CN116503755A - Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning - Google Patents

Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning Download PDF

Info

Publication number
CN116503755A
CN116503755A CN202310570811.4A CN202310570811A CN116503755A CN 116503755 A CN116503755 A CN 116503755A CN 202310570811 A CN202310570811 A CN 202310570811A CN 116503755 A CN116503755 A CN 116503755A
Authority
CN
China
Prior art keywords
shoreline
coastline
remote sensing
area
cloud platform
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
CN202310570811.4A
Other languages
Chinese (zh)
Inventor
陈博伟
张丽
左健
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.)
Aerospace Information Research Institute of CAS
Original Assignee
Aerospace Information Research Institute of CAS
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 Aerospace Information Research Institute of CAS filed Critical Aerospace Information Research Institute of CAS
Priority to CN202310570811.4A priority Critical patent/CN116503755A/en
Publication of CN116503755A publication Critical patent/CN116503755A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • 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
    • G06N3/0455—Auto-encoder networks; Encoder-decoder 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A10/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE at coastal zones; at river basins
    • Y02A10/40—Controlling or monitoring, e.g. of flood or hurricane; Forecasting, e.g. risk assessment or mapping

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a shoreline remote sensing automatic identification analysis method based on cloud platform and deep learning, which comprises the following steps: realizing the block downloading and preprocessing of the large-area multi-source remote sensing image based on the cloud platform; coastline sample dataset development across a wide geographic area involving different coastline type scenarios under high water level and low water level conditions; feature importance screening based on a random forest average precision decreasing analysis method; deep learning network training and precision evaluation are carried out by applying image features and coastline sample data sets; generating a shoreline binary map and converting an actual shoreline vector based on a Canny edge detection algorithm by using a network model self-adaptive sea Liu Fenge; and (5) extracting a long-time sequence shoreline result and performing tidal correction post-treatment. The invention makes the extraction error of the shoreline smaller than the theoretical maximum allowable error, and can meet the requirement of automatic detection of long-time sequence large-scale complex shorelines under multiple scenes.

