CN115620132A - Unsupervised comparative learning ice lake extraction method - Google Patents
Unsupervised comparative learning ice lake extraction method Download PDFInfo
- Publication number
- CN115620132A CN115620132A CN202211216115.5A CN202211216115A CN115620132A CN 115620132 A CN115620132 A CN 115620132A CN 202211216115 A CN202211216115 A CN 202211216115A CN 115620132 A CN115620132 A CN 115620132A
- Authority
- CN
- China
- Prior art keywords
- ice lake
- ice
- transformation
- image
- lake
- 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
Links
Images
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
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- 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
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- 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
- G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an unsupervised comparative learning ice lake extraction method, which mainly solves the technical problems that in the existing ice lake extraction method, training sample labels are complex to manufacture, time and labor are wasted, and models are difficult to directly transfer to other data. The method is based on a convolutional neural network, original images of the ice lake remote sensing images are subjected to transformation processing to obtain transformation images, and the original images and the transformation images form sample pairs containing two branches; then respectively carrying out downsampling processing and mapping processing on the sample pairs to further obtain a comparison learning module of the ice lake; meanwhile, a water body index NDWI spectral feature map is used as a pseudo label for comparison learning, a position learning module of the ice lake is obtained by calculating position loss, an ice lake extraction model is finally obtained, and ice lake information can be automatically extracted by inputting any ice lake remote sensing image into the model. The method is more convenient in extraction process, time-saving and labor-saving, and the extraction efficiency of the ice lake is greatly improved.
Description
Technical Field
The invention belongs to the field of remote sensing image processing and freezing circles, and particularly relates to an unsupervised comparative learning ice lake extraction method.
Background
The iced lake which uses glacier melt water as a main material source is frequently raised in a plateau area. In recent years, as global climate becomes warm, the state of the ice lake changes. When the area of the ice lake is gradually enlarged and the moraine dam cannot bear the pressure of the ice lake, the state of the ice lake is changed, and the ice lake is broken. Therefore, the method has important significance for accurately monitoring the state of the ice lake, preventing ice lake burst disasters, reducing property loss in downstream areas and the like. With the development of remote sensing technology in recent decades, the remote sensing images record abundant ground feature information, the types of the remote sensing images become more diverse, and the spatial resolution and the temporal resolution are gradually improved, so that the large-scale continuous monitoring of the ice lake is gradually possible, and a method for quickly and accurately extracting the ice lake information is urgently needed.
With the development of machine learning and deep learning in the field of computer vision, the target detection method also makes great progress in ice lake extraction. For example, the method for extracting the ice lake based on the random forest (Wangchuk S.; bolch T.mapping of cementitious lakes using Sentinel-1and Sentinel-2data and a random forest flavor. For another example, a Deep Learning segmentation model U-net is applied to the extraction of ice lake, and the results are obtained by referring to (Qayyum N.; ghuffar S.; ahmad H.M.; yousaf A.; shahid I.Glacal Lakes Mapping Using Multi Satellite plan image and Deep Learning, ISPRS International Journal of Geo-Information,2020,9) and (Wu R.; liu G.; zhuang R.; wang X.; li Y.; zhuang B.; cai J.; xing 5262 zxft Left Learning Method for Mapping Global Lakes front composite of Synthetic-lake radial and 3763), and the results are also better. As another example, a generation-resistant network (ZHao H.; zhang M.; chen F. GAN-GL: genetic adaptive Networks for Global Lake mapping. Remote Sensing,2021,13) is combined in a generating manner to obtain a region of the ice Lake. However, existing ice lake Extraction methods, such as NDWI (Li, j.; sheng, Y.an automated scheme for a cementitious laboratory Mapping Using Landsat Image and Digital Elevation Models A Case Study in the Himalayas.int.J.Remote Sens.2012,33, 5194-5213), C-V model (Zhao H.; chen F.; zhang M.A Systematic Extraction application for Mapping Glacial Lakes in High Mountain areas of Asia. IEEE Journal of Selected Topics in Applied Earth requirements and remotes Sensing,2018,11,2788-2799), object-oriented methods (Mitkari K.V.; arora M.K.; tiwari R.K. Extraction of Glacial Lakes in gap ceramics Using Object-Based analysis. IEEE Journal of Selected Topics in Applied Earth objectives and remotes Sensing,2017,10,5275-5283), all of which affect the cloud, such as removing the interference factors, such as ice shadows, and the like. Meanwhile, the methods are all methods based on data learning, the test results of models of the methods mostly depend on the completeness of training data, but a great deal of effort is usually required to be invested in obtaining the training data with labels, and if once new remote sensing data is encountered, training samples must be made again, so that the applicability of the models in different data is greatly limited by the methods.
