CN113435484A - U-Net neural network ice lake extraction method combined with self-attention mechanism - Google Patents

U-Net neural network ice lake extraction method combined with self-attention mechanism Download PDF

Info

Publication number
CN113435484A
CN113435484A CN202110659969.XA CN202110659969A CN113435484A CN 113435484 A CN113435484 A CN 113435484A CN 202110659969 A CN202110659969 A CN 202110659969A CN 113435484 A CN113435484 A CN 113435484A
Authority
CN
China
Prior art keywords
ice lake
research area
neural network
net
self
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110659969.XA
Other languages
Chinese (zh)
Inventor
何毅
张立峰
姚圣
陈宝山
杨旺
陈毅
蒲虹宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Lanzhou Jiaotong University
Original Assignee
Lanzhou Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Lanzhou Jiaotong University filed Critical Lanzhou Jiaotong University
Priority to CN202110659969.XA priority Critical patent/CN113435484A/en
Publication of CN113435484A publication Critical patent/CN113435484A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a U-Net neural network ice lake extraction method combined with a self-attention mechanism, which mainly comprises the following steps: firstly, acquiring a remote sensing image of a research area and drawing label data; then, a supervised classification training grid data set is manufactured, a contrast test input waveband is selected, a neural network model is constructed, and effects are contrastively analyzed; and finally, extracting the ice lake in the research area by using the model with the best effect. The invention combines Landsat 8 full-waveband remote sensing data with the U-Net network fused with self attention, can be used for assisting decision, early warning and reducing harm, and provides a new idea for disaster prevention and reduction.