Description

Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning
Technical Field
The invention belongs to the field of coastline remote sensing identification, and particularly relates to a cloud platform and deep learning-based automatic coastline remote sensing identification analysis method.
Background
The sea-land interaction zone at the coastal zone is not only affected by global environmental problems such as climate change, ocean acidification, sea Liu Shengjing change and the like, but also severely affected by human activities such as construction of artificial islands around the filled sea, cultivation of offshore ponds, and urban construction of ports and the like. The coastline is taken as a coastal boundary, is a most visual coastal element reflecting the current situation of the ecological environment of the coastal zone, and the position, length and type changes of the coastline are closely related to the climate change and human activity problems affecting the coastal zone. The remote sensing is used as a non-contact remote detection technology, and particularly the continuous increase of medium-high resolution remote sensing satellites in recent years gradually becomes an important data source for coastline monitoring, and coastline change monitoring based on long-time sequence remote sensing data becomes an important means and key indexes for coastal zone environment monitoring and evaluation.
Coastline remote sensing extraction is initially accomplished by field surveying or visual interpretation of unified interpretation standards, which is also the method with the highest accuracy of coastline extraction, but severely limited by efficiency, cost and subjectivity. Then, with the development of remote sensing computer interpretation technology, a large number of coastline automatic extraction methods appear: zhang et al propose a combination of integrated image segmentation, region growing and multispectral image edge detection to improve shoreline extraction accuracy; chen Weitong and the like, and multi-temporal shoreline data is extracted by combining an improved water side line method; dai et al, repeatedly measuring superimposed normalized differential water index shoreline extraction method by statistical analysis to mitigate image offset and cloud image errors. In a word, a semi-automatic extraction method based on water edge extraction and post correction of an optical remote sensing image, an extraction method based on LiDAR data for solving a tide level and an intelligent extraction method based on a remote sensing image and interpretation standards are widely used shore remote sensing extraction methods at present, and meanwhile, development of a remote sensing intelligent cloud platform brings new possible and low-cost solutions to shore extraction research.
The deep learning is widely applied to various fields due to the capability of extracting image features and fitting complex problems, a large number of deep learning algorithms are also generated in the field of shoreline remote sensing extraction, a sea island shoreline remote sensing image segmentation model is provided by using an improved deep network such as Wang Zhenhua, and the result shows that the method overcomes the problem of sea island shoreline discontinuity caused by other models and reduces the phenomenon of sea island shoreline erroneous segmentation; seale et al propose a Sentinel-2 water line dataset for training and testing shorelines automatically extracted from the Sentinel-2 images, and based on U-Net model architecture four convolutional neural network models were trained, tested and optimized to obtain a shoreline type and element dataset from around the world. The coastline extraction algorithm based on the deep learning remote sensing image segmentation algorithm is limited by the following reasons: (1) The image segmentation algorithm based on deep learning is seriously dependent on long-time training, a large number of data sets and manual labeling, and challenges are presented to the automatic acquisition capability of remote sensing data; (2) The shorelines in different scenes have different characteristics, and the characteristics of multiple scenes and multiple wave bands provide challenges for the input of the deep learning model.
Disclosure of Invention
In order to solve the technical problems, the invention provides a shoreline remote sensing automatic identification analysis method based on a cloud platform and deep learning, so that the extraction error of the shoreline is smaller than the theoretical maximum allowable error, and the long-time sequence large-range complex shoreline automatic detection requirement under multiple scenes can be met.
In order to achieve the above purpose, the invention adopts the following technical scheme:
a shoreline remote sensing automatic identification analysis method based on cloud platform and deep learning comprises the following steps:
step S1, carrying out remote sensing image block downloading of a large area based on a cloud platform;
s2, preprocessing the remote sensing image of the designated area, wherein the preprocessing comprises cloud processing, full-color sharpening and downsampling;
s3, developing a coastline sample data set for training a coastline detection method based on Landsat and Sentinel series data, wherein the coastline sample data set spans a wide geographic area and comprises various coastline type scenes under high water level and low water level conditions;
s4, carrying out importance analysis on the input features based on an average precision decreasing analysis method of the random forest, and determining the input features most suitable for shoreline extraction;
s5, constructing and parameter setting selected U-net, deep Labv3+, deep Res U-net, R2U-net, attention U-net, res U-net++, SAnet and U-net3+ neural network models;
step S6, training the neural network model in the step S5 by taking training data sets and verification data sets in different scenes as inputs to obtain the neural network model under the optimal weight of each scene;
s7, sea-land segmentation is carried out on the test data of the neural network model under the optimal weight of different scenes, a Canny edge detection algorithm is used for generating a coastline binary image of the segmented result, the coastline binary image is converted into an actual coastline vector, and finally accuracy comparison is carried out on the integrity complex and accuracy correct index of the reference coastline vector which is interpreted visually, so that the neural network model with the highest accuracy under different scenes is obtained;
