CN111008664A - Hyperspectral sea ice detection method based on space-spectrum combined characteristics - Google Patents

Abstract

The invention discloses a hyperspectral sea ice detection method based on space-spectrum combined characteristics, which overcomes the defect that the traditional method limits the improvement of the classification precision of hyperspectral sea ice images, and the technical scheme is characterized by comprising the following steps: acquiring original data through an original hyperspectral image; extracting spatial features from the original data and obtaining a label sample; selecting and processing neighbor unlabeled samples of each labeled sample from original data; fusing the labeled and unlabeled exemplars to form input data; randomly dividing input data into training samples and testing samples, training the pre-built three-dimensional CNN network model through the training samples, and testing through the testing samples to obtain classification accuracy; the hyperspectral remote sensing images are detected and classified through the three-dimensional CNN network model after training and testing, and the hyperspectral sea ice detection method based on the space-spectrum combined characteristics can effectively overcome the existing difficulties and improve the classification precision of hyperspectral sea ice.

Description

Hyperspectral sea ice detection method based on space-spectrum combined characteristics

Technical Field

The invention relates to the field of sea ice detection, in particular to a hyperspectral sea ice detection method based on space-spectrum combined characteristics.

Background

Sea ice is an important component of the earth's climate system and also one of the indicators of global climate change. It plays an important role in the heat exchange between the sea and the atmosphere. Meanwhile, sea ice is also one of the causes of marine disasters in polar regions and high and medium latitude areas. The method has important research significance for the research of sea ice detection. Sea ice detection needs to acquire timely and effective data, and the remote sensing technology has the characteristics of timeliness, accuracy and capability of acquiring large-range data, and becomes an important means for analyzing and researching sea ice. The hyperspectral remote sensing data is large in data volume and multiple in wave bands, continuous and high-resolution one-dimensional spectral information representing physical attributes of target ground objects can be obtained while two-dimensional spatial image information of the target ground objects is obtained, the remote sensing data image dimension and the spectral dimension information are organically fused, namely the hyperspectral remote sensing data has the characteristic of 'map integration', and important data support is provided for accurate sea ice classification.

1) According to the spectral characteristics, the hyperspectral remote sensing has multiple wave bands and narrow wave width, nearly continuous spectral information can be obtained, and finer spectral characteristics are provided for sea ice detection. Common methods include conventional minimum distance, maximum likelihood estimation, decision tree, Support Vector Machine (SVM), and the like.

2) According to the spatial characteristics, a hyperspectral image can obtain a wide range of data images, and different ground object types have abundant and diverse spatial information. Different types of sea ice tend to have different surface roughness and grain direction, exhibiting different shapes and elevations, while the same type of sea ice surface tends to exhibit approximately the same textural characteristics at a scale. Common spatial feature extraction methods include Gray-level Co-occurrence Matrix (GLCM), Gabor filter, morphological contour, and the like.

The problems of strong correlation between wave bands, redundant mixed pixels and data information and the like are also brought by the high dimensionality and large data volume of the hyperspectral data, the Hughes phenomenon is easily caused, and the problem of sea ice detection by using hyperspectral remote sensing is solved. Because the physical characteristics of the sea ice and the environmental factor influence of the distribution of the sea ice, under specific conditions, the phenomenon of 'same-spectrum foreign matter, same-object different spectrum' can occur when the optical remote sensing data is used for sea ice detection, so that the spatial characteristics are effectively utilized and combined with the spectral characteristics by introducing the spatial characteristics to carry out the sea ice detection, the information complementation is realized, and the improvement of the sea ice detection precision is facilitated. However, deep information is often ignored in the method of extracting shallow manual features based on expert experience and priori knowledge, which is widely adopted at present, so that the improvement of the classification accuracy of the hyperspectral sea ice image is limited.

