A cloud platform and deep learning-based shore line remote sensing automatic identification analysis method

Abstract

The present application provides a kind of cloud platform and deep learning-based shoreline remote sensing automatic identification analysis method, comprising: based on cloud platform, the block downloading and pre-processing of large-area multi-source remote sensing image are realized;Cross extensive geographical area, involve different shoreline type scene of shoreline sample data set development under high water level and low water level condition;Feature importance screening based on average precision reduction analysis method of random forest;Deep learning network training and precision evaluation of image feature and shoreline sample data set are applied;Network model adaptive sea-land segmentation and shoreline binary graph generation based on Canny edge detection algorithm and actual shoreline vector conversion;Long time series shoreline result extraction and tidal correction post-processing.The present application makes the extraction error of shoreline less than the maximum allowable error, and can meet the long time series large range complex shoreline automatic detection demand under multiple scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of coastline remote sensing recognition, and particularly relates to a coastline remote sensing automatic recognition and analysis method based on a cloud platform and deep learning. BACKGROUND

[0002] The coastal zone is located in the land-sea interaction zone, and is not only affected by global environmental problems such as climate change, ocean acidification and land-sea habitat changes, but also seriously affected by human activities such as reclamation artificial beach construction, near-shore pond aquaculture and port city construction. As the boundary between land and sea, the coastline is the most intuitive coastal element reflecting the current situation of the coastal zone ecological environment, and the position, length and type changes of the coastline are closely related to the problems of climate change and human activities affecting the coastal zone. As a non-contact remote sensing technology, especially in recent years with the increasing of medium and high resolution remote sensing satellites, remote sensing has gradually become an important data source for coastline monitoring. The coastline change monitoring based on long-term remote sensing data has become an important means and key indicator for the environmental monitoring and evaluation of the coastal zone.

[0003] The initial coastline remote sensing extraction is completed by field survey or visual interpretation with unified interpretation standards, which is also the method with the highest coastline extraction accuracy, but is severely limited by efficiency, cost and subjectivity. With the development of remote sensing computer interpretation technology, a large number of automatic coastline extraction methods have emerged. Zhang et al. proposed a combination method integrating image segmentation, region growing and multispectral image edge detection to improve the coastline extraction accuracy. Chen Weitong et al. integrated multi-source remote sensing data and measured tidal level and slope data, and combined with the improved water edge line method to extract multi-temporal coastline data. Dai et al. analyzed the normalized difference water index coastline extraction method by statistical analysis of repeated measurement superposition to reduce image offset and cloud shadow image error. In summary, the semi-automatic extraction method based on optical remote sensing image water edge line extraction and post-correction, the extraction method based on LiDAR data to solve the tidal level surface, and the intelligent extraction method based on remote sensing image and interpretation standard are the most widely used coastline remote sensing extraction methods at present. At the same time, the development of remote sensing intelligent cloud platform also brings new possibilities and low-cost solutions for coastline extraction research.

[0004] Deep learning is widely used in various fields due to its ability to extract image features and fit complex problems. In the field of coastline remote sensing extraction, a large number of deep learning algorithms have emerged. Wang Zhenhua et al. applied an improved Deeplab network to propose a sea island coastline remote sensing image segmentation model. The results show that this method overcomes the problem of discontinuous sea island coastline caused by other models and reduces the misclassification of sea island coastline. Seale et al. proposed a Sentinel-2 water edge dataset for training and testing automatically extracted coastlines from Sentinel-2 images and trained, tested and optimized four convolutional neural network models based on the U-Net model architecture to obtain a coastline type and element dataset from all over the world. The current coastline extraction algorithm based on deep learning remote sensing image segmentation algorithm is limited by the following reasons: (1) The deep learning-based image segmentation algorithm relies heavily on long-term training, large amounts of data sets and manual annotation, which poses a challenge to the automatic acquisition of remote sensing data; (2) Coastlines in different scenarios have different characteristics, and the characteristics of multiple scenarios and multiple bands pose a challenge to the input of deep learning models. SUMMARY