and S8, large-area shoreline result extraction and tide correction post-processing, classifying and marking the downloaded large-area remote sensing image as input, adaptively selecting a neural network model with highest classification precision under different scenes, carrying out long-time-sequence shoreline extraction, and carrying out tide correction on the shoreline extraction result.
Further, in the step S1, the method includes: the images are acquired, downloaded and cropped through the GEE API, and the blocking process required due to memory limitations is achieved by building a grid of 10km x 10 km.
Further, in the step S3, the method for constructing the coastline sample data set includes:
the sample area relates to different kinds of shoreline distribution areas of all the big continents except antarctic, and comprises 6 shoreline type scenes including an artificial shoreline, a sandy shoreline, a biomass shoreline, a bedrock shoreline, a river mouth shoreline and a silt shoreline.
Further, in the step S4, the index features in the input features include a simple ratio vegetation index SR, an enhanced vegetation index EVI, a normalized difference vegetation index NDVI, a normalized difference water index NDWI, and an improved normalized difference water index MNDWI, which are respectively:
wherein, B2 is a visible blue light wave band, B3 is a visible green light wave band, B4 is a visible red light wave band, B8 is a near infrared wave band, and B11 is a short wave infrared wave band.
Further, in the step S7, an extracted shoreline is obtained from an actual shoreline, buffer areas are respectively made for the reference shoreline and the extracted shoreline, and the in-area part of the extracted shoreline is denoted as TP1 and the out-of-area part is denoted as FP in the buffer area centered on the reference shoreline; in the buffer area taking the extracted shoreline as the center, the part in the area of the reference shoreline is marked as TP2, the part outside the area is marked as FN, and the calculation formulas of the integrity and the accuracy are as follows:
the integrity Complete describes the percentage of the correct coastline in the extracted coastline, and the integrity of the coastline extraction result is evaluated; accuracy Correct describes the specific gravity of the Correct extraction in the reference shoreline.
The beneficial effects are that:
the cloud platform and the deep learning algorithm are applied to semantic segmentation of the remote sensing image and applied to shoreline extraction of the remote sensing image, so that more manpower and material resources are saved. Aiming at the problem of lack of a large number of data sets and manual annotation in large-area automatic extraction, the GEE API-based automatic acquisition of long-time-sequence large-area remote sensing images is realized, and a coastline sample data set which spans a wide geographic area and relates to different coastline type scenes under high water level and low water level conditions is constructed. Aiming at characteristics of multiple scenes and multiple wave bands of the shoreline, the method fully adapts to different scenes by training the self-adaptive neural network model under different scenes, and improves the shoreline extraction precision under different scenes. By the research, not only can the coastline be monitored rapidly and the information such as the change dynamics of the coastline be mastered, but also the coast erosion and accumulation degree can be monitored, and the coastal environment can be tracked and monitored dynamically.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the description of the embodiments or the prior art will be briefly described, and it is obvious that the drawings in the description below are some embodiments of the present invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a shoreline remote sensing automatic identification analysis method based on a cloud platform and deep learning according to an embodiment of the invention;
FIG. 2 is a diagram illustrating accuracy verification of shoreline extraction result integrity complete and accuracy correct according to an embodiment of the present invention; wherein a is a matched extraction result diagram, and b is a matched reference coastline diagram;
fig. 3 is a schematic diagram of a U-net original network structure according to an example of the present invention.
Detailed Description
Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily implement them. In addition, for the sake of clarity, portions irrelevant to description of the exemplary embodiments are omitted in the drawings.
In this disclosure, it should be understood that terms such as "comprises" or "comprising," etc., are intended to indicate the presence of features, numbers, steps, acts, components, portions, or combinations thereof disclosed in this specification, and do not preclude the presence or addition of one or more other features, numbers, steps, acts, components, portions, or combinations thereof.
In addition, it should be noted that, without conflict, the embodiments of the present disclosure and features of the embodiments may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.
The invention discloses a shoreline remote sensing automatic identification analysis method based on cloud platform and deep learning, which provides remote sensing image data with strong cloud computing capability and multi-source long time sequence based on a global scale remote sensing cloud computing platform GEE, can automatically acquire, screen and cut remote sensing images required by a selected area, and integrates remote sensing image preprocessing methods such as cloud processing, full color sharpening, downsampling and the like. On the basis, sea-land segmentation samples with full semantic level under different scenes are constructed, 9 kinds of neural network models including U-net and deep Labv3+ are trained, the optimal neural network model is obtained by integrating the extraction precision under different scenes, the identification of complex shorelines is realized, and real shoreline data are obtained after post-processing operations such as tide correction and the like.
The invention can realize automatic extraction of the shoreline under different scenes, and the blocking processing technology can ensure the operation speed under the condition of limited memory, thereby saving the running cost of the system; according to the method, the neural network models in different scenes are trained, so that the model with optimal performance is obtained to extract the complex shoreline, and the operation precision is greatly improved. Therefore, the invention can extract seeds in a large range at a complex shoreline automatic long time sequence to ensure higher speed and accuracy.