In addition, due to the particularity of the geographic environment of the sea ice coverage area, field investigation is difficult, actual ground object class information is difficult to obtain, the sea ice type needs to be manually marked through priori knowledge, however, manual marking is long in time consumption and high in cost, the problems that training samples are difficult to obtain and sample quality is difficult to guarantee are easily caused, and improvement of classification accuracy of hyperspectral sea ice images is limited to a certain extent

Disclosure of Invention

The invention aims to provide a hyperspectral sea ice detection method based on space-spectrum combined characteristics, which can effectively overcome the existing difficulties and improve the classification precision of hyperspectral sea ice.

The technical purpose of the invention is realized by the following technical scheme:

a hyperspectral sea ice detection method based on space-spectrum combined characteristics comprises the following steps:

acquiring original data through an original hyperspectral image;

extracting spatial features from the original data through a GLCM algorithm, and stacking and fusing the extracted spatial features and the original data to obtain a label sample;

selecting neighbor unlabeled samples of all labeled samples from the original data according to a K neighbor algorithm, and processing the unlabeled samples through a waveband selection algorithm and correlation analysis;

fusing the labeled and unlabeled samples to form input data;

randomly dividing input data into training samples and testing samples, training the pre-built three-dimensional CNN network model through the training samples, and testing the trained three-dimensional CNN network model through the testing samples to obtain classification accuracy;

and detecting and classifying the hyperspectral remote sensing images through the trained and tested three-dimensional CNN network model.

Preferably, the extracting of the spatial features specifically includes:

processing the original data by adopting a principal component analysis algorithm to obtain a first principal component;

sliding a sliding window on a first principal component according to a set step length and a set direction angle according to a GLCM (global solution for communications) algorithm, calculating a gray level co-occurrence matrix of the sliding window and obtaining a texture characteristic value every time the sliding window slides, giving the texture characteristic value of the window to a central pixel point of the window, and sliding the window until the first principal component is covered;

and summing and averaging all direction angles of each texture feature to obtain a texture feature matrix so as to finish spatial feature extraction.

Preferably, the obtaining of the neighbor unlabeled sample specifically includes:

and calculating Euclidean distances between the label samples in the original data and all the unlabeled samples, sequencing, and finishing the calculation of all the label samples so as to finish the extraction of the adjacent unlabeled samples of all the label samples.

Preferably, the treatment of the unlabeled sample is specifically as follows:

performing dimensionality reduction on original data according to a wave band selection algorithm of the hyperspectral remote sensing sea ice image to obtain spectral characteristics of an unlabeled sample;

performing correlation analysis on the texture features of the unlabeled samples, and removing components with high correlation to obtain spatial features;

and (4) superposing and fusing the spatial features and the spectral features extracted from the unlabeled samples to complete the processing and acquisition of the unlabeled samples.

Preferably, the training of the pre-established three-dimensional CNN network model specifically comprises:

randomly dividing input data formed by fusing the labeled samples and the unlabeled samples into training samples and testing samples according to a training strategy, wherein the input size of each sample is KxKxB, wherein K is the space dimension size of a pixel block and is an odd number; b is the size of the pixel block in the channel dimension, which is the depth of the input data;

training is carried out: randomly inputting a plurality of training samples from the training samples to a pre-established three-dimensional CNN network model for training each time;

assuming that the first layer includes n convolution kernels of size C × C × D, each sample of size K × K × B is subjected to convolution operation of the first layer, and then n data cubes of size (K-C +1) × (K-C +1) × (B-D +1) are output; the output of the first layer is used as the input of the second layer, convolution operation is continued, the analogy is repeated, the final output is converted into a feature vector to be input into the full-connection layer, local feature mapping extracted in the convolution process is mapped and fused, after loss rate is calculated through a Softmax cross entropy function, the gradient of each parameter is calculated through back propagation, and network parameters are dynamically updated through an Adam algorithm;

and (4) repeatedly training the three-dimensional CNN network model by each training sample until the preset iteration times are finished, and finishing the training.

