a plurality of spectral bands and spectral indexes after being preprocessed by Landsat-8 multi-spectral remote sensing images are respectively obtained aiming at spectral features, 2 backscattering coefficients VH and VV of two polarization modes of a Sentinel image are firstly obtained aiming at texture features, a 5 multiplied by 5 filter window is selected, and feature indexes under all angles are calculated by utilizing a gray level co-occurrence matrix method.
5. The lake water surface extraction method for multi-source data fusion and deep feature screening according to claim 1 is characterized in that 9 wave bands and 8 spectral features are extracted from Landsat-8 multi-spectral images, 17 variables are taken in total, 2 backscattering coefficients and 9 texture features are taken from Sentinel images, 11 variables are taken in total, high-resolution images and resource satellite images are subjected to multi-scale segmentation to obtain image objects, the average value of each variable taken by the multi-spectral images and the Sentinel images in each object is calculated according to each segmented object obtained after segmentation to obtain two column vectors, boundary and morphological features of ground objects are extracted through a multi-scale segmentation technology by utilizing Ecognition software, and then the minimum outsourcing rectangle of each segmented object is created for the multi-scale segmentation result of the high-resolution images and the resource satellite images, so that high-resolution image blocks are generated for semantic feature extraction.
6. The lake water surface extraction method for multi-source data fusion and deep feature screening of claim 1, wherein in S2, an MD-DFE neural network model is a hierarchical parallel convolutional neural network and comprises 1D-CNN and HRnet-AM feature extraction modules, wherein the 2 modules are used for deep semantic feature extraction of a multi-source data set, the extraction process comprises (1) taking pixels as object units, respectively processing Landsat-8 and Sentinel-1 images by using 1D-CNN to extract two feature column vectors, the network is totally designed with 4 layers, and comprises 2 convolutional layers, 1 flattening layer and 1 full-connection layer, all feature images of each initial input vector are combined into 1 one-dimensional column vector through the flattening layer, finally, after some neurons are randomly discarded through a random inactivation regularization module, the features are re-extracted, and semantic features obtained by the feature column vectors are output, (2) processing high-resolution image blocks to obtain multi-scale information and need to expand a sense image field.
7. The lake water surface extraction method of multi-source data fusion and deep feature screening of claim 1, wherein in S2, the calculation process of the adaptive weight quantization module is to multiply the weight value with the semantic feature column vector extracted by hierarchical parallel convolution, and fuse the weighted feature column vector by Concat method, so as to increase the feature dimension describing the image block, but the information quantity under each dimension feature is unchanged.
8. The lake water surface extraction method of multi-source data fusion and deep feature screening of claim 6, wherein in S3, the training process is that firstly, the full-connection layer output classification type information is added for the hierarchical parallel convolution neural network, then the MD-DFE neural network model is trained by utilizing the constructed training set, the model precision after each iteration training is checked by utilizing the verification set, the iteration times are set, multiple iteration training is carried out, the model parameters are updated by utilizing the gradient descent method in each iteration, the loss value of the model is reduced until the iteration times are reached, and the trained MD-DFE neural network model is obtained.
9. The lake water surface extraction method of multi-source data fusion and deep feature screening of claim 1, wherein in S4, the process of constructing OGMT is to form a binary tree-form choice tree structure by cyclic analysis of training data sets formed by test variables and target variables, and to use the coefficient of the kenel as the standard for selecting the most attribute features and dividing threshold values, namely, the feature with the minimum coefficient of the kenel as the standard for node splitting, and stop splitting when the samples in each sub-node belong to the same category or reach the preset threshold value, and finally obtain OGMT.
10. A storage medium having stored thereon a computer program, which when executed by a processor performs the steps of the method for multi-source data fusion and deep feature screening of lake water surface extraction of any one of claims 2-9.

Description

Lake water surface extraction method for multi-source data fusion and deep feature screening Technical Field The invention relates to the fields of remote sensing image processing, lake water surface extraction and deep learning, in particular to a lake water surface extraction method for multi-source data fusion and deep feature screening. Background The extraction of lake water surface by remote sensing technology is a crucial working means at present. However, this process usually relies on a single image source, such as Landsat-8 satellite images, resource satellite images, high-resolution satellite images, synthetic aperture radar images, and the like, to distinguish and extract based on spectral features, texture characteristics, and shape contours exhibited by lakes in the remote sensing images. Although the remote sensing technology has wide application prospect in the aspect of lake water surface extraction, various image sources have inherent advantages and limitations. For example, landsat-8 satellite images, although rich in spectral information, have relatively low spatial resolution, limiting to some extent the precise identification in terms of ground object texture. The resource satellite and the high-resolution satellite have higher spatial resolution, can clearly capture the earth surface details, but have fewer wave bands, relatively lack spectrum information, are easily limited by satellite orbits and revisit periods, and can influence the continuity of data. In addition, the lake region represented by the eastern plain in China is rich in precipitation, and as the cloud and fog shielding phenomenon is obvious, the optical remote sensing satellites such as Landsat-8, resources and high scores are difficult to meet the requirements of real-time observation and low cloud cover. The synthetic aperture radar image (SYNTHETIC APERTURE RADAR, SAR) has good cloud penetration and fog penetration capability, can provide good texture information for ground objects, but also has the problems of dispersion and deletion in the aspects of observation angle and resolution due to the limitation of a track and a space domain, and on the other hand, the presence of speckle noise can influence the definition of the SAR image. In order to solve the problems of low lake identification accuracy, unmatched boundaries and the like caused by a single data source, expert scholars at home and abroad propose a multi-source data fusion method. However, the current research still stays in extracting the data characteristics after data fusion by using the traditional manual method, and the advantages of multi-source data fusion are not fully utilized, so that the extraction precision does not meet the expected requirements yet. Disclosure of Invention In order to solve the problem that the automatic extraction of the lake water surface is difficult in the prior art, the invention provides a lake water surface extraction method for multi-source data fusion and deep feature screening, which mainly comprises the following steps: S1, acquiring multi-source remote sensing image data, preprocessing the multi-source remote sensing image data, extracting the characteristics of the data, cutting the preprocessed remote sensing image data to acquire image blocks, and constructing a training set; S2, constructing an MD-DFE neural network model, wherein the model comprises a 1D-CNN and HRnet-AM feature extraction module, a self-adaptive weight quantization module A-WCM and a redundant feature rejection module REM, wherein the 1D-CNN and HRnet-AM feature extraction module is used for extracting semantic features of different image blocks; s3, realizing multi-source semantic feature fusion by adopting a deep semantic fusion technology; and S4, training the MD-DFE neural network model by utilizing a training set, extracting the fused semantic features, inputting the fused semantic features into the optimal growth model OGMT of the decision tree, acquiring nodes and threshold values by an optimal path selection and feature screening mechanism, constructing a decision tree classification model, and realizing automatic extraction of the lake water surface by utilizing the decision tree classification model. A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the steps of the method for multi-source data fusion and deep feature screening. The technical scheme provided by the invention has the beneficial effects that in the aspect of lake water surface extraction, the prior art generally faces the problems of low identification accuracy, mismatching of boundaries, insufficient robustness and the like caused by a single data source, and the accuracy and the intelligent level of lake water surface extraction are seriously influenced. In order to overcome the technical bottlenecks, the invention provides a solution for multi-source data fusion and deep feature scree