Description

U-Net neural network ice lake extraction method combined with self-attention mechanism
Technical Field
The invention relates to the field of information extraction of deep learning methods, in particular to a U-Net neural network ice lake extraction method combined with a self-attention mechanism.
Background
The glacier lake is a natural water body formed by glacier molten water staying in a sunken terrain on the ground surface, is an incubator for the alpine glacier disaster, and has an important role in the mountain disaster research. In recent years, a plurality of ice lake collapse events occur, and great life and property losses are caused. Therefore, real-time monitoring of the ice lake is of great importance. The ice lake is a special lake and has various characteristics in a high-altitude area. The geographic information contained in the optical remote sensing image is complex, and a sensor of the optical remote sensing satellite has multiband data, is influenced by different terrains and also carries complex semantic information. In the remote sensing image, the ice lake may have a phenomenon similar to some ground features (such as mountain shadow), so that it is difficult to accurately extract the ice lake from the remote sensing image in a large range.
In recent years, deep learning is taken as a novel multi-layer neural network learning algorithm, high-level features can be obtained from original features without manual intervention, and the classification accuracy is improved. The U-Net network is a U-shaped network structure and can acquire context information and position information simultaneously. The design is originally designed to solve the problem of medical image segmentation. U-Net is a more refined design based on the basic structure of a full convolution neural network (FCN), replacing the optimized FCN network. The U-net employs a network structure that includes down-sampling and up-sampling. The down-sampling is used to gradually present the environment information, and the up-sampling is a process of restoring detail information by combining the down-sampled layer information and the up-sampled input information, and gradually restoring the image precision.
Self-attention U-net is a deep learning method combined with an attention mechanism. The self-attention U-Net network encoder and decoder parts use the same scale and are connected with the AG module to realize an attention mechanism. Firstly, the AG module combines the feature maps of the encoder part in the same proportion with the up-sampling result of the decoder part to carry out 3 multiplied by 3 convolution operation, and the number of output channels is half of that of the proportion encoder and decoder parts, thereby extracting the fusion feature on the coarse scale and eliminating irrelevant noise. The ReLU activation function has an activation function that suppresses AG module overfitting, and then outputs attention weights using 1 output channel and convolution layer of 1 × 1 convolution kernel size.
The invention provides a U-Net neural network ice lake extraction method combined with a self-attention mechanism.
In the prior art, the following documents are mainly relevant to the present application.
Document 1, an invention of a remote sensing satellite image-based method for extracting ice lake in plateau region, applied by the remote sensing of the Chinese academy of sciences and the Chen of the digital Earth research institute, and the like, application number: CN 201710957162.8. The method is mainly used for large-scale crude extraction of the ice lake and partial fine extraction of the ice lake, and the crude extraction method comprises the following steps: obtaining a normalized ice lake index NDWI value of the whole area according to the preprocessed remote sensing satellite image, and setting thresholds of NDWI, NIR and SWIR to remove the influence of glaciers and other parts of ground objects on the whole image by combining a near infrared band NIR value and a short wave infrared band SWIR value; the method comprises the following steps of generating a slope map of an area by utilizing preprocessed DEM data, setting a slope threshold to remove the influence of mountain shadow, and roughly extracting all ice lakes and mixed partial backgrounds, wherein the fine extraction method comprises the following steps: accurately extracting the ice lakes one by using an C V model introducing a symbolic pressure function on the result of the coarse extraction of the ice lakes; the extraction method can effectively remove various factors influencing the extraction of the ice lake, achieves accurate extraction, and has better extraction effect aiming at the smaller ice lake in the image.
Document 2, an invention of image segmentation method and system for green apples based on U-Net network, applied by jordan university, jiawei width, etc., application no: CN 202011427191.1. The method obtains the green apple segmentation image through the image segmentation model,
inputting the green apple image into the trained image segmentation model to obtain a green apple segmentation image; the image cutting model is a green apple Edge feature extraction model which is built on the basis of a U-Net network and is fused into an Edge structure, then green apple Edge features and the green apple features extracted on the basis of the U-Net network are fused, green apple segmentation images are output, and rapid and accurate segmentation of green apples is achieved.
Document 3, songtianqiang, et al, applied to Qingdao science and technology university, the invention, a method for extracting a remote sensing image road with an improved U-Net network, has the following application numbers: CN 202011251522.0. According to the method, a channel attention mechanism is added in a U-Net network encoder, and prominent target features are screened from abundant low-level features, so that the influence of background noise is inhibited, and the accuracy of depth information fusion is improved; secondly, a spatial pyramid pooling module sensitive to road size features is added; and finally, a space attention mechanism is added into a decoder, so that the image restoring capability is improved.
Document 4, the invention of hangzhou electronic technology university chengzgang and others, an optical image phase unwrapping method based on U-Net segmentation network, application number: CN 201910003354.4. The method utilizes a Zernike polynomial to generate an optical image with an unwrapped phase, and performs phase wrapping operation to obtain a phase wrapping image; and (3) utilizing a U-Net segmentation network training model, and carrying out phase knot winding operation by using the trained model to obtain a phase unwrapping image. The method has high pixel segmentation accuracy, so that the solving accuracy is high.
Document 5, the invention of the southwest university of transportation yellow go, et al, "a land use type identification method based on U-Net neural network", application No.: CN 202011236409.5. According to the method, satellite pictures are cut, the cut pictures with standard sizes are respectively input into a U-Net network for feature extraction, upsampling reduction is carried out, and softmax classification is carried out on the reduction results. The method improves the accuracy of identifying the land use type after training a small number of samples.