As shown in fig. 1, the method for automatically identifying and analyzing the shoreline remote sensing based on the cloud platform and the deep learning according to the embodiment of the invention comprises the following steps:
step one, carrying out large-area remote sensing image block downloading based on a cloud platform, wherein the step one comprises the following steps:
automatic acquisition of multi-source satellite images: an integrated Google Earth Engine (GEE) API interface, providing top-of-the-atmosphere reflectivity (TOA) images from Landsat5TM, landsat7 ETM+, landsat8 OLI Tier 1 collection, and the Sentinel-2MSI level-1C product, can crop images within a specified time frame downloaded from the GEE server by defining a region of interest. In addition, the downloaded image only contains the spectrum bands after the coastline detection feature is preferable, namely three visible light bands (R, G, B), a near infrared band (NIR) and two short wave infrared bands (SWIR), so that the processing speed is remarkably improved and the memory use on a local computer is reduced.
Specifically, GEE (Google Earth Engine) is a tool for google that can process satellite image data in bulk, providing an offline python API and an online JavaScript API. First, the python API library of the GEE is locally configured, and the python API client of Google, the authentication dependency verification library and the python library of the GEE are installed. The local integration of GEE is completed by importing a geemap package, the automatic downloading of images is completed by defining the area range, the date range and the satellite type which need to be downloaded, meanwhile, the images are cut according to the range of the defined area, the downloading range which is larger than 100 square kilometers is subjected to blocking processing due to the limitation of a memory, and the method is realized by establishing a grid of 10km multiplied by 10km on a research area.
Step two, preprocessing the data of the multi-source satellite image: and carrying out pretreatment works such as cloud processing, full-color sharpening, downsampling and the like before the shoreline is extracted.
The cloud processing includes: the cloud coverage area is supplemented by a quality assessment band pre-calculated by a data provider (Landsat is USGS, sentinel-2 is ESA), wherein a cloud mask containing all cloud area pixels is included, cloud coverage is calculated according to the number of cloud pixels of a defined region of interest on the basis of the cloud mask, and cloud area images exceeding a certain coverage are discarded based on a defined cloud coverage threshold.
Specifically, the Landsat and Sentinel series data acquired by the GEE will have QA (data product quality assessment) band, which is used to evaluate the quality of each pixel and extract the pixel meeting the requirement, and is mainly used to extract the cloud pixel information and mask, the value stored in the QA band is generally displayed in decimal, and the value needs to be converted into binary value to judge the pixel state of the region. After judging the input remote sensing image, judging the band characteristics of the input image, screening out a cloud band, generating a cloud mask through a create_closed_mask function, checking other bands of the image, extracting and merging a null region into the cloud mask, and finally removing pixels at corresponding positions through the cloud mask.
The full color sharpening and downsampling includes: the spatial resolution of the full-color band of the image is enhanced to achieve the best coastline monitoring effect, for Landsat7 and Landsat8 images, the full-color band with higher resolution is used to increase the spatial resolution of the multispectral band from 30m to 15m by applying a data fusion method of principal component analysis, the multispectral band is downsampled to 15m by bilinear interpolation and decomposed into principal components, then the first principal component is replaced by the full-color band after histogram matching, and the original multispectral space is reconverted. For Landsat5 images, full-color wave bands are not available, 30m is directly downsampled to 15m through bilinear interpolation, and therefore accuracy of shoreline detection is improved. For a Sentinel-2 image, 20m short wave infrared bands are downsampled to 10m through bilinear interpolation, and the spatial resolution of all bands is unified to 10m.
Step three, training and verifying data set labeling: coastline sample data sets for coastline detection method training were developed based on Landsat and Sentinel series data, which span a wide geographic area, involving various coastline type scenarios in high water level and low water level conditions.
The image data is downloaded by the cloud platform image automatic acquisition technology, the selected images basically meet the characteristics of clearness, no cloud and the like, and the sample area relates to different kinds of shoreline distribution areas of all the global continents except antarctic continents, including 16 training sites and 49 testing sites, and 6 shoreline type scenes of artificial shorelines, sandy shorelines, biomass shorelines, bedrock shorelines, estuary shorelines and silt shorelines. Labeling the selected images, and creating dense pixel-level labels of two categories of 'water' and 'non-water', wherein the specific process is as follows:
creating a label by a semi-supervised clustering method, and firstly, performing false color synthesis on an image to improve the contrast between water and non-water pixels; secondly, applying a K-means clustering method to the image after pseudo-color rendering, obtaining the best distinguishing effect by continuously optimizing the clustering class number K, and then merging the classes until the last two classes of water and non-water; and finally, comparing the image with a high-definition image provided by Google Earth, and manually correcting the misplaced pixels.
And (3) after the data set is manufactured, randomly dividing the data set into a training set and a verification set according to the proportion of 7:3. Before the training set image is used as the training algorithm input, the training set image is normalized, and the training data set is subjected to data expansion, namely, the image color is adjusted, and the image is subjected to rotational symmetry operation to generate a new image.
And fourthly, analyzing the importance of the input features based on an average precision decreasing analysis method of the random forest, and determining the input features most suitable for shoreline extraction.