[0005] To solve the above technical problems, the present application proposes a coastline remote sensing automatic identification analysis method based on cloud platform and deep learning, which can meet the demand of long-term and large-scale complex coastline automatic detection in multiple scenarios with an extraction error less than the theoretical maximum allowable error.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A coastline remote sensing automatic identification analysis method based on cloud platform and deep learning, comprising the following steps:

[0008] Step S1, downloading remote sensing images in large areas based on cloud platform;

[0009] Step S2, preprocessing remote sensing images in a specified area, which includes cloud processing, panchromatic sharpening and downsampling;

[0010] Step S3, developing a coastline sample dataset for coastline detection method training based on Landsat and Sentinel series data, which spans a wide geographical area and includes various coastline type scenarios under high and low water level conditions;

[0011] Step S4, analyzing the importance of input features based on the average precision reduction analysis method of random forest to determine the most suitable input features for coastline extraction;

[0012] Step S5, constructing and setting parameters of the selected U-net, DeepLabv3+, Deep Res U-net, Res U-net, R2 U-net, Attention U-net, Res U-net++, SAnet and U-net 3+ neural network models;

[0013] Step S6, training the neural network model in step S5 by taking the training data set and the verification data set in different scenes as input, to obtain the neural network model with the optimal weight value in each scene;

[0014] Step S7, performing sea-land segmentation on the test data by using the neural network model with the optimal weight value in different scenes, generating a coastline binary graph by using a Canny edge detection algorithm, and then converting the coastline binary graph into an actual coastline vector, and finally comparing the accuracy of the reference coastline vector obtained by visual interpretation with the two indexes of completeness and correctness, to obtain the neural network model with the highest accuracy in different scenes;

[0015] Step S8, large-area coastline result extraction and tide correction post-processing, taking the downloaded large-area remote sensing image as input, classifying the image and marking it as different scenes, and then adaptively selecting the neural network model with the highest classification accuracy in different scenes, performing long-time coastline extraction, and performing tide correction on the coastline extraction result.

[0016] Further, in the step S1, the image is obtained, downloaded and cropped through a GEE API, and the blocking processing required due to memory limitation is realized by establishing a 10km×10km grid.

[0017] Further, in the step S3, the construction method of the coastline sample data set comprises:

[0018] The sample area involves different types of coastline distribution areas in each continent in the world except Antarctica, including artificial coastline, sandy coastline, biomass coastline, bedrock coastline, estuary coastline and silt coastline.

[0019] 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:

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] Wherein, B2 is visible blue light band, B3 is visible green light band, B4 is visible red light band, B8 is near-infrared band, B11 is short-wave infrared band.

[0026] Further, in the step S7, the extraction coastline is obtained from the actual coastline, and buffer zones are made for the reference coastline and the extraction coastline respectively, in the buffer zone with the reference coastline as the center, the part of the extraction coastline in the zone is recorded as TP1, and the part outside the zone is recorded as FP; in the buffer zone with the extraction coastline as the center, the part of the reference coastline in the zone is recorded as TP2, and the part outside the zone is recorded as FN, and the calculation formula of the completeness Complete and the correctness Correct is as follows:

[0027]

[0028]

[0029] Wherein, the completeness Complete describes the percentage of correct coastlines in the extraction coastline, and evaluates the completeness of the extraction result of the coastline; the correctness Correct describes the proportion of the correct extraction in the reference coastline.

[0030] Beneficial effects:

[0031] The application applies cloud platform and deep learning algorithm to semantic segmentation of remote sensing images, and applies to coastline extraction of remote sensing images, so that more manpower and material resources are saved. In view of the problem that a large number of data sets and artificial annotation are lacked in automatic extraction of a large area, automatic acquisition of long-time sequence remote sensing images of a large area is realized based on GEE API, and a coastline sample data set spanning a wide geographical area and involving different coastline types under high water level and low water level conditions is constructed. In view of the characteristics of multiple scenes and multiple bands of the coastline, an adaptive neural network model under different scenes is trained, so that the model is fully adapted to different scenes and the precision of coastline extraction under different scenes is improved. Through the research, not only the coastline can be quickly monitored, and information such as change dynamic is mastered, but also the degree of coastal erosion and accumulation can be monitored, and dynamic tracking and monitoring of the coastal environment are realized. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the drawings needed to be used in the specific embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0033] Figure 1 A flow chart of a shore line remote sensing automatic recognition and analysis method based on a cloud platform and deep learning according to an embodiment of the present application;

[0034] Figure 2 A complete and correct accuracy verification schematic diagram of shore line extraction result according to an embodiment of the present application; wherein a is a matched extraction result map, and b is a matched reference shore line map.