Document 6 "a dam crack detection method based on U-net network and SC-SAM attention machine system" applied by liufan, et al, river and sea university, application No.: CN 202011049216.9. According to the method, a dam data set is collected and expanded, a deep learning segmentation network U-net model is constructed, an SC-SAM (Standard template-sequence analysis) attention mechanism is added on the basis, the attention mechanism is composed of two parts, CAM improves crack channel weight in a characteristic diagram, SAM improves crack area weight in a space domain in the characteristic diagram, and the accuracy of dam crack detection of the model is greatly improved due to mutual cooperation of the two parts.
The above patent applications do not relate to the extraction of the U-Net neural network ice lake combined with the self-attention mechanism, so the invention provides the method for extracting the U-Net neural network ice lake combined with the self-attention mechanism, and the accurate extraction of the ice lake is realized.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to extract large-range ice lake distribution information, provides a U-Net neural network ice lake extraction method combined with a self-attention mechanism, and provides scientific basis for scientific research on understanding of spatial distribution, climate and water resources of ice lakes.
The technical scheme is as follows: the invention relates to a U-Net neural network ice lake extraction method combined with a self-attention mechanism, which sequentially comprises the following steps.
(1) And acquiring Landsat 8 OLI and TIRS images of a research area, resampling to the same pixel size after radiometric calibration and atmospheric correction, and embedding. A portion of the ice lake surface of the study area was mapped using visual interpretation and converted to a grid.
(2) And manufacturing a supervised classification training grid data set, cutting all band images and visual interpretation images by using a sliding window, expanding the data set through rotation and mirror image operation, and dividing the data set into an 80% training set, a 10% testing set and a 10% verifying set.
(3) And selecting an input waveband of a comparison experiment, and selecting Landsat 8 OLI 2, 5 and 6 wavebands as input wavebands of a comparison full-waveband input experiment according to the wavelength adopted by the algorithm of the normalized differential water body index and the normalized differential snow cover index.
(4) And constructing a standard U-Net network model with Landsat 8 OLI 3, 5 and 6 waveband inputs, a U-Net network model with 3 waveband inputs and a self-attention module, a standard U-Net network model with Landsat 8 OLI and TIRS full waveband inputs, and a U-Net network model with a full waveband input and a self-attention module. And (5) comparing the segmentation precision of the four network models through experiments.
(5) The method comprises the steps of performing model training by adopting a U-Net network model with a full-waveband input and a self-attention module, constructing a data set required by the model, inputting the data set into the U-Net network model with the self-attention module to learn the spectrum and texture characteristics of the ice lake, selecting an optimal training model by adjusting hyper-parameters, and finally extracting the ice lake in the whole research area.
The invention has the beneficial effects that: the invention discloses a U-Net neural network ice lake extraction method combined with a self-attention mechanism, which efficiently and accurately extracts ice lake space distribution information in a large range by combining the U-Net neural network combined with the self-attention mechanism, and provides reference data and technical support for ice lake disasters in alpine regions.
Drawings
FIG. 1 is a flow of U-Net neural network ice lake extraction combined with a self-attention mechanism.
FIG. 2 shows the results of the U-Net neural network ice lake extraction in combination with the self-attention mechanism.
Detailed Description
The technical solution of the present invention will be clearly and completely described in conjunction with the drawings in the present invention, and the described embodiment is only a part of the embodiment of the present invention, and not all of the embodiment.
A U-Net neural network ice lake extraction method combined with a self-attention mechanism is implemented.
The method comprises the following steps: and acquiring Landsat 8 images of the research area and drawing Label data for training.
(1) Data preprocessing: carrying out radiometric calibration and atmospheric correction on 3 scene Landsat 8 OLI and TIRS images covering a research area, resampling to the same pixel size (30 m), and finally inlaying the three scene images, wherein the overlapped part adopts a MAX (maximum value selection) mode.
(2) Training label data production: visually interpreting an ice lake region in ArcGIS, drawing a boundary polygon, using a surface-to-grid tool, reclassifying the ice lake region into 255 (white) with the pixel size being the same as that of the remote sensing image, setting the rest regions to be 0 (black), and storing as a single-waveband 8-bit unsigned integer grid image.
Step two: and manufacturing a supervised classification training grid data set.
(1) Cutting data: the two data are cut into images by using three sliding windows with scales of 1, 2 and 4 respectively, and the output image sizes are 256 multiplied by 256, 128 multiplied by 128 and 64 multiplied by 64.
(2) Augmented data: and carrying out mirror image and rotation operations on the two kinds of cut data.
(3) The output image not only contains the original slice image, but also contains the image which mirrors and rotates the slice image to enrich the training data. Finally, the data set is divided into 80% training set, 10% testing set and 10% validation set. In order to simplify and balance the training data set and improve the final prediction precision, if the original image position of the slice remote sensing image does not have the ice lake region, the image is not output.
Step three: and selecting an input waveband of a comparison experiment.
According to the wavelengths adopted by the algorithm of the normalized differential water body index (NDWI) and the Normalized Differential Snow Index (NDSI), Landsat 8 OLI 3, 5 and 6 wave bands are selected as input wave bands of a contrast full-wave band input experiment.
Step four: and constructing four neural network models for effect comparison.
The system comprises a standard U-Net network model with Landsat 8 OLI 3, 5 and 6 waveband inputs, a U-Net network model with 3 waveband inputs and a self-attention module, a standard U-Net network model with Landsat 8 OLI and TIRS full waveband inputs, and a U-Net network model with a full waveband input and a self-attention module. And comparing the segmentation precision of the four network models in an experiment, and selecting the model with the best effect to carry out hyper-parameter adjustment so as to optimize the output effect.
Step five: and performing ice lake extraction on the research area by using the selected optimal model.
And (3) carrying out ice lake region extraction on the complete research region by utilizing a U-Net network model with self-attention modules and Landsat 8 OLI and TIRS full-wave band input.