Disclosure of Invention
The invention provides an unsupervised comparative learning ice lake extraction method, which mainly solves the technical problems that in the existing ice lake extraction method, training sample labels are complex to manufacture, time and labor are wasted, and models are difficult to directly transfer to other data.
In order to solve the technical problems, the technical scheme provided by the invention is as follows:
an unsupervised comparative learning ice lake extraction method is characterized by comprising the following steps:
step 1, transforming an original image of the ice lake remote sensing image to obtain a transformation image, and combining the original image and the transformation image into a sample pair comprising two branches;
step 2, performing down-sampling processing on the sample in a network with shared input weights, and respectively extracting corresponding feature maps of two branches under different scales;
step 3, respectively inputting the corresponding feature maps of the two branches obtained in the step 2 under different scales into a projection layer for mapping to obtain feature vectors mapped by the two branches under corresponding scales, and measuring the similarity degree of the two branches by adopting similar loss based on the feature vectors, wherein the similar loss is calculated by adopting cosine similarity, so as to obtain a comparison learning module of the ice lake;
step 4, calculating a water body index NDWI spectral feature map of the original image, setting a water body index threshold value as T, further obtaining a rough ice lake distribution area map of the original image, and using the rough result map as a pseudo label learned by the comparison learning module obtained in the step 3 for guiding the recognition of the ice lake features;
step 5, performing up-sampling processing on the corresponding ice lake characteristic graphs of the original image in the two branches obtained in the step 2 under different scales in a network with shared weight, and outputting an ice lake extraction result corresponding to the original image;
step 6, calculating the position loss by combining the pseudo label obtained in the step 4 and the ice lake extraction result obtained in the step 5 to further obtain an ice lake position learning module, and combining the position learning module with the comparison learning module obtained in the step 3 to obtain an ice lake extraction model;
and 7, inputting any ice lake remote sensing image into the ice lake extraction model obtained in the step 6, and outputting an ice lake extraction result.
Further, in step 1, the transformation process is performed by any one of color mapping transformation, inversion transformation, gradation transformation, and blurring transformation, and noise is randomly added while processing the image size to 448 × 448.
Further, in step 2, the scales include 4 scales, which are 224 × 224, 112 × 112, 56 × 56, and 28 × 28, respectively, where one scale is a downsampling block, each downsampling block includes two convolutional layers and one downsampling layer, each convolutional layer is activated using a ReLU, and each downsampling layer uses a convolution operation with a step size of 2; the ice lake characteristic diagram of the original image and the ice lake characteristic diagram of the transformation image are respectively the result of activation of the last convolution layer before each downsampling layer under the corresponding scale.
Further, in step 3, the projection layer includes three fully-connected layers, the mapping is to activate the feature maps entering the two fully-connected layers by using ReLU, so as to obtain a mapping result at the third fully-connected layer, the mapping result is a feature map vector corresponding to different scales, the vector is used to calculate a similarity loss of the mapping result, the similarity loss is calculated by using cosine similarity, and the calculation formula is as follows:
wherein q represents the ice lake characteristic diagram of the original image, q' represents the ice lake characteristic diagram of the transformation diagram, i represents different scales, i is a positive integer, and i is more than or equal to 1and less than or equal to 4; p (q) i ) Representing the feature vector after the ice lake feature map of the original image is mapped; p (q' i ) Representing the feature vector after the ice lake feature map of the transformation map is mapped; | | p (q) i )|| 2 Representing the L2 norm of the feature vector after the ice lake feature map of the original image is mapped; l | p (q' i )|| 2 Map of ice lake features representing transformation mapL2 norm of the fired eigenvector.