Preferably, the testing of the pre-established three-dimensional CNN network model specifically comprises:

and inputting the test sample into the trained three-dimensional CNN network model, calculating a confusion matrix according to a predicted label obtained by prediction and a directly obtained real label, obtaining classification precision and completing the test.

In conclusion, the invention has the following beneficial effects:

aiming at the main problem in hyperspectral sea ice detection, the problem of insufficient sea ice information mining in traditional sea ice detection is solved by utilizing a three-dimensional CNN network model, the defect of manually marking sample information is made up by utilizing a large amount of unlabeled sample information, the marking pressure is reduced, and the diversity of the sample information is increased;

on the basis of combining a three-dimensional CNN algorithm and a GLCM algorithm, a small number of labeled samples and a large number of unlabeled samples are combined, the diversity of training sample information is increased, and the hyperspectral sea ice classification precision is improved;

redundant spectral information and spatial information in unlabeled samples are respectively eliminated through a band selection algorithm and correlation analysis, and the training time of the model is effectively shortened.

Drawings

FIG. 1 is a schematic block flow diagram of the process;

FIGS. 2a-2h are schematic diagrams of classification results of various classification methods;

FIG. 3 is a comparison graph of the overall classification accuracy of the method and other methods under different size samples;

fig. 4 is a graph showing the influence of the K value on the classification accuracy.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings.

According to one or more embodiments, a hyperspectral sea ice detection method based on space-spectrum combined characteristics is disclosed, which comprises the following steps:

acquiring original data through an original hyperspectral image;

extracting spatial features from the original data through a GLCM algorithm, and stacking and fusing the extracted spatial features and the original data to obtain a label sample;

selecting neighbor unlabeled samples of all labeled samples from the original data according to a K neighbor algorithm, and processing the unlabeled samples through a waveband selection algorithm and correlation analysis;

fusing the labeled and unlabeled samples to form input data;

randomly dividing input data into training samples and testing samples, training the pre-built three-dimensional CNN network model through the training samples, and testing the trained three-dimensional CNN network model through the testing samples to obtain classification accuracy;

and detecting and classifying the hyperspectral remote sensing images through the trained and tested three-dimensional CNN network model.

Wherein, the extraction of the spatial characteristics specifically comprises:

processing the original data by adopting a principal component analysis algorithm to obtain a first principal component;

sliding a sliding window on a first principal component according to a set step length and a set direction angle according to a GLCM (global solution for communications) algorithm, calculating a gray level co-occurrence matrix of the sliding window and obtaining a texture characteristic value every time the sliding window slides, giving the texture characteristic value of the window to a central pixel point of the window, and sliding the window until the first principal component is covered;

and summing and averaging all direction angles of each texture feature to obtain a texture feature matrix so as to finish spatial feature extraction.

The acquisition of the neighbor unlabeled sample specifically comprises the following steps:

and calculating Euclidean distances between the label samples in the original data and all the unlabeled samples, sequencing, and finishing the calculation of all the label samples so as to finish the extraction of the adjacent unlabeled samples of all the label samples.

The method comprises the following specific steps:

inputting: raw hyperspectral image

A. Feature extraction

1) Processing the original data set by adopting a principal component analysis algorithm, and taking a first Principal Component (PC);

2) according to a GLCM algorithm, a sliding window slides on a PC according to a step length d (d is equal to 1) and a direction angle q (q is equal to 0 degrees, 45 degrees, 90 degrees and 135 degrees), a gray level co-occurrence matrix of the sliding window is calculated every time the sliding window slides, a texture characteristic value is obtained, and the texture characteristic value of the window is assigned to a central pixel point of the window;

3) repeating the step 2) until the sliding window covers the PC;

4) for a certain texture scalar quantity, summing and averaging texture feature matrixes under 4 direction angles to serve as a final texture feature matrix of the texture feature, and the same is true for other texture features;