By an average decreasing precision (Mean decrease accuracy) analysis method based on random forests, 5 multispectral wave bands (R, G, B, NIR and SWIR 1), 5 common spectral indexes (simple ratio vegetation index SR, enhanced vegetation index EVI, normalized difference vegetation index NDVI, normalized difference water index NDWI and improved normalized difference water index MNDWI) and variances of each spectral wave band and index are selected as input features of a subsequent model.
Wherein, B2 is a visible blue light wave band, B3 is a visible green light wave band, B4 is a visible red light wave band, B8 is a near infrared wave band, and B11 is a short wave infrared wave band.
Step five, constructing a neural network model and setting parameters: 9 neural network models are selected for coastline detection under different scenes, wherein the structures of other models (Deep Res U-net, R2U-net, attention U-net, res U-net++, SAnet and U-net3+) basically take the U-net model as a main body, and different modules and combinations are added for optimization. The modules of these model applications are respectively: residual (Residual), recursive (return), attention Gate (Attention Gate), compression stimulus (Squeeze Excitation), and Dense Connect (Dense Connect) structures, as shown in table 1.
TABLE 1
Wherein: publishing time for each model under the first row name; "" indicates that the model is provided with this module.
As shown in FIG. 3, the U-Net network is composed of a downsampling part and an upsampling part, is similar to a 'U', and based on an Encoder-Decoder structure, realizes feature fusion in a splicing mode, and is simple and stable in structure. The first part performs feature extraction on the image through convolution layers and max-pooling layers, each 3 x 3 convolution layer followed by an activation function ReLU and a 2 x 2 max-pooling operation. The second part is subjected to deconvolution operation, then the result and the corresponding feature map are spliced, and a 1 multiplied by 1 convolution kernel is adopted at the final output layer. The body of the deep labv3+ Encoder is DCNN with hole convolution followed by a spatial pyramid pooling module (Atrous Spatial Pyramid Pooling, ASPP) with hole convolution to introduce multi-scale information. Compared with other DeepLab networks, the method introduces a Decoder module, fuses the bottom layer features with the high layer features, and further improves the accuracy of the segmentation boundary. The Residual (Residual) structure solves a series of problems that after the network depth is continuously increased, the model is difficult to train, the model is easy to overfit, the gradient is easy to disappear or explode, and the like. The conventional convolution module is split into two parts of direct mapping and residual short circuit, and the two parts are added at the output part, so that the information lost in the convolution of a network is effectively compensated, meanwhile, the gradient has the participation of the residual part in the back propagation, and the stability of the gradient range in a controllable range is ensured. The recursive residual convolution (Recurrent Residual Convolutional) structure is a combination of recursive convolution and residual structure, called a cyclic residual U-network (R2U-net). The recursive convolution is arranged at the tail part of each layer of the U-net coding part, two recursions are set by default, the layer number of the convolution of each layer is increased, and more context information and more deep features of the image participate in learning. And a residual structure is introduced, so that the problems of gradient gradually disappearing and data original information losing during back propagation caused by the increase of the network depth of the recursion structure are solved. An Attention Gate (AG) structure introduces an Attention mechanism into U-net, called Attention U-net. The 'attention gate' module belongs to soft attention of a space domain, the structure is arranged at a jump joint of the U-net decoding part, the feature map after up sampling is fused with the feature map after the down sampling of the upper layer, attention coefficients of the fused features are obtained through 'point-by-point convolution' learning, and information of a key region (with high attention coefficients) is amplified. Dense connection (Dense connection) structures, a portion of which is "Dense connected" to other feature layers, is stripped from the conventional convolution. Unlike the residual structure, the feature layers are spliced at the channel level when they are "densely connected", and almost all layers are connected from the entire network, which enables the network information to be utilized to the maximum. The method has the advantages of high effective information utilization, network gradient disappearance reduction and redundancy parameter reduction.
Training a neural network model: training the neural network model by taking training data sets and verification data sets in different scenes as input to obtain the neural network model under the optimal weight of each scene.
The experimental environment is a graphic workstation running a Windows 10 operating system 64 bit, and the specific hardware environment parameters are: the Core Processor (CPU) isW-2223@3.60GHz, the image processor (GPU) is NVIDIA GeForce RTX 309024GB, and the memory (RAM) is 64GB; the software environment is as follows: deep learning frameworks PyTorch (version GPU1.7.0) and CUDA (version 11.1), extended libraries GDAL (version 2.3.3) and OpenCV (version 1.4) for data enhancement.
The experiment trains 9 neural networks in a unified environment to compare the extraction precision under different network structures. The specific strategy is as follows:
(1) The weights of all layers (particularly the activation function layer) of the network are initialized by uniformly using a Kaiming initialization method, so that the loss convergence speed of each network is improved, and the problems of gradient disappearance and the like of each network are prevented;
(2) Training is carried out by adjusting the training learning rate from the main body through an Adam learning rate optimizer so as to accelerate the gradient descent of the network and avoid the network loss from being stagnated at a local minimum value;
(3) The initial learning rate of the model is set to be 0.001, and once the network trains, the model is verified on a verification set, and verification accuracy (verification accuracy is achieved through an average cross-over ratio (mIoU) and an average pixel accuracy (mPA)) is recorded and monitored.