Claims (8)

1. A method for extracting an ice lake region by combining a U-Net neural network model of a self-attention mechanism is characterized by comprising the following steps: acquiring remote sensing images and label data of a research area; and step two, making a data set, selecting a comparison test input waveband, constructing a neural network model, comparing effects, and extracting the ice lake in the research area by using the model with the best effect.
2. The method for acquiring remote sensing images and label data of a research area according to claim 1, wherein: and acquiring Landsat 8 images of the research area and drawing Label data for training.
3. And acquiring Landsat 8 OLI and TIRS images of the research area, and acquiring full-waveband images of the research area through steps of radiometric calibration, atmospheric correction, resampling, embedding and the like.
4. And (3) performing vector drawing on part of the ice lake in the research area by utilizing visual interpretation, and obtaining an 8-bit unsigned integer single-waveband grid image with the white ice lake area and the black rest part through the steps of surface-to-grid, reclassification and the like.
5. A method of producing a data set according to claim 1, wherein: and manufacturing a supervised classification training grid data set, cutting an original image by utilizing three scaling sliding windows and performing data augmentation.
6. The method of selecting a contrast trial input band as in claim 1, wherein: according to the wavelengths adopted by the algorithm of the normalized differential water body index (NDWI) and the Normalized Differential Snow Index (NDSI), Landsat 8 OLI 3, 5 and 6 wave bands are selected as input wave bands of a contrast full-wave band input experiment.
7. The method of constructing a neural network model and comparing effects of claim 1, wherein: and constructing four U-Net and variant neural network models thereof, carrying out precision evaluation on the output image, selecting the optimal model and carrying out parameter adjustment to optimize the output effect.
8. The best effort model for research area ice lake extraction as claimed in claim 1, characterized by: and (3) carrying out ice lake extraction on the complete research area by utilizing a U-Net network model with self-attention modules and input in Landsat 8 OLI and TIRS full wave bands.
CN202110659969.XA 2021-06-15 2021-06-15 U-Net neural network ice lake extraction method combined with self-attention mechanism Pending CN113435484A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110659969.XA CN113435484A (en) 2021-06-15 2021-06-15 U-Net neural network ice lake extraction method combined with self-attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110659969.XA CN113435484A (en) 2021-06-15 2021-06-15 U-Net neural network ice lake extraction method combined with self-attention mechanism

Publications (1)

Publication Number Publication Date
CN113435484A true CN113435484A (en) 2021-09-24

Family

ID=77755946

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110659969.XA Pending CN113435484A (en) 2021-06-15 2021-06-15 U-Net neural network ice lake extraction method combined with self-attention mechanism

Country Status (1)

Country Link
CN (1) CN113435484A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113936226A (en) * 2021-11-23 2022-01-14 河南大学 Global glacier search identification method based on remote sensing cloud computing
CN116000945A (en) * 2022-12-05 2023-04-25 湖北工业大学 Intelligent control method of cable deicing robot
CN117233764A (en) * 2023-11-14 2023-12-15 兰州交通大学 InSAR phase unwrapping method based on R2AU-Net