Further, in step 4, the calculation formula of the water body index NDWI spectral feature map is as follows:
where ρ is Green The apparent reflectivity of the atmosphere top layer of the green light wave band is expressed, and is more than 0 and less than rho Green ＜1，ρ NIR The apparent reflectivity of the top atmosphere layer in the near infrared band is expressed, and is more than 0 and less than rho NIR Is less than 1; the threshold value T of the water body index is 0.7.
Further, step 5 specifically comprises: and (3) respectively taking the corresponding ice lake characteristic diagrams of the original image in the two branches obtained in the step (2) at different scales as up-sampling blocks, wherein each up-sampling block comprises an up-sampling layer and two convolution layers, the up-sampling layer uses deconvolution operation, each convolution layer is activated by using a ReLU function, and the final reduction result is the ice lake extraction result of the original image.
Further, in step 6, the position loss is calculated by using an L2 norm, where the calculation formula of the L2 norm is:
wherein I represents an input image; b represents an ice lake mask obtained after threshold according to the water body index NDWI; f (-) represents an ice lake extraction model; u represents an arbitrary position in the image; f. of u (I) An extraction result representing any position in the input image; b is u Representing the result of the masking at any location of the image.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention relates to an unsupervised contrast learning ice lake extraction method, which is based on a convolutional neural network, adopts an unsupervised training mode, obtains a transformation graph by transforming an original image of an ice lake remote sensing image, and combines the original image and the transformation graph into a sample pair containing two branches; then respectively carrying out downsampling processing and mapping processing on the sample pairs to further obtain a comparison learning module of the ice lake; meanwhile, a water body index NDWI spectral feature map is used as a pseudo label for comparison learning, a position learning module of the ice lake is obtained by calculating position loss, an ice lake extraction model is finally obtained, and ice lake information can be automatically extracted by inputting any ice lake remote sensing image into the model. According to the method, a label of training data is not required to be made, the preparation work of the model training process is greatly simplified, the extraction process of the ice lake is more convenient, time and labor are saved, and the extraction efficiency of the ice lake is greatly improved.
2. In order to ensure that all noise interference is eliminated, the water body index threshold is set to be 0.7, and if the NDWI pixel is larger than the threshold, the NDWI pixel is marked as a water body, so that the accuracy of the model for extracting the ice lake information is improved.
3. According to the unsupervised comparison learning ice lake extraction method, the pseudo label is used as the label for comparison learning, the obtained ice lake model has high applicability, and any image is input into the trained model to obtain a relatively accurate ice lake extraction result, so that the automatic extraction of the ice lake is realized.
Drawings
FIG. 1 is a schematic diagram of an extraction process of an embodiment of an unsupervised comparative learning ice lake extraction method of the present invention.
Detailed Description
Contrast learning is a deep learning model newly emerging in recent two years, which transforms input images and inputs the transformed images into a network shared by two weights, so as to discover targets with similar characteristics in the images. Based on the principle, the invention provides an ice lake extraction method without unsupervised comparative learning of a large number of training labels, avoids using auxiliary data or a large number of preprocessing and post-processing works, obtains a better ice lake extraction effect and realizes the automatic extraction of the ice lake.
The method for extracting the ice lake through unsupervised comparative learning is explained in detail below with reference to the attached drawings.
As shown in fig. 1, the unsupervised comparative learning method for extracting the ice lake, provided by the invention, is used for realizing automatic extraction of the ice lake from the remote sensing image under the condition of no tag data based on a convolutional neural network, and specifically comprises the following steps:
step 1, transforming an original image of the ice lake remote sensing image to obtain a transformation image, and combining the original image and the transformation image into a sample pair containing two branches.