Average cross-over (Mean Intersection over Union, mIoU), i.e., the ratio of the intersection of the predicted and true results of each class in space to the union (abbreviated cross-over, ioU) is calculated, followed by the arithmetic average of all classes. The method can reflect the accuracy of the semantic segmentation result in space, and the higher the intersection ratio is, the better the segmentation effect is, and the calculation formula is as follows:
wherein TP is the intersection part of the true label and the intersection result, FN is the true label, FP is the predicted result, and k is the category number.
Average pixel accuracy (Mean Pixel Accuracy, mPA), i.e. the proportion of the number of pixels with correct segmentation in each class in the prediction result to the total number of pixels in the class is calculated, and the arithmetic average of the pixel accuracy of each class is calculated. The index reflects the capability of algorithm segmentation by dividing correct pixel proportion in the result, the higher the accuracy is, the better the segmentation effect is, and the following calculation formula is adopted:
where pii represents the number of pixels that would belong to class i but are predicted to be class j, i being the correct class, j being the predicted class, and k being the number of classes.
The mPA is mainly used for evaluating the effect of the model on the land type judgment and the position of each pixel on the verification set, the mPOU is used for evaluating the effect of the model on the land area and the boundary segmentation on the verification set, the change trend of the model is basically the same, the combination of the model and the boundary segmentation can comprehensively represent the precision of the model, if the verification precision is not obviously improved in two rounds, the learning rate is reduced by 1%, the network loss is accelerated to converge, and the local minimum value is jumped out as soon as possible;
(4) Setting the initial training round as 20, and continuing training the model which is not fully converged in a 'Warm start' (Warm up) mode until the model approaches to a fitting state;
(5) And selecting the round with optimal verification set precision as the final training result of the network.
Step seven, sea-land segmentation is carried out on the neural network model under the optimal weight of different scenes on test data, a binary coastline graph is generated by a Canny edge detection algorithm on the segmentation result, the binary coastline graph is converted into a real coastline vector, and finally accuracy comparison is carried out on the binary coastline graph and a visually interpreted reference coastline vector on two indexes of integrity complex and accuracy correction, so that the neural network model with the highest accuracy under different scenes is obtained;
the method for calculating the integrity and the accuracy of the extracted shoreline based on the real shoreline reference data by adopting a line target matching method comprises the following specific steps:
firstly, respectively taking a reference shoreline and an extracted shoreline as buffer areas, and taking the part in the extracted shoreline area as TP1 and the part outside the extracted shoreline area as FP in the buffer areas of the reference shoreline; in the buffer area taking the extracted shoreline as the center, the part in the reference shoreline area is marked as TP2, the part outside the area is marked as FN, and thus the calculation formulas of the integrity and the accuracy for evaluating the result precision are as follows:
the integrity is used for describing the percentage of the coastline which is correctly divided in the extracted coastline, and evaluating the integrity of the result; the accuracy describes the specific gravity of the extracted shoreline that the reference shoreline was extracted correctly. For a common shoreline extraction result without a real shoreline reference, a mode of acquiring a verification point from a Google Earth high definition image is adopted to calculate a shoreline average Offset Mean Offset and a root Mean square error RMSE (Root Mean Square Error), and the formula is as follows:
wherein n is the verification point number, D n The euclidean distance from the point to the shoreline results is verified. The average offset is used to describe the degree of offset of the shoreline result from the verification point, and the root mean square error is the standard deviation of the shoreline result and the verification point offset. The relationship between the theoretical maximum allowable error U of the shoreline extracted based on the remote sensing image and the image resolution r is as follows:
when the average offset and the root mean square error of the extracted shoreline are smaller than the theoretical maximum allowable error calculated by the image resolution, the accuracy of the extracted shoreline can be considered to accord with the standard; the results are shown in FIG. 2.
Step eight, step S8, large-area shoreline result extraction and tide correction post-processing, classifying and marking the downloaded large-area remote sensing image as input, adaptively selecting a neural network model with highest classification precision under different scenes, extracting a long-time-sequence shoreline, and carrying out tide correction on the shoreline extraction result.
In order to ensure that the shoreline extraction result is closer to the shoreline in the true sense, tidal correction is required to be carried out on the shoreline extraction result, and the correction method is as follows:
firstly, the shore lines of two scenes at different imaging moments are required to be extracted to be C1 and C2 respectively, the horizontal distance between the C1 and the C2 is delta L, theta is the coast gradient, H1 and H2 are the tide level heights (H2 > H1) of the two scenes of image satellites when the two scenes of the image satellites pass through the border, H is the high tide level of the average big tide, L is the correction distance of the shore lines, and the gradient of the beach can be calculated by the following formula:
θ＝arctan[(h 2 -h 1 )/ΔL]
the distance from the shoreline C1 to the actual shoreline is:
L＝(H-h 2 )/tanθ
according to the shoreline tide correction formula, the extracted shoreline C1 is moved to the land by a distance L to obtain an actual shoreline, wherein the artificial, bedrock and estuary shoreline are insensitive to the tide influence and can be ignored. The biomass, silt and sandy shoreline algorithm takes the inside as a shoreline, and the influence of tides can be avoided to a certain extent by combining a tides correction formula.