[0035] Figure 3 A U-net original network structure schematic diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] 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. Also, parts irrelevant to the description of the exemplary embodiments are omitted in the drawings for the sake of clarity.

[0037] In the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate that there is presence of features, numbers, steps, actions, parts, or combinations thereof disclosed in the specification, and do not exclude the presence or addition of one or more other features, numbers, steps, actions, parts, or combinations thereof.

[0038] It should also be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the drawings and in combination with the embodiments.

[0039] This invention discloses an automatic shoreline remote sensing identification and analysis method based on a cloud platform and deep learning. This method utilizes the global-scale remote sensing cloud computing platform GEE, leveraging Google's powerful cloud computing capabilities and multi-source long-term remote sensing image data. It can automatically acquire, filter, and crop remote sensing images for selected areas and integrates remote sensing image preprocessing methods such as cloud processing, pan-color sharpening, and downsampling. Based on this, semantic-level land-sea segmentation samples for different scenarios are constructed, and nine neural network models, including U-net and DeepLabv3+, are trained. The optimal neural network model is obtained by combining the extraction accuracy across different scenarios, enabling the identification of complex shorelines. After post-processing operations such as tidal correction, realistic shoreline data is obtained.

[0040] This invention enables automated shoreline extraction in various scenarios. The block-based processing technology ensures processing speed under limited memory conditions, saving system operating costs. By training neural network models for different scenarios, the optimal model is obtained for extracting complex shorelines, significantly improving computational accuracy. Therefore, this invention can guarantee high speed and accuracy in the automated, long-term, large-scale extraction of complex shorelines.

[0041] like Figure 1 As shown, an embodiment of the present invention provides a method for automatic shoreline remote sensing identification and analysis based on cloud platform and deep learning, which includes the following steps:

[0042] Step 1: Download large-area remote sensing imagery in chunks using a cloud platform, including:

[0043] Automatic acquisition of multi-source satellite imagery: Integrating the Google Earth Engine (GEE) API, it provides top-of-atmosphere (TOA) reflectance images from Landsat 5TM, Landsat 7 ETM+, Landsat 8 OLI Tier 1 collections, and Sentinel-2 MSI level-1C products. Images downloaded from the GEE server within a specified time range can be cropped by defining regions of interest. Furthermore, the downloaded images only contain spectral bands optimized for coastline detection features: three visible bands (R, G, B), near-infrared (NIR), and two shortwave infrared (SWIR) bands, significantly improving processing speed and reducing memory usage on the local computer.

[0044] Specifically, GEE (Google Earth Engine) is a tool provided by Google that can process satellite image data in batches, and it provides an offline Python API and an online JavaScript API. First, the Python API library of GEE is configured locally, and the Google Python API client, authentication dependency verification library, and GEE Python library are installed. The local integration of GEE is completed by importing the geemap package, and the automatic download of images is completed by defining the area range, date range, and satellite type to be downloaded. At the same time, these images will also be cropped according to the defined area range. Due to the limitation of memory, the download range larger than 100 square kilometers will be processed in blocks, and this is achieved by establishing a 10 km x 10 km grid on the study area.

[0045] Step two, preprocessing of multi-source satellite image data: cloud processing, panchromatic sharpening, and downsampling are performed before shoreline extraction.

[0046] The cloud processing includes: the quality assessment band pre-calculated by the data provider (Landsat for USGS, Sentinel-2 for ESA) to supplement the cloud coverage area, which includes a cloud mask containing all cloud area pixels. On this basis, the cloud coverage is calculated according to the number of cloud pixels in the defined area of interest, and the cloud area image exceeding a certain coverage is discarded based on the defined cloud coverage threshold.