5) repeating the step 4) until 8 texture features are obtained;

6) completing the spatial feature extraction;

B. selection of neighbor samples

7) Calculating Euclidean distance between a certain labeled sample and all unlabeled samples in the original data, and calculating the formula

And arranged in ascending order;

8) repeating the step 7) until all the label samples are calculated;

9) extracting the neighbor unlabeled samples of all the labeled samples;

C. input data pre-processing

i. Label sample

10) Stacking the spatial features of the label sample extracted in the stage A with the original data;

11) completing the acquisition of label sample data;

unlabeled samples

12) Performing dimensionality reduction on original hyperspectral data according to a waveband selection algorithm of the hyperspectral remote sensing sea ice image to obtain spectral characteristics;

13) performing correlation analysis on the spatial features extracted from the unlabeled samples in the stage A, and removing components with high correlation to obtain the spatial features;

14) superposing the spectral features and the spatial features of the unlabeled sample extracted in the steps 12) and 13);

15) completing the acquisition of unlabeled sample data;

16) fusing the label sample data in the i and the unlabeled sample data in the ii according to the spatial spectrum characteristics (spatial characteristics and spectral characteristics) of the adjacent unlabeled sample extracted by the adjacent relation in the stage B;

17) completing the acquisition of input data;

D. three-dimensional CNN

18) Randomly dividing input data into training samples and testing samples according to a training strategy, wherein the input size of each sample is KxKxB (K is the size of a space dimension of a pixel block and is an odd number, and B is the size of the pixel block in a through dimension and is the depth of the input data);

a training stage:

19) randomly inputting batch (20) training samples from the training samples to a pre-established three-dimensional CNN network for training each time;

20) assuming that the first layer includes n convolution kernels of size C × C × D, each sample of size K × K × B is subjected to convolution operation of the first layer, and then n data cubes of size (K-C +1) × (K-C +1) × (B-D +1) are output; the output of the first layer is used as the input of the second layer, and the convolution operation (similar to the first layer) is continued; and so on; converting the final output into a feature vector, inputting the feature vector into a full-connection layer, mapping and fusing local features extracted in the convolution process, calculating the loss rate through a Softmax cross entropy function, calculating the gradient of each parameter through back propagation, and dynamically updating the network parameters by using an Adam algorithm;

21) repeating steps 19) and 20) until a predetermined number of iterations is completed;

22) finishing the model training;

and (3) a testing stage:

23) inputting the test sample in the step 18) into the trained three-dimensional CNN model, calculating a confusion matrix according to the predicted label and the real label, and obtaining classification precision;

24) the test is finished;

E. and (3) outputting: confusion matrix, Overall classification accuracy (OA), Average classification accuracy (AA), Kappa value (Kappa static).

And detecting and classifying the hyperspectral remote sensing images through the three-dimensional CNN network model which is trained and tested.

The three-dimensional CNN network model can simultaneously extract spectral characteristics and spatial characteristics with a one-dimensional CNN model based on spectral characteristics and a two-dimensional CNN model based on spatial characteristics, and sea ice characteristic information hidden in remote sensing data is fully utilized. And the hyperspectral data is usually represented by a three-dimensional data cube, an input mode of extracting features by a three-dimensional convolution filter in the CNN is matched, and a simple, convenient and effective method is provided for simultaneously extracting the spatial-spectral combined features, so that the three-dimensional CNN network model is a classification model suitable for hyperspectral sea ice remote sensing images.