The method includes the steps of performing transformation processing on an original image of an obtained ice lake remote sensing image to obtain a transformation image, converting the original image and the transformation image into a sample pair, and providing 4 image transformation modes for obtaining the sample pair which is mutually compared, wherein the image transformation modes specifically include:
(1) and (3) color mapping transformation: the specific parameters comprise the transformation of brightness, contrast, saturation and hue, wherein the maximum adjustment range of the brightness is 0.4; the maximum adjustment range of the contrast is 0.4; the maximum adjustment range of the saturation is 0.4; the maximum adjustment range of the color tone is 0.2;
(2) turning and transforming: namely, horizontal transformation and vertical transformation are carried out;
(3) gray level transformation: performing gray scale transformation with a probability of 0.2;
(4) fuzzy transformation: this embodiment may employ a blurring transformation of a gaussian kernel with the blurring parameter sigma set to 1.
The image transformation is generally implemented by using an algorithm for writing image transformation in Python language, one of the four ways is randomly adopted during image transformation, noise is randomly added, meanwhile, the image is processed into 448 × 448 dimensions, and then the transformed image and the original image obtained after processing jointly form a sample pair as the input of the network.
And 2, performing downsampling processing on the sample in a network with shared input weights, and respectively extracting the corresponding feature maps of the two branches under different scales.
The present embodiment selects a network in which weights are shared, and in this network, the parameters for performing downsampling on two branches are kept the same, and downsampling is performed on each of the two branches. The different scales include 4 scales, which are 224 × 224, 112 × 112, 56 × 56, and 28 × 28, respectively, one scale represents one downsampling block, each downsampling block includes two convolution layers and one downsampling layer, each convolution layer is activated using ReLU, each downsampling layer uses convolution operation with step size of 2, the result of activation of the last convolution layer before each downsampling layer is the ice lake feature map of different scales corresponding to the original image and the transformation map, and the ice lake feature maps of different scales include different ice lake details.
And 3, respectively inputting the corresponding feature maps of the two branches obtained in the step 2 under different scales into a projection layer for mapping to obtain feature vectors mapped by the two branches under the corresponding scales, and measuring the similarity degree of the two branches by adopting similarity loss based on the feature vectors, wherein the similarity loss is calculated by adopting cosine similarity, so as to obtain a comparison learning module of the ice lake.
Specifically, the feature maps corresponding to the two branches obtained in step 2 under different scales are respectively input into a projection layer for mapping, that is, the ice lake feature maps of different scales corresponding to the original image and the ice lake feature maps of different scales corresponding to the transformed map are respectively input into the projection layer, the projection layer includes three full-connected layers, the mapping means that ReLU is respectively adopted to activate the output of the feature maps entering the two full-connected layers, so that the mapping results of the two branches under the corresponding scales are obtained at the third full-connected layer, the mapping results are feature vectors corresponding to different scales, the feature vectors are used for calculating the similarity loss of the mapping results of the two branches under the corresponding scales, the similarity loss means the similarity degree of the features of the two similar images after passing through the comparison branch shared by weight values, and thus a comparison learning module of the ice lake is obtained, and the module can learn the similar features of the two branches in the feature maps corresponding to different scales.
In this embodiment, the similarity loss is measured by cosine similarity, and the calculation formula is as follows:
wherein q represents an ice lake characteristic diagram of the original image, q' represents an ice lake characteristic diagram of the transformation diagram, i represents different scales, i is a positive integer, and i is more than or equal to 1and less than or equal to 4; p (q) i ) Representing the feature vector after the ice lake feature map of the original image is mapped; p (q' i ) Representing the feature vector after the ice lake feature map of the transformation map is mapped; | | p (q) i )|| 2 Representing the L2 norm of the feature vector after the ice lake feature map of the original image is mapped; l | p (q' i )|| 2 And the L2 norm of the feature vector after mapping of the ice lake feature map representing the transformed map is provided with 4 features with different scales, so that the similarity under 4 scales is accumulated to further obtain a comparison learning module of the ice lake.