[0047] Specifically, the Landsat and Sentinel series data obtained by GEE will have a QA (data product quality assessment) band to evaluate the quality of each pixel and extract the required pixels. It is mainly used to extract cloud pixel information and shielding. The value stored in the QA band is generally displayed in decimal, which needs to be converted to binary value to determine the pixel state in the area. After judging the input remote sensing image, the band characteristics of the input image are judged, and the cloud band is selected. The create_cloud_mask function is used to generate a cloud mask, and the other bands of the image are also checked. The null value area is extracted and incorporated into the cloud mask. Finally, the corresponding pixel position is removed through the cloud mask.

[0048] The panchromatic sharpening and downsampling includes: enhancing the spatial resolution of the panchromatic band of the image to achieve the best effect of coastline monitoring, for Landsat7 and Landsat8 images, by applying the data fusion method of principal component analysis, using the higher resolution panchromatic band to increase the spatial resolution of the multispectral band from 30m to 15m, downsampling the multispectral band to 15m by bilinear interpolation and decomposing into principal components, then replacing the first principal component with the panchromatic band after histogram matching, and transforming back to the original multispectral space. For Landsat5 images, which do not have a panchromatic band, directly downsample 30m to 15m by bilinear interpolation to improve the accuracy of coastline detection. For Sentinel-2 images, downsample the 20m short-wave infrared band to 10m by bilinear interpolation, and unify the spatial resolution of all bands to 10m.

[0049] Step three, training and validation dataset annotation: based on Landsat and Sentinel series data, a coastline sample dataset for training coastline detection method is developed, which covers a wide geographical area and involves various coastline type scenes under high and low water level conditions.

[0050] The image data is downloaded through the above-mentioned cloud platform image automatic acquisition technology, the selected images basically meet the characteristics of clearness, no cloud, etc., the sample area involves different types of coastline distribution areas in all continents except Antarctica, including 16 training sites and 49 test sites, as well as artificial coastline, sandy coastline, biomass coastline, bedrock coastline, estuary coastline and silt coastline. Six kinds of coastline type scenes. The selected images are labeled, creating dense pixel-level labels of two classes of "water" and "non-water", the specific process is as follows:

[0051] Create labels through semi-supervised clustering method, first perform false color synthesis on the image to improve the contrast between water and non-water pixels; secondly, apply the K-means clustering method to the image after false color rendering, optimize the number of clustering categories k to obtain the best discrimination effect, then merge the categories until the last two categories of water and non-water; finally, compare with the high-definition images provided by Google Earth, manually correct the misclassified pixels.

[0052] At this point, the dataset is complete, and the dataset is randomly divided into training set and validation set in the ratio of 7:3. Before using the training set image as input to train the algorithm, it is first normalized, and the training dataset is data augmented, that is, new images are generated by adjusting the image color and performing rotational symmetry operation on the image.

[0053] Step four, analyze the importance of input features based on the average precision reduction analysis method of random forest, and determine the most suitable input features for coastline extraction.

[0054] Based on the Mean decrease accuracy analysis method of random forest, 5 multispectral bands (R, G, B, NIR and SWIR1), 5 commonly used spectral indices (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 the variance of each spectral band and index were selected as the input features of the subsequent model.

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Wherein, B2 is visible blue band, B3 is visible green band, B4 is visible red band, B8 is near-infrared band, B11 is short-wave infrared band.

[0061] Step five, neural network model construction and parameter setting: nine kinds of neural network models were selected for coastline detection in different scenes. In addition to U-net and DeepLabv3+, the structures of other models (Deep Res U-net, Res U-net, R2 U-net, Attention U-net, Res U-net++, SAnet and U-net 3+) are basically based on U-net model as the main body, and are optimized by adding different modules and combinations. The modules applied by these models are: residual (Residual), recurrent (Recurrent), attention gate (Attention Gate), squeeze excitation (Squeeze Excitation) and dense connection (Dense Connect) structures, as shown in Table 1.

[0062] Table 1

[0063]

[0064] Wherein: the first row name is the publication time of each model; "√" represents that the model has this module.