Due to the obvious difference of the texture characteristics of different types of sea ice, the GLCM is used for enhancing the texture characteristics, so that the identification and classification of the sea ice are more facilitated. Sea ice texture features are extracted based on a GLCM method, texture information is combined with sea ice spectral space information, and the quality of a marked sample is further enhanced by utilizing the spatial spectral features of a large number of adjacent unmarked samples. The near neighbor sample is selected through the KNN algorithm which is free of any parameter to be trained and is simple and easy to achieve, the empty spectrum information of the unlabeled sample and the empty spectrum information of the labeled sample are fused, the quality of the labeled sample is improved, the diversity of the sample is increased, the problem of high labeling cost of hyperspectral sea ice data is solved, the quantity of the parameters to be trained in the three-dimensional CNN is reduced, and the sea ice classification precision is obviously improved under the condition of small samples.

The addition of a large amount of unlabeled sample information increases the computational effort. In order to reduce the temporal complexity of the model, the spectral-spatial information of the unlabeled samples is preprocessed. In the aspect of spectral information, a band algorithm is adopted for dimension reduction, a large number of redundant bands are removed, and a selected band combination with large information amount and low similarity is selected for deep spectral feature extraction. In the aspect of spatial information, low-correlation textural features are selected based on correlation analysis to perform deep spatial feature extraction, so that training time is effectively shortened.

As shown in fig. 1, it mainly comprises 4 parts: the method comprises the steps of spatial feature extraction based on GLCM, unlabeled sample data processing based on a waveband selection algorithm and correlation analysis, fusion of labeled samples and adjacent unlabeled samples, and deep spectrum spatial feature extraction and classification based on three-dimensional CNN. Firstly, extracting spatial features from original hyperspectral data by using GLCM (global matching pursuit), and fusing the extracted spatial features and the original data into label sample data. Selecting a Neighbor unlabeled sample of the labeled sample from the original hyperspectral data according to a K Neighbor (KNN, K-Nearest Neighbor) algorithm which does not contain any training parameter and is simple and easy to implement, and respectively removing redundant information in the spectral feature and the spatial feature of the Neighbor unlabeled sample through a waveband selection algorithm and correlation analysis. And finally, fusing the labeled sample data and the unlabeled sample data to be used as the input of the three-dimensional CNN for extracting the deep space spectrum characteristics.

For clarity, an example is given:

1) description of data

The experimental data is a Hyperion hyperspectral image of a Bohai Bay area shot by an earth observation satellite (EO-1) in 2008, 1 month and 23 days, the longitude and latitude of the upper left corner is 120 degrees, 45 '12' W, 41 degrees, 39 '7' N, and the longitude and latitude of the lower right corner is 121 degrees, 13 '9' E and 39 degrees, 44 '42' N. The experimental data are subjected to system geometric correction, projection registration and terrain correction, the image level is L1Gst level, the wave band spectrum range is 356-2578 nm, the wave band spectrum range totally comprises 242 wave bands, the spatial resolution is 30m, and the spectral resolution reaches the nanometer level. The original image size was 7061 × 2001, and after the cropping, a scene covered with sea ice was selected as an experimental area whose image size was 272 × 159. In the 242 wave band image data, after removing part of the low signal-to-noise ratio and water absorption wave band, 176 wave bands are finally used for analysis. A certain number of label samples are manually marked to serve as a sample library, the sample library is randomly divided into training samples and testing samples, and table 1 shows sample information of a Bohai Bay data set.

TABLE 1 training sample number (number of pixels) in each class of Bohai Bay dataset

2) Experimental setup

In the aspect of spectral characteristics of unlabeled samples, the similarity between the information of wave bands and the spectral characteristics of hyperspectral sea ice are comprehensively considered, wave band combinations with large information amount and low similarity between the wave bands are selected through a wave band selection algorithm, and finally 3 wave bands are selected as spectral wave bands, wherein the wave band numbers are 21, 120 and 83.

In addition, a correlation analysis method is adopted to exclude highly correlated components in the texture features extracted by the GLCM algorithm, quantitative analysis is carried out through a correlation coefficient matrix, and table 2 shows the correlation of 8 texture components in the Bohai Bay data set. When two texture features are highly correlated (the absolute value of the correlation coefficient is greater than 0.7), the texture component with the smaller average absolute correlation is selected for further study.