And 4, calculating a water body index NDWI spectral feature map of the original image, setting a water body index threshold value as T, and obtaining a rough distribution area map of the ice lake of the original image, wherein the value of T is 0.7, and the rough distribution area map of the ice lake is used as a pseudo label for comparison learning in the step 3 and is used for guiding the identification of the features of the ice lake.
Because the comparison learning can only learn similar features in the image, but cannot learn whether an image element in the image is the ice lake, in this embodiment, a relatively high water body index threshold T is used to obtain a rough ice lake distribution area map in combination with a Normal Differential Water Index (NDWI) spectral feature map, and the result is used as a pseudo label learned by the comparison learning module in step 3, so as to guide the ice lake training model to further recognize the ice lake features, and improve the accuracy of the ice lake training model. The calculation formula of the water body index NDWI spectral characteristic diagram is as follows:
where ρ is Green The apparent reflectivity of the atmosphere top layer of the green light wave band is expressed, and is more than 0 and less than rho Green ＜1，ρ NIR The apparent reflectivity of the top atmosphere layer in the near infrared band is expressed, and is more than 0 and less than rho NIR ＜1。
In order to ensure that all noise interference can be eliminated, the embodiment selects to set a higher water body index threshold T to be 0.7, and if the NDWI pixel is greater than the threshold, the NDWI pixel is marked as a water body.
And 5, performing up-sampling processing on the corresponding ice lake characteristic graphs of the original image in the two branches obtained in the step 2 under different scales in a network with shared weight values, and outputting an ice lake extraction result corresponding to the original image.
And (3) respectively taking the original image with 4 scales corresponding to the ice lake characteristic diagram obtained in the step (2) as up-sampling blocks, wherein each up-sampling block comprises an up-sampling layer and two convolution layers, the up-sampling layer uses deconvolution operation, each convolution layer is activated by using a ReLU function, and the final reduction result is the ice lake extraction result of the original image.
And 6, combining the pseudo label obtained in the step 4 and the ice lake extraction result obtained in the step 5 to calculate the position loss of the pseudo label and the ice lake extraction result, calculating the position loss by adopting an L2 norm to further obtain a position learning module of the ice lake, and combining the position learning module and the ice lake comparison learning module obtained in the step 3 to obtain the ice lake extraction model.
In this embodiment, the position loss is measured by using an L2 norm, and the formula is as follows:
wherein I represents an input image; b represents an ice lake mask obtained after threshold according to the water body index NDWI; f (-) represents an ice lake extraction model; u represents an arbitrary position in the image; f. of u (I) An extraction result representing any position in the input image; b is u Representing the result of the masking at any location of the image.
And 7, inputting any ice lake remote sensing image into the ice lake extraction model in the step 6, and outputting an ice lake extraction result.
Generally, the method is an unsupervised comparative learning ice lake extraction method, can effectively and accurately extract ice lake boundaries, and does not need any label training data.
The effects of the present invention can be further explained by the following experiments.
1. Conditions of the experiment
The invention is in the central processing unit
The I5-9400F 2.9GHz CPU, the GTX 1660T 6G GPU and the memory 16G, WINDOWS are realized by python language programming on an operating system. The model also relates to a deep learning framework, and the deep learning framework adopted in the experiment is tensorflow 1.14. The data used in the experiment were collected from Landsat-8 images, all of which were 256 × 256 × 7 in size.
