A small sample mangrove biomass estimation method based on self-supervised perturbation feature reconstruction network

Abstract

The invention relates to a small sample mangrove biomass estimation method based on a self-supervised disturbance feature reconstruction network. First, Landsat8 remote sensing images and DEM elevation data are used to extract 22 original feature quantities of band information, vegetation index, texture characteristics and pressure height index; then, for the original feature quantities, some feature quantities are randomly selected to disturb the order and add interference, and the remaining feature quantities are used to guide the reconstruction of the feature quantities before the disturbance, so that the characteristics of the mangrove are deeply extracted from a small amount of data; then a learning task model is established, and a multi-view convolutional neural network is constructed based on the deep stacking of multiple multi-view cascade convolution modules to extract the deep characteristics of the mangrove; finally, a biomass estimator is constructed and the model parameters are fine-tuned to obtain the final model and obtain the accurate estimation value of the mangrove biomass under a small sample; so that the extraction performance of mangrove feature data is effectively improved under data scarcity, and then the mangrove biomass is accurately estimated.

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning, and in particular relates to a small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network. Background Art

[0002] Mangroves are woody plant communities that grow in tropical and subtropical coastal zones where land and sea meet. They play an extremely important role in purifying seawater, preventing wind and waves, maintaining biodiversity, and fixing and storing carbon. Accurately estimating mangrove biomass is an important part of studying the carbon cycle of mangrove ecosystems, which is conducive to better assessing the carbon sink potential of mangrove ecosystems.

[0003] Traditional mangrove survey methods have problems such as high cost and low efficiency, and will cause certain damage to mangroves, which greatly restricts the efficiency of mangrove biomass distribution estimation and monitoring. Due to the interference of various external environmental factors, the difficulty of mangrove biomass estimation has been significantly increased. Traditional linear regression and nonlinear regression models have poor performance in mangrove biomass estimation and are not suitable for practical applications.

[0004] With the development of artificial intelligence and machine learning, some researchers have adopted a data-driven approach to achieve mangrove biomass estimation. Common machine learning models such as decision trees, artificial neural networks, support vector machines, convolutional neural networks, etc. have high requirements on the quantity and quality of training data. However, in actual situations, it is difficult to obtain an ideal data source due to the existing survey methods, that is, there are non-ideal conditions of data scarcity, which leads to the inability of traditional machine learning models to effectively extract the characteristics of mangrove data for application.

[0005] Therefore, designing a small-sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network that can improve the effect of mangrove feature extraction under the condition of scarce sample data has become a technical problem that needs to be solved urgently. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a small sample mangrove biomass estimation method based on a self-supervised disturbance feature reconstruction network.

[0007] The technical solution of the present invention is a small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network, comprising the following steps:

[0008] s1, use Landsat8 remote sensing images and DEM elevation data to process small sample mangrove data, extract original feature quantities, use random window feature perturbation and noise masking on the original feature quantities, and obtain auxiliary data , perform data splicing on the features of the unmasked part to obtain the residual data and record the position information P of the disturbance window;

[0009] s2, construct a learning task model, use multiple multi-view convolution modules MVCCM to build a multi-view convolution neural network MVCNN, extract the deep features of mangroves from the residual data for output; then use multiple one-dimensional convolution and pooling layers to cascade to form a simplified decoder, flatten the deep features, consider the position information of the disturbance window, and splice with the auxiliary data as the input of the simplified decoder to obtain the reconstructed feature data of mangroves and the model loss function representing the reconstruction error of the disturbance feature in stage 1, and adjust the convolution weight and bias of the multi-view convolution module with the loss value of the reconstructed feature;

[0010] s3, a mangrove biomass identifier is constructed through two fully connected layers, and the deep characteristics of the mangrove are input to obtain the estimated biomass of the mangrove; the estimated biomass value is compared with the actual biomass value, the biomass estimation loss of stage 2 is calculated, and the model parameters are fine-tuned with the estimated loss value to obtain the final biomass estimation model.

[0011] As a further improvement of the present invention, the original feature quantities extracted in step s1 include four types: band information, vegetation index, texture feature and air pressure height index; the feature quantities included in the band information are coastal band, blue band, green band, red band, near infrared wave, short infrared wave 1 and short infrared wave 2; the feature quantities included in the vegetation index are normalized vegetation index NDVI, ratio vegetation index RVI, difference vegetation index DVI, soil adjusted vegetation index SAVI, enhanced vegetation index EVI and green band normalized difference vegetation index GNDVI; the feature quantities included in the texture feature are variance VAR, homogeneity HOM, contrast CON, heterogeneity DIS, entropy ENT, angular second moment SM, correlation COR and mean ME; the air pressure height index is terrain factor TOF.

[0012] As a further improvement of the present invention, the random window feature disturbance sets the window length, randomly selects the number of features corresponding to the window length in each original feature quantity data, and uses Gaussian noise to perform feature masking on the selected features. The expression is:

[0013] , , where P is the randomly selected position information of the disturbance window, It means randomly selecting an integer from a to b, L is the total number of input features, W is the window length, represents the generated auxiliary data, represents Gaussian noise, x\left [ {a,b} \right ] Represents the features of segment a to b of the original feature data, Indicates feature disturbance;

[0014] The expression of data splicing is: ,in, Represents the residual data after splicing, Represents a data concatenation operation.