[0065] As Figure 3As shown, the U-Net network consists of two parts of down-sampling and up-sampling, shaped like a 'U', based on the Encoder-Decoder structure, through splicing to realize feature fusion, the structure is simple and stable. The first part extracts features from the image through convolutional layers and max-pooling layers, and each 3x3 convolutional layer is followed by an activation function ReLU and a 2x2 max-pooling operation. The second part first performs deconvolution operation, then splices the result with the corresponding feature map, and finally uses a 1x1 convolution kernel in the output layer. The main body of the Encoder of DeepLabv3+ is a DCNN with atrous convolution, followed by an Atrous Spatial Pyramid Pooling (ASPP) module with atrous convolution to introduce multi-scale information. Compared with other DeepLab networks, it introduces a Decoder module to fuse low-level features with high-level features, further improving the accuracy of segmentation boundaries. The Residual structure solves a series of problems such as difficulty in training, easy overfitting, and gradient vanishing or explosion when the network depth is constantly increasing. It splits the regular convolution module into "direct mapping" and "residual short connection" two parts, and adds them at the output, effectively making up for the information lost during convolution, and the gradient also participates in the residual part during back propagation, ensuring that the gradient range is stable within a controllable range. The Recurrent Residual Convolutional structure is a combination of recurrent convolution and residual structure, called the R2U-net. Recurrent convolution is set at the end of each layer in the encoding part of the U-net, with a default setting of two recursions, increasing the number of layers in each convolution, allowing more contextual information and more deep features of the image to participate in learning. The introduction of the residual structure solves the problem of gradient vanishing during back propagation and loss of original data information when the depth of the recurrent structure network increases. The Attention Gate (AG) structure introduces attention mechanisms into the U-net, called the Attention U-net. The Attention Gate module is a soft attention in the spatial domain, which is set in the skip connection of the decoding part of the U-net, fuses the feature map after up-sampling with the feature map after down-sampling of the previous layer, and learns the attention coefficient of the fused features through "point-wise convolution", and uses it to amplify the information in the key area (high attention coefficient). The Dense Connect structure separates a part of the regular convolution from the "dense connection" with other feature layers. Unlike the residual structure, the feature layers are "densely connected" in the channel layer, and from the entire network, almost all layers are connected, which can maximize the use of network information. Highly effective information utilization reduces network gradient vanishing and reduces redundant parameter quantity.

[0066] Step six, training of the neural network model: training the neural network model with the training data set and the validation data set in different scenarios as input, obtaining the neural network model with the optimal weight value in each scenario.

[0067] The experimental environment is a graphics workstation running "Windows 10 operating system 64-bit", and the specific hardware environment parameters are: the core processor (CPU) is W-2223@3.60GHz, the image processor (GPU) is NVIDIA GeForce RTX 3090 24GB, and the memory (RAM) is 64GB; the software environment is: deep learning framework PyTorch (version GPU1.7.0) and CUDA (version 11.1), and the expansion library GDAL (version 2.3.3) and OpenCV (version 1.4) for data enhancement.

[0068] The experiment will train 9 neural networks in a unified environment to compare the extraction accuracy under different network structures. The specific strategy is as follows:

[0069] (1) The "Kaiming" initialization method is used to initialize the weights of each layer of the network (especially the activation function layer), so as to improve the loss convergence speed of each network and prevent the network from appearing gradient disappearance and other problems;

[0070] (2) The training is adjusted by the "Adam" learning rate optimizer to automatically adjust the training learning rate, so as to speed up the network gradient descent and avoid the network loss stagnating at a local minimum;

[0071] (3) The initial learning rate of the model is set to 0.001, and the network is verified on the validation set once every training round, while the verification accuracy (verified by the average intersection over union mIoU and the average pixel accuracy mPA) is recorded and monitored.

[0072] Mean Intersection over Union (mIoU), that is, the ratio of the intersection and union of the prediction result and the true result in space for each class (abbreviated as IoU), and then the arithmetic mean of all classes is calculated. It can reflect the accuracy of the semantic segmentation result in space, the higher the intersection over union, the better the segmentation effect, and the calculation formula is as follows:

[0073]

[0074] Where TP is the intersection of the true label and the intersection result, FN is the true label, FP is the prediction result, and k is the number of classes.