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107730527A (en) * 2017-10-16 2018-02-23 中国科学院遥感与数字地球研究所 A kind of highlands glade extracting method based on satellite-remote-sensing image
CN111932567A (en) * 2020-07-30 2020-11-13 中国科学院空天信息创新研究院 Satellite image-based ice lake contour automatic extraction method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107730527A (en) * 2017-10-16 2018-02-23 中国科学院遥感与数字地球研究所 A kind of highlands glade extracting method based on satellite-remote-sensing image
CN111932567A (en) * 2020-07-30 2020-11-13 中国科学院空天信息创新研究院 Satellite image-based ice lake contour automatic extraction method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
YI HE 等: "An Extraction Method for Glacial Lakes Based on Landsat-8 Imagery Using an Improved U-Net Network", 《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》 *
赵航等: "基于改进C-V模型的冰湖轮廓提取方法研究", 《遥感技术与应用》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113936226A (en) * 2021-11-23 2022-01-14 河南大学 Global glacier search identification method based on remote sensing cloud computing
CN116000945A (en) * 2022-12-05 2023-04-25 湖北工业大学 Intelligent control method of cable deicing robot
CN116000945B (en) * 2022-12-05 2024-04-12 湖北工业大学 Intelligent control method of cable deicing robot
CN117233764A (en) * 2023-11-14 2023-12-15 兰州交通大学 InSAR phase unwrapping method based on R2AU-Net
CN117233764B (en) * 2023-11-14 2024-01-30 兰州交通大学 InSAR phase unwrapping method based on R2AU-Net

Similar Documents

Publication Publication Date Title
Wieland et al. Multi-sensor cloud and cloud shadow segmentation with a convolutional neural network
Ji et al. Simultaneous cloud detection and removal from bitemporal remote sensing images using cascade convolutional neural networks
CN113435484A (en) U-Net neural network ice lake extraction method combined with self-attention mechanism
Hou et al. Marine floating raft aquaculture extraction of hyperspectral remote sensing images based decision tree algorithm
CN107463901B (en) Multi-scale regional flood disaster risk remote sensing evaluation method and system
Wu et al. A geographic information-driven method and a new large scale dataset for remote sensing cloud/snow detection
CN112288647B (en) Remote sensing image cloud and shadow restoration method based on gating convolution
CN111932567B (en) Satellite image-based ice lake contour automatic extraction method
CN106600574B (en) A kind of landslide extracting method based on remote sensing image and altitude data
Ursani et al. Fusion of textural and spectral information for tree crop and other agricultural cover mapping with very-high resolution satellite images
CN107392130A (en) Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks
Mohajerani et al. Cloud and cloud shadow segmentation for remote sensing imagery via filtered jaccard loss function and parametric augmentation
CN113033670A (en) Method for extracting rice planting area based on Sentinel-2A/B data
Sun et al. Characterizing tree species of a tropical wetland in southern china at the individual tree level based on convolutional neural network
CN113239736B (en) Land coverage classification annotation drawing acquisition method based on multi-source remote sensing data
Wang et al. Fusion of HJ1B and ALOS PALSAR data for land cover classification using machine learning methods
Zhu et al. Estimating and mapping mangrove biomass dynamic change using WorldView-2 images and digital surface models
Kim et al. Detection of swimming pools by geographic object-based image analysis to support west Nile virus control efforts
Gonzalez et al. Application of semantic segmentation with few labels in the detection of water bodies from perusat-1 satellite’s images
CN116452872A (en) Forest scene tree classification method based on improved deep pavv3+
Sah et al. Semi-automated cloud/shadow removal and land cover change detection using satellite imagery
Xie et al. Object-oriented random forest classification for Enteromorpha prolifera detection with SAR images
Ouerghemmi et al. Urban vegetation mapping using hyperspectral imagery and spectral library
Huang et al. Comparing the effects of temporal features derived from synthetic time-series NDVI on fine land cover classification
Viana et al. Shoreline extraction using unsupervised classification on Sentinel-2 imagery

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210924

WD01 Invention patent application deemed withdrawn after publication