Claims (5)

1. A shoreline remote sensing automatic identification analysis method based on cloud platform and deep learning is characterized by comprising the following steps:
step S1, carrying out remote sensing image block downloading of a large area based on a cloud platform;
s2, preprocessing the remote sensing image of the designated area, wherein the preprocessing comprises cloud processing, full-color sharpening and downsampling;
s3, developing a coastline sample data set for training a coastline detection method based on Landsat and Sentinel series data, wherein the coastline sample data set spans a wide geographic area and comprises various coastline type scenes under high water level and low water level conditions;
s4, carrying out importance analysis on the input features based on an average precision decreasing analysis method of the random forest, and determining the input features most suitable for shoreline extraction;
s5, constructing and parameter setting selected U-net, deep Labv3+, deep Res U-net, R2U-net, attention U-net, res U-net++, SAnet and U-net3+ neural network models;
step S6, training the neural network model in the step S5 by taking training data sets and verification data sets in different scenes as inputs to obtain the neural network model under the optimal weight of each scene;
s7, sea-land segmentation is carried out on the test data of the neural network model under the optimal weight of different scenes, a Canny edge detection algorithm is used for generating a coastline binary image of the segmented result, the coastline binary image is converted into an actual coastline vector, and finally accuracy comparison is carried out on the integrity complex and accuracy correct index of the reference coastline vector which is interpreted visually, so that the neural network model with the highest accuracy under different scenes is obtained;
and S8, large-area shoreline result extraction and tide correction post-processing, classifying and marking the downloaded large-area remote sensing image as input, adaptively selecting a neural network model with highest classification precision under different scenes, carrying out long-time-sequence shoreline extraction, and carrying out tide correction on the shoreline extraction result.
2. The automatic recognition analysis method of shoreline remote sensing based on cloud platform and deep learning according to claim 1, wherein the step S1 comprises: the images are acquired, downloaded and cropped through the GEE API, and the blocking process required due to memory limitations is achieved by building a grid of 10km x 10 km.
3. The automatic recognition analysis method of shoreline remote sensing based on cloud platform and deep learning according to claim 2, wherein in the step S3, the construction method of the shoreline sample data set comprises:
the sample area relates to different kinds of shoreline distribution areas of all the big continents except antarctic, and comprises 6 shoreline type scenes including an artificial shoreline, a sandy shoreline, a biomass shoreline, a bedrock shoreline, a river mouth shoreline and a silt shoreline.
4. The cloud platform and deep learning based shoreline remote sensing automatic identification analysis method according to claim 3, wherein in the step S4, the index features in the input features include a simple ratio vegetation index SR, an enhanced vegetation index EVI, a normalized difference vegetation index NDVI, a normalized difference water body index NDWI, and an improved normalized difference water body index MNDWI, which are respectively:
wherein, B2 is a visible blue light wave band, B3 is a visible green light wave band, B4 is a visible red light wave band, B8 is a near infrared wave band, and B11 is a short wave infrared wave band.
5. The automatic recognition analysis method of shoreline remote sensing based on cloud platform and deep learning according to claim 4, wherein in the step S7, an extracted shoreline is obtained from an actual shoreline, buffer areas are respectively made for a reference shoreline and the extracted shoreline, and in the buffer area with the reference shoreline as a center, the in-area part of the extracted shoreline is denoted as TP1, and the out-of-area part is denoted as FP; in the buffer area taking the extracted shoreline as the center, the part in the area of the reference shoreline is marked as TP2, the part outside the area is marked as FN, and the calculation formulas of the integrity and the accuracy are as follows:
the integrity Complete describes the percentage of the correct coastline in the extracted coastline, and the integrity of the coastline extraction result is evaluated; accuracy Correct describes the specific gravity of the Correct extraction in the reference shoreline.
CN202310570811.4A 2023-05-19 2023-05-19 Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning Pending CN116503755A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310570811.4A CN116503755A (en) 2023-05-19 2023-05-19 Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310570811.4A CN116503755A (en) 2023-05-19 2023-05-19 Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning

Publications (1)

Publication Number Publication Date
CN116503755A true CN116503755A (en) 2023-07-28

Family

ID=87320227

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310570811.4A Pending CN116503755A (en) 2023-05-19 2023-05-19 Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning

Country Status (1)

Country Link
CN (1) CN116503755A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117649598A (en) * 2023-10-23 2024-03-05 广东省国土资源测绘院 Offshore culture space distribution information monitoring method and system based on SAR (synthetic aperture radar) images

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117649598A (en) * 2023-10-23 2024-03-05 广东省国土资源测绘院 Offshore culture space distribution information monitoring method and system based on SAR (synthetic aperture radar) images

Similar Documents

Publication Publication Date Title
CN109934200B (en) RGB color remote sensing image cloud detection method and system based on improved M-Net
CN111598174B (en) Model training method based on semi-supervised antagonistic learning and image change analysis method
CN113160234B (en) Unsupervised remote sensing image semantic segmentation method based on super-resolution and domain self-adaptation
CN114202696A (en) SAR target detection method and device based on context vision and storage medium
CN113343789A (en) High-resolution remote sensing image land cover classification method based on local detail enhancement and edge constraint
CN112801270B (en) Automatic U-shaped network slot identification method integrating depth convolution and attention mechanism
CN110298211A (en) A kind of Methods Deriving Drainage Network based on deep learning and high-resolution remote sensing image
CN110991430B (en) Ground feature identification and coverage rate calculation method and system based on remote sensing image
Liu et al. Aquaculture area extraction and vulnerability assessment in Sanduao based on richer convolutional features network model
Zhang et al. Mask-R-FCN: A deep fusion network for semantic segmentation
CN114612769A (en) Integrated sensing infrared imaging ship detection method integrated with local structure information
CN116503755A (en) Automatic recognition analysis method for shoreline remote sensing based on cloud platform and deep learning
CN114494821A (en) Remote sensing image cloud detection method based on feature multi-scale perception and self-adaptive aggregation
CN117079163A (en) Aerial image small target detection method based on improved YOLOX-S
CN112686184A (en) Remote sensing house change detection method based on neural network
CN112037225A (en) Marine ship image segmentation method based on convolutional nerves
CN115937707A (en) SAR image water body extraction method based on multi-scale residual error attention model
Xia et al. Submesoscale oceanic eddy detection in SAR images using context and edge association network
Chen et al. A novel lightweight bilateral segmentation network for detecting oil spills on the sea surface
CN115240066A (en) Remote sensing image mining area greening monitoring method and system based on deep learning
Huan et al. MAENet: multiple attention encoder–decoder network for farmland segmentation of remote sensing images
Li et al. Long-term spatio-temporal changes of wetlands in Tibetan Plateau and their response to climate change
CN112330562B (en) Heterogeneous remote sensing image transformation method and system
Moradi et al. Potential evaluation of visible-thermal UAV image fusion for individual tree detection based on convolutional neural network
CN117058367A (en) Semantic segmentation method and device for high-resolution remote sensing image building

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