[0075] Mean Pixel Accuracy (mPA), i.e. calculating the proportion of the number of pixels correctly segmented in each class in the prediction result to the total number of pixels in the class, and taking the arithmetic mean of the pixel accuracy of each class. This index reflects the ability of the algorithm to segment by the proportion of pixels correctly segmented in the result. The higher the accuracy, the better the segmentation effect. The calculation formula is as follows:

[0076]

[0077] Wherein, pii represents the number of pixels that belong to class i but are predicted as class j, i is the correct class, j is the predicted class, and k is the number of classes.

[0078] mPA is mainly used to evaluate the effect of the model on the land type determination and the position of each pixel in the validation set, while mIoU is used to evaluate the effect of the model on the land area and boundary segmentation in the validation set. The change trend of the two is basically the same. The combination of the two can fully represent the accuracy of the model. If the verification accuracy does not improve significantly within two rounds, reduce the learning rate by 1% and accelerate the network loss convergence to quickly jump out of the "local minimum";

[0079] (4) Set the initial training round to 20. For the model that has not completely converged, continue training through the "warm-up" method until it approaches the fitting state.

[0080] (5) Select the round with the best validation set accuracy as the final training result of the network.

[0081] Step seven, perform sea-land segmentation on the test data using the neural network model with the optimal weight value in different scenarios, generate a coastline binary graph using the Canny edge detection algorithm, and then convert it into a real coastline vector. Finally, compare the accuracy of the segmented result with the reference coastline vector obtained by visual interpretation in terms of completeness and correctness, and obtain the neural network model with the highest accuracy in different scenarios.

[0082] The line target matching method is used to calculate the completeness and correctness of the extracted coastline based on the real coastline reference data. The specific method is as follows:

[0083] First, buffer the reference coastline and the extracted coastline. In the buffer zone of the reference coastline, mark the part inside the extracted coastline as TP1 and the part outside as FP. In the buffer zone centered on the extracted coastline, mark the part inside the reference coastline as TP2 and the part outside as FN. The calculation formulas of the completeness and correctness for result accuracy evaluation are as follows:

[0084]

[0085]

[0086] wherein the completeness is used to describe the percentage of the extracted coastline that is correctly divided, and the completeness of the evaluation result; the correctness describes the proportion of the reference coastline that is correctly extracted in the extracted coastline. For the general coastline extraction result without real coastline reference, the verification points are obtained from the high-definition images of Google Earth to calculate the mean offset and the root mean square error (RMSE), and the formulas are as follows:

[0087]

[0088]

[0089] wherein n is the number of verification points, D n is the Euclidean distance from the verification point to the coastline result. The mean offset is used to describe the offset degree of the coastline result from the verification point, and the root mean square error is the standard deviation of the coastline result and the offset of the verification point. The relationship between the maximum allowed error U of the coastline extraction based on the remote sensing image and the image resolution r is as follows:

[0090]

[0091] When the mean offset and the root mean square error of the extracted coastline are less than the maximum allowed error calculated by the image resolution, it is considered that the accuracy of the extracted coastline meets the standard; and the result is shown in Figure 2 .

[0092] Step eight, step S8, large area coastline result extraction and tide correction post-processing, taking the downloaded large area remote sensing image as input, classifying and marking the image as different scenes, then adaptively selecting the neural network model with the highest classification accuracy in different scenes, performing long-time coastline extraction, and performing tide correction on the coastline extraction result.

[0093] In order to ensure that the extracted coastline result is closer to the real coastline, it is necessary to perform tide correction on the coastline extraction result, and the correction method is as follows:

[0094] First, two scenes of coastline at different imaging times are extracted, C1 and C2, the horizontal distance between C1 and C2 is ΔL, θ is the slope of the coast, h1 and h2 are the tide heights when the two scenes of satellite pass, respectively (h2>h1), H is the high tide height of the average spring tide, L is the correction distance of the coastline, and the slope of the beach can be calculated by the following formula:

[0095] θ=arctan[(h2-h1) / ΔL]

[0096] The distance from the coastline C1 to the actual coastline is:

[0097] L = (H - h2) / tan θ

[0098] According to the above coastline tide correction formula, the extracted coastline C1 is moved to the land direction by a distance L to obtain the actual coastline, wherein the artificial, bedrock and estuary coastline is not sensitive to the influence of the tide and can be ignored. The biomass, silt and sandy coastline algorithm takes the inside as the coastline, and combines the tide correction formula to avoid the influence of the tide to a certain extent.