Table 2. correlation coefficient matrix of 8 texture components in Bohai Bay data set

In Table 2, Mea Mean (Mean); var Variance (Variance); hom Homogeneity (Homogeneity); con: Contrast (Contrast); dis: Dissimilarity (Dissimilarity); ent is Entropy (Encopy); ASM Angular second moment (Angular second moment); correlation (Correlation); ABC average absolute correlation (Averageabsolute correlation).

Combining the 3 bands after the band selection and the 5 texture components reserved after the correlation analysis, finally using 8 attribute features to represent the empty spectrum information of the unlabeled sample, selecting 20 neighboring samples of the labeled sample (K equals to 20), using the empty spectrum feature as the patch information of the labeled sample, and finally inputting the channel (depth) dimension of the data to be 176+8+8 × 20 equals to 344. The model structure of the Bohai Bay data set is shown in Table 3, the model comprises two convolutional layers, a Dropout layer, a full connection layer, an input layer and an output layer, the data input size is 5 multiplied by 344, the data input size is normalized to [0,1], the learning rate is 0.001, the Dropout value is 0.5, the data volume size of batch input network is 20, and the iteration number is 2000.

Table 3 model structure of Bohai Bay data set

3) Example results

Table 4 shows the comparison of the classification results of the method of this patent with other classification methods, each algorithm was performed 10 times under randomly selected samples, and the experimental results were in the form of mean ± standard deviation. The GLCM-CNN combines the spatial features extracted by the GLCM algorithm with the original data, and then inputs the fused data into a CNN network. In the experiment, the input size of the 2D-CNN is 19 × 19, the input size of the 3D-CNN is 9 × 9 × 176, and the input size of the GLCM-CNN is 5 × 5 × (176+8), where 176 is the number of original bands and 8 is 8 texture components extracted from the first principal component by the GLCM algorithm. As can be seen from Table 4, compared with other algorithms, the algorithm provided by the patent obtains the best classification result, the overall classification precision is 97.91%, which is 7.32% and 3.83% higher than that of 1D-CNN only using spectral features and 2D-CNN only using spatial features, respectively, and thus, the precision of sea ice classification can be effectively improved by using the spatial-spectral combination features; in the algorithm based on the space-spectrum features, the GLCM-CNN is higher than the 3D-CNN, which shows the effectiveness of texture information extracted by the GLCM algorithm, but the precision is improved by only 1.48%, and the method is higher than the 3D-CNN by 2.96%, compared with the result (96.43%) of the GLCM-CNN which is not fused with neighbor sample information, the method is improved by 1.48%, the AA and Kappa coefficients are respectively improved by 1.30% and 2.05%, which shows that the fusion of the spatial information of neighbor unlabeled samples can effectively enhance the quality of training samples and improve the sea ice classification precision.

Among the 4 categories, Seawater is easier to distinguish than sea Ice because of its low spectral reflectance due to its own characteristics, while Gray Ice and Gray-White Ice, which are intermediate categories between Seawater and White Ice, have a wide thickness range and are seriously classified by mistake. In the table 4, the classification accuracy of the Thin Ice in 7 algorithms is relatively low, the classification in a decision tree, an SVM and a 1D-CNN is difficult to distinguish only through spectral features, certain effects are obtained for classifying the Gray Ice and the Gray-White Ice through spatial features in a 2D-CNN and a GLCM-CNN, the spatial information of unlabeled adjacent samples is fused on the basis of the spatial features of the original labeled samples by the algorithm, the optimal classification result is obtained, the Gray Ice is respectively improved by 2.96% and 1.41%, and the Gray-White Ice is respectively improved by 3.23% and 1.77%. Fig. 2a-2h are corresponding classification result images, wherein fig. 2a is a composite image, and fig. 2b to 2h are a decision tree, a Support Vector Machine (SVM), a one-dimensional convolutional neural network (1D-CNN), a two-dimensional convolutional neural network (2D-CNN), a three-dimensional convolutional neural network (3D-CNN), a gray level co-occurrence matrix-convolutional neural network (GLCM-CNN), and a classification result image of the method adopted in the present invention, respectively. Because the method has the advantages of classifying the intermediate categories (Gray Ice and Gray-White Ice), the classification result of the method can effectively eliminate noise points, so that the classification result graph is smoother, and the edge regions of different categories are more accurately distinguished.