2. Content of the experiment
The experimental accuracy evaluation was calculated by using F1 score. The specific calculation of the F1 score is as follows:
precision = correctly extracted ice lake image element/all extracted image elements
Recall = correctly extracted ice lake picture element/all ice lake picture elements
F1 score =2 × precision × recall/(precision + recall)
The size of the input image is set to 224 × 224 × 7. With respect to the training parameters, the batch size is set to 8, the epoch is set to 100, the dropout is set to 0.5 to prevent overfitting, the optimizer selects AdamaOptimizer, and the learning rate is set to 0.0005. <xnotran> , , , , (Wangchuk S.; bolch T.Mapping of glacial lakes using Sentinel-1and Sentinel-2data and a random forest classifier:Strengths and challenges.Science of Remote Sensing,2020,2.), U-net , (Qayyum N.; ghuffar S.; ahmad H.M.; yousaf A.; shahid I.Glacial Lakes Mapping Using Multi Satellite PlanetScope Imagery and Deep Learning.ISPRS International Journal of Geo-Information, 5754 zxft 5754.) (Wu R.; liu G.; zhang R.; wang X.; li Y.; zhang B.; cai J.; xiang 3252 zxft 3252 Deep Learning Method for Mapping Glacial Lakes from the Combined Use of Synthetic-Aperture Radar and Optical Satellite Images.Remote Sensing, 3532 zxft 3532.) GAN-GL, (Zhao H.; zhang M.; chen F.GAN-GL: generative Adversarial Networks for Glacial Lake Mapping.Remote Sensing, 3425 zxft 3425.). </xnotran> Unsupervised ice lake Extraction algorithms include the modified C-V model, references (Zhao H.; chen F.; zhang M.A Systematic Extraction application for Mapping Global lakes in High Mountain Regions of Asia. IEEE Journal of Selected Topics in Applied Earth objectives and Remote Sensing,2018,11,2788-2799.), global-local iterative segmentation algorithms, references (Li, J.; sheng, Y.an automated scheme for Global lake absorption Mapping in Landsat imaging and Digital Elevation model 2012: A e Study in the Himalayas. J. Remote. Registration, 94-5113). In the random forest method, about 198 ten thousand sample points are randomly sampled, wherein the number of ice lake pixels and background pixels is kept consistent, and 70% of the sample points are used for training and 30% are used for testing. The improved C-V model algorithm is based on the principle of region segmentation, the model has better noise immunity, and simultaneously, the quadratic term in the C-V model is simplified into the primary term, so that the model iteration process is accelerated. The global-local iterative segmentation algorithm is to firstly use the water body index NDWI to carry out rough extraction on the ice lake, then establish a buffer area for the rough extraction result, and further extract ice lake information in the buffer area by using a double-peak threshold segmentation algorithm. In the embodiment, the data set is a Landsat-8 remote sensing image, and a size of 256 × 256 × 7 ice lake image is obtained by cutting, and similarly, 70% of the images in the data set are used for supervising the training process in the model, and 30% of the images are used for testing. The unsupervised model was tested directly with 30% images, and the final results of comparing the accuracy of different models are shown in table 1 below:
TABLE 1 results of accuracy comparisons of typical model ice lake extractions
As can be seen from table 1, compared with the conventional ice lake extraction model and deep learning model, the model provided by the invention can learn similar objects in an image by a comparison method due to the introduction of unsupervised comparison learning, and can directly output ice lake information in the model by introducing a pseudo label, so that a heavy process of making a large number of training labels is avoided, and the model can achieve an effect close to that of a supervision method and far exceeds the result of an unsupervised method.
Although embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that changes and modifications may be made in the above embodiments without departing from the spirit of the invention, and the scope of the appended claims is to be construed as being limited only by the scope of the claims.