Claims

1. A cloud platform and deep learning based automatic identification and analysis method for shoreline remote sensing, characterized in that, Comprise the following steps: Step S1, based on cloud platform for large area remote sensing image block download; Step S2, the remote sensing image of the specified area is pretreated, and the pretreatment includes cloud processing, panchromatic sharpening and downsampling; Step S3, based on Landsat and Sentinel series data, a coastline sample data set for training the coastline detection method is developed, the coastline sample data set spans a wide geographical area, including various coastline type scenes under high water level and low water level conditions; Step S4, the importance of input features is analyzed based on the average precision reduction analysis method of random forest, and the input features most suitable for coastline extraction are determined; Step S5, the selected U-net, DeepLabv3+, Deep Res U-net, Res U-net, R2 U-net, Attention U-net, Res U-net++, SAnet and U-net 3+ neural network models are constructed and parameterized; Step S6, the training data set and the verification data set under different scenes are taken as input to train the neural network model in step S5, and the neural network model under the optimal weight of each scene is obtained; Step S7, the neural network model under the optimal weight of different scenes is used to perform sea-land segmentation on the test data, and the segmented results are used to generate coastline binary graph by Canny edge detection algorithm, and then converted into actual coastline vector, finally compared with the reference coastline vector interpreted visually in complete and correct indexes, and the neural network model with the highest accuracy under different scenes is obtained; Step S8, large area coastline result extraction and tide correction post-processing, taking the downloaded large area remote sensing image as input, classifying the image and marking it as different scenes, then adaptively selecting the neural network model with the highest classification accuracy under different scenes, performing long time sequence coastline extraction, and tide correction on the coastline extraction result; Firstly, the shorelines of the two images are extracted as C1 and C2, respectively, and the horizontal distance between C1 and C2 is , is the slope of the coast, h1 and h2 are the heights of the tide at the time of satellite overpass, h2>h1, H is the height of the high tide of the average spring tide, and L is the corrected distance of the shoreline. The slope of the beach is calculated by the following formula: ； The distance from the coastline C1 to the actual coastline is: ； According to the above formula, the extracted coastline C1 is moved to the land direction by a distance L to obtain the actual coastline. 2.The shore line remote sensing automatic identification and analysis method based on cloud platform and deep learning according to claim 1, characterized in that, In the step S1, the image is obtained, downloaded and cropped through GEE API, and the blocking processing required due to memory limitation is realized by establishing a 10km*10km grid. 3.The shore line remote sensing automatic identification and analysis method based on cloud platform and deep learning according to claim 2, characterized in that, In the step S3, the construction method of the coastline sample data set comprises: The sample area involves different types of coastline distribution areas in each continent in the world except Antarctica, including artificial coastline, sandy coastline, biomass coastline, bedrock coastline, estuary coastline and silt coastline.

4. The cloud platform and deep learning-based automatic shore line remote sensing identification analysis method according to claim 3, characterized in that, In the step S4, the index features in the input features include 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, respectively: ； ； ； ； ； Wherein, B2 is visible blue band, B3 is visible green band, B4 is visible red band, B8 is near infrared band, B11 is short wave infrared band.

5. The cloud platform and deep learning-based automatic shore line remote sensing identification analysis method according to claim 4, characterized in that, In the step S7, the extracted shoreline is obtained from the actual shoreline, and a buffer zone is made for the reference shoreline and the extracted shoreline. In the buffer zone with the reference shoreline as the center, the part within the zone of the extracted shoreline is recorded as TP1, and the part outside the zone is recorded as FP. In the buffer zone with the extracted shoreline as the center, the part within the zone of the reference shoreline is recorded as TP2, and the part outside the zone is recorded as FN. The completeness and the correctness are calculated according to the following formulas: ； ； where completeness Described is the percentage of the correct coastline in the extracted coastline, evaluating the completeness of the coastline extraction result; correctness Described is the proportion of the correctly extracted in the reference coastline.

Images