TABLE 4 classification results of Bohai Bay data set (%)

Because the CNN contains a large number of parameters to be trained, a large number of training samples are required to ensure the diversity of the samples, so as to extract the features with robustness and effectiveness. The method is researched in the aspect of training sample size, 10 times of experiments are carried out under different training sample sizes, and the average value of 10 times of experiment results is taken as a classification precision value. The Bohai Bay is located in a middle latitude area, does not contain perennial accumulated ice, is thin in ice thickness and easy to melt, causes the confusion of sea water and sea ice, is low in separability among different kinds of sea ice, and needs a certain amount of training samples for distinguishing. The method randomly selects the same number of training samples for different categories, as shown in fig. 3, the method and the decision tree, the Support Vector Machine (SVM), the one-dimensional convolutional neural network (1D-CNN), the two-dimensional convolutional neural network (2D-CNN), the three-dimensional convolutional neural network (3D-CNN), and the gray level co-occurrence matrix-convolutional neural network (GLCM-CNN) of the method are overall classification precision contrastive graphs under different size samples, the classification precision increases with the amount of the training samples, but the algorithm provided by the patent is superior to other algorithms under different training sample amounts, for example, when only 10% of training samples are selected in each category, the average precision value of 10 experiments performed by the algorithm is 92.74%, the comparison decision (82.64%), the SVM (65.33%), the 1D-CNN (86.85%), and the 2D-CNN (88.72%), 3D-CNN (85.99%), GLCM-CNN (89.13%), increased by 10.10%, 27.41%, 5.89%, 4.02%, 6.75% and 3.61%, respectively.

The KNN algorithm does not contain any parameter to be trained, and only the selection of the K value influences the result. The size of the K value represents the number of the neighbor samples in the selected unlabeled samples, the influence of the K value on the algorithm is mainly explored, and 5 experiments are carried out by randomly sampling under different K values, as shown in FIG. 4. With the increase of the K value, the classification accuracy value is in an increasing trend, but with the increase of the K value, the number of selected neighbor samples is also increased, and although abundant information is brought to the label samples, the channel (depth) dimension of input data is increased, and more time is required for training a robust and effective classification model, so that the influence on the classification accuracy and the model calculation complexity needs to be comprehensively considered when the appropriate K value is selected. In summary, in this patent, the K value is taken to be 20.

In addition, the traditional sea ice classification method mostly only utilizes spectral characteristics, does not fully utilize abundant spatial information contained in the hyperspectral remote sensing sea ice image, and limits further improvement of sea ice classification precision. The utility model provides a space spectrum characteristic is united sea ice image classification model based on convolution neural network combines together a small amount of label samples and a large amount of not label samples, fully excavates space and spectral information in the hyperspectral remote sensing sea ice data, improves sea ice classification precision. Compared with a classification method based on single characteristics (spectrum or space) and other space-spectrum combined characteristics, experimental analysis shows that sea ice space spectrum characteristics can be effectively extracted through a small amount of training samples by means of a large amount of unlabeled sample information, a better classification result is obtained on the whole, and a new thought is provided for remote sensing sea ice image classification.

The present embodiment is only for explaining the present invention, and it is not limited to the present invention, and those skilled in the art can make modifications of the present embodiment without inventive contribution as needed after reading the present specification, but all of them are protected by patent law within the scope of the claims of the present invention.