Claims (7)
1. An unsupervised comparative learning ice lake extraction method is characterized by comprising the following steps:
step 1, transforming an original image of the ice lake remote sensing image to obtain a transformation image, and combining the original image and the transformation image into a sample pair comprising two branches;
step 2, carrying out downsampling processing on a sample in a network with shared input weights, and respectively extracting corresponding feature maps of two branches under different scales;
step 3, respectively inputting the corresponding feature maps of the two branches obtained in the step 2 under different scales into a projection layer for mapping to obtain feature vectors mapped by the two branches under corresponding scales, and measuring the similarity degree of the two branches by adopting similar loss based on the feature vectors, wherein the similar loss is calculated by adopting cosine similarity, so as to obtain a comparison learning module of the ice lake;
step 4, calculating a water body index NDWI spectral feature map of the original image, setting a water body index threshold value as T, further obtaining a rough ice lake distribution area map of the original image, and using the rough result map as a pseudo label learned by the comparison learning module obtained in the step 3 for guiding the recognition of the ice lake features;
step 5, performing up-sampling processing on the corresponding ice lake characteristic graphs of the original image in the two branches obtained in the step 2 under different scales in a network with shared weight values, and outputting an ice lake extraction result corresponding to the original image;
step 6, calculating the position loss by combining the pseudo label obtained in the step 4 and the ice lake extraction result obtained in the step 5 to further obtain an ice lake position learning module, and combining the position learning module with the comparison learning module obtained in the step 3 to obtain an ice lake extraction model;
and 7, inputting any ice lake remote sensing image into the ice lake extraction model obtained in the step 6, and outputting an ice lake extraction result.
2. The unsupervised comparative learning ice lake extraction method according to claim 1, characterized in that:
in step 1, the transformation process is performed by any one of color mapping transformation, inversion transformation, gray-scale transformation, and fuzzy transformation, and noise is randomly added while the image size is processed to 448 × 448.
3. The unsupervised comparative learning ice lake extraction method according to claim 2, characterized in that:
in step 2, the scales include 4 scales, the sizes of the scales are respectively 224 × 224, 112 × 112, 56 × 56 and 28 × 28, one scale is a downsampling block, each downsampling block respectively includes two convolutional layers and one downsampling layer, each convolutional layer is activated by using a ReLU, and each downsampling layer uses a convolution operation with a step size of 2; the ice lake characteristic diagram of the original image and the ice lake characteristic diagram of the transformation image are respectively the result of activation of the last convolution layer before each downsampling layer under the corresponding scale.
4. The unsupervised comparative learning ice lake extraction method according to claim 3, characterized in that:
in step 3, the projection layer includes three fully-connected layers, the mapping is to activate the feature maps entering the two fully-connected layers by using ReLU, so as to obtain a mapping result at the third fully-connected layer, the mapping result is a feature map vector corresponding to different scales, the vector is used for calculating a similarity loss of the mapping result, the similarity loss is calculated by using cosine similarity, and the calculation formula is as follows:
wherein q represents the ice lake characteristic diagram of the original image, q' represents the ice lake characteristic diagram of the transformation diagram, i represents different scales, i is a positive integer, and i is more than or equal to 1and less than or equal to 4; p (q) i ) Representing the feature vector after the ice lake feature map of the original image is mapped; p (q' i ) Representing the feature vector after the ice lake feature map of the transformation map is mapped; | | p (q) i )|| 2 Representing the L2 norm of the feature vector after the ice lake feature map of the original image is mapped; l | p (q' i )|| 2 And the L2 norm of the feature vector after the ice lake feature map of the transformed map is mapped is shown.
5. The unsupervised comparative learning ice lake extraction method according to claim 4, characterized in that:
in step 4, the calculation formula of the water body index NDWI spectral feature map is as follows:
where ρ is Green The apparent reflectivity of the atmosphere top layer of green light wave band is expressed, and is more than 0 and less than rho Green ＜1，ρ NIR The apparent reflectivity of the top atmosphere layer in the near infrared band is expressed, and is more than 0 and less than rho NIR Less than 1; the threshold value T of the water body index is 0.7.
6. The unsupervised comparative learning ice lake extraction method according to claim 5, characterized in that:
the step 5 specifically comprises the following steps: and (3) respectively taking the corresponding ice lake characteristic diagrams of the original image in the two branches obtained in the step (2) at different scales as up-sampling blocks, wherein each up-sampling block comprises an up-sampling layer and two convolution layers, the up-sampling layer uses deconvolution operation, each convolution layer is activated by using a ReLU function, and the final reduction result is the ice lake extraction result of the original image.
7. The unsupervised comparative learning ice lake extraction method according to claim 6, characterized in that:
in step 6, the position loss is calculated by using an L2 norm, and a calculation formula of the L2 norm is as follows:
wherein I represents an input image; b represents an ice lake mask obtained after a water body index NDWI threshold value is obtained; f (-) represents an ice lake extraction model; u represents an arbitrary position in the image; f. of u (I) An extraction result representing any position in the input image; b is u Representing the result of the masking at any location of the image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211216115.5A CN115620132A (en) | 2022-09-30 | 2022-09-30 | Unsupervised comparative learning ice lake extraction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211216115.5A CN115620132A (en) | 2022-09-30 | 2022-09-30 | Unsupervised comparative learning ice lake extraction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115620132A true CN115620132A (en) | 2023-01-17 |
Family
ID=84861204
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211216115.5A Pending CN115620132A (en) | 2022-09-30 | 2022-09-30 | Unsupervised comparative learning ice lake extraction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115620132A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116481600A (en) * | 2023-06-26 | 2023-07-25 | 四川省林业勘察设计研究院有限公司 | Plateau forestry ecological monitoring and early warning system and method |
-
2022
- 2022-09-30 CN CN202211216115.5A patent/CN115620132A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116481600A (en) * | 2023-06-26 | 2023-07-25 | 四川省林业勘察设计研究院有限公司 | Plateau forestry ecological monitoring and early warning system and method |
| CN116481600B (en) * | 2023-06-26 | 2023-10-20 | 四川省林业勘察设计研究院有限公司 | Plateau forestry ecological monitoring and early warning system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108230329B (en) | Semantic segmentation method based on multi-scale convolution neural network | |
| CN108154192B (en) | High-resolution SAR terrain classification method based on multi-scale convolution and feature fusion | |
| CN110599537A (en) | Mask R-CNN-based unmanned aerial vehicle image building area calculation method and system | |
| CN108596108B (en) | Aerial remote sensing image change detection method based on triple semantic relation learning | |
| CN109035172B (en) | Non-local mean ultrasonic image denoising method based on deep learning | |
| CN107808138B (en) | Communication signal identification method based on FasterR-CNN | |
| CN109753996B (en) | Hyperspectral image classification method based on three-dimensional lightweight depth network | |
| CN112464745B (en) | Feature identification and classification method and device based on semantic segmentation | |
| CN111178438A (en) | ResNet 101-based weather type identification method | |
| CN110598613A (en) | Expressway agglomerate fog monitoring method | |
| CN111008664A (en) | Hyperspectral sea ice detection method based on space-spectrum combined characteristics | |
| Chen et al. | Scene segmentation of remotely sensed images with data augmentation using U-net++ | |
| CN115620132A (en) | Unsupervised comparative learning ice lake extraction method | |
| CN114841992A (en) | Defect detection method based on cyclic generation countermeasure network and structural similarity | |
| Wang | Remote sensing image semantic segmentation algorithm based on improved ENet network | |
| CN111222576B (en) | High-resolution remote sensing image classification method | |
| CN116843938A (en) | Mixed classification method for high-spatial-resolution remote sensing image | |
| Harika et al. | Extracting water bodies in rgb images using deeplabv3+ algorithm | |
| CN114862883A (en) | Target edge extraction method, image segmentation method and system | |
| CN111046861B (en) | Method for identifying infrared image, method for constructing identification model and application | |
| CN114187477A (en) | Small sample hyperspectral image classification method based on supervised self-contrast learning | |
| Peng et al. | Recognition and classification of water surface targets based on deep learning | |
| Que et al. | Low altitude, slow speed and small size object detection improvement in noise conditions based on mixed training | |
| Nakhatovich et al. | Applications of classical and deep learning techniques for polar bear detection and recognition from aero photography | |
| CN117036893B (en) | Image fusion method based on local cross-stage and rapid downsampling |
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 |