Claims (6)

1. A hyperspectral sea ice detection method based on space-spectrum combined characteristics is characterized by comprising the following steps:

acquiring original data through an original hyperspectral image;

extracting spatial features from the original data through a GLCM algorithm, and stacking and fusing the extracted spatial features and the original data to obtain a label sample;

selecting neighbor unlabeled samples of all labeled samples from the original data according to a K neighbor algorithm, and processing the unlabeled samples through a waveband selection algorithm and correlation analysis;

fusing the labeled and unlabeled samples to form input data;

randomly dividing input data into training samples and testing samples, training the pre-built three-dimensional CNN network model through the training samples, and testing the trained three-dimensional CNN network model through the testing samples to obtain classification accuracy;

and detecting and classifying the hyperspectral remote sensing images through the trained and tested three-dimensional CNN network model.

2. The hyperspectral sea ice detection method according to claim 1, wherein: the extraction of the spatial features specifically comprises:

processing the original data by adopting a principal component analysis algorithm to obtain a first principal component;

sliding a sliding window on a first principal component according to a set step length and a set direction angle according to a GLCM (global solution for communications) algorithm, calculating a gray level co-occurrence matrix of the sliding window and obtaining a texture characteristic value every time the sliding window slides, giving the texture characteristic value of the window to a central pixel point of the window, and sliding the window until the first principal component is covered;

and summing and averaging all direction angles of each texture feature to obtain a texture feature matrix so as to finish spatial feature extraction.

3. The hyperspectral sea ice detection method according to claim 2, wherein: the acquisition of the neighbor unlabeled sample specifically comprises the following steps:

and calculating Euclidean distances between the label samples in the original data and all the unlabeled samples, sequencing, and finishing the calculation of all the label samples so as to finish the extraction of the adjacent unlabeled samples of all the label samples.

4. The hyperspectral sea ice detection method according to claim 3, wherein the processing of unlabeled samples is specifically:

performing dimensionality reduction on original data according to a wave band selection algorithm of the hyperspectral remote sensing sea ice image to obtain spectral characteristics of an unlabeled sample;

performing correlation analysis on the texture features of the unlabeled samples, and removing components with high correlation to obtain spatial features;

and (4) superposing and fusing the spatial features and the spectral features extracted from the unlabeled samples to complete the processing and acquisition of the unlabeled samples.

5. The hyperspectral sea ice detection method according to claim 4, wherein the training of the pre-built three-dimensional CNN network model specifically comprises:

randomly dividing input data formed by fusing the labeled samples and the unlabeled samples into training samples and testing samples according to a training strategy, wherein the input size of each sample is KxKxB, wherein K is the space dimension size of a pixel block and is an odd number; b is the size of the pixel block in the channel dimension, which is the depth of the input data;

training is carried out: randomly inputting a plurality of training samples from the training samples to a pre-established three-dimensional CNN network model for training each time;

if the first layer comprises n convolution kernels with the size of C multiplied by D, and each sample with the size of K multiplied by B is subjected to convolution operation of the first layer, n data cubes with the size of (K-C +1) × (K-C +1) × (B-D +1) are output; the output of the first layer is used as the input of the second layer, convolution operation is continued, the analogy is repeated, the final output is converted into a feature vector to be input into the full-connection layer, local features extracted in the convolution process are mapped and fused, after loss rate is calculated through a Softmax cross entropy function, the gradient of each parameter is calculated through back propagation, and network parameters are dynamically updated through an Adam algorithm;

and (4) repeatedly training the three-dimensional CNN network model by each training sample until the preset iteration times are finished, and finishing the training.

6. The hyperspectral sea ice detection method according to claim 5, wherein the testing of the pre-built three-dimensional CNN network model specifically comprises:

and inputting the test sample into the trained three-dimensional CNN network model, calculating a confusion matrix according to a predicted label obtained by prediction and a directly obtained real label, obtaining classification precision and completing the test.

Images