[0015] As a further improvement of the present invention, after inputting the residual data, the multi-view convolution module first uses small-scale convolution to obtain small-view features using different convolution step sizes, and the expression is: ,in, represents the i-th step output of the small-scale convolution, Represents a small-scale convolution with a step size of i;

[0016] Based on the small field of view features, medium-scale convolution is used to obtain the medium field of view features of different branches using different convolution step sizes. The expression is: ,in, represents the i-th step output of the medium-scale convolution, represents a medium-scale convolution with a step size of i;

[0017] Finally, the small field of view features and the medium field of view features are obtained. Based on the cascade convolution mode, large-scale convolution is used to obtain the large field of view features under different convolution step sizes. The expression is: ,in, represents the i-th step output of the large-scale convolution, Represents a large-scale convolution with a step size of i; Represents feature concatenation operation;

[0018] The feature concatenation operation is used to process the large-view features to obtain the deep features output by a single multi-view convolution module, which can be expressed as ,

[0019] in, Represents a single MVCCM module output feature.

[0020] As a further improvement of the present invention, the multi-view convolutional neural network is constructed by cascading multiple multi-view convolutional modules and pooling operations; the output feature expression of the multi-view convolutional neural network is: ,in, represents the output features of the multi-view convolutional neural network, It represents the cascade operation of n modules, and Pool represents the pooling operation.

[0021] As a further improvement of the present invention, the expression of the input of the simplified decoder is: ,in, To simplify the input features of the decoder after concatenation, represents the feature tiling operation; the expression of the simplified decoder is ,in, represents the output features of the decoder; the learning task model sets a fully connected layer to process the output features of the decoder to obtain the reconstructed feature data of the mangrove forest, which is expressed as ,in, Represents the reconstructed feature data, Represents the fully connected layer processing the output features of the decoder.

[0022] As a further improvement of the present invention, the calculation formula of the model loss function is: ,in, represents the stage 1 loss, and They represent the reconstructed value and true value of the ith feature of the jth mangrove sample, respectively, and M is the total number of samples.

[0023] As a further improvement of the present invention, the expression of the biomass identifier is: ,in, represents the estimated mangrove biomass, represents the i-th fully connected layer; the calculation formula for the estimated biomass loss is ,in, represents the estimated biomass loss in phase 2, and represent the estimated value and true value of the biomass of the jth sample respectively.

[0024] After adopting the above method, 22 original feature quantities of band information, vegetation index, texture characteristics and pressure height index were extracted through Landsat8 remote sensing images and DEM elevation data; then, for the original feature quantities, some feature quantities were randomly selected to perturb the order and add interference, and the remaining feature quantities were used to guide the reconstruction of the feature quantities before disturbance, so that the characteristics of mangroves were deeply extracted from a small amount of data; then a learning task model was established, and a multi-view convolutional neural network was constructed based on the deep stacking of multiple multi-view cascade convolution modules to extract the deep features of mangroves; finally, a biomass estimator was constructed and the model parameters were fine-tuned to obtain the final model and obtain the accurate estimation value of mangrove biomass under small samples, forming a complete self-supervised perturbation feature reconstruction network; the estimation method based on the self-supervised perturbation feature reconstruction network composed of the above technical scheme was realized, which can avoid the limitation of the original data sample size when extracting the deep features of mangroves, solve the problem of data scarcity, improve the generalization of the model, and effectively improve the accuracy of mangrove biomass estimation under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Shown is a flow chart of the forest biomass estimation method of the present invention.

[0026] Figure 2 Shown is a structural diagram of the self-supervised disturbance feature reconstruction network model of the present invention.

[0027] Figure 3 Shown is a diagram of the multi-view convolutional neural network structure of the present invention.

[0028] Figure 4 Shown is a diagram illustrating the original characteristic quantities of mangroves according to the present invention.

[0029] Figure 5 Shown is a data description diagram of the present invention.

[0030] Figure 6 Shown is a graph of biomass estimation loss and self-supervised learning loss of the present invention.

[0031] Figure 7 Shown is a comparison diagram of the mangrove biomass estimation results of the present invention. DETAILED DESCRIPTION

[0032] like Figure 1-Figure 7 A small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network is shown, comprising the following steps:

[0033] s1, use Landsat8 remote sensing images and DEM elevation data to process small sample mangrove data, extract original feature quantities, use random window feature perturbation and noise masking on the original feature quantities, and obtain auxiliary data , perform data splicing on the features of the unmasked part to obtain the residual data and record the position information P of the disturbance window;

[0034] s2, construct a learning task model, use multiple multi-view convolution modules MVCCM to build a multi-view convolution neural network MVCNN, extract the deep features of mangroves from the residual data for output; then use multiple one-dimensional convolution and pooling layers to cascade to form a simplified decoder, flatten the deep features, consider the position information of the disturbance window, and splice with the auxiliary data as the input of the simplified decoder to obtain the reconstructed feature data of mangroves and the model loss function representing the reconstruction error of the disturbance feature in stage 1, and adjust the convolution weight and bias of the multi-view convolution module with the loss value of the reconstructed feature;

[0035] s3, a mangrove biomass identifier is constructed through two fully connected layers, and the deep characteristics of the mangrove are input to obtain the estimated biomass of the mangrove; the estimated biomass value is compared with the actual biomass value, the biomass estimation loss of stage 2 is calculated, and the model parameters are fine-tuned with the estimated loss value to obtain the final biomass estimation model.

[0036] The original feature quantities extracted in step s1 include four types: band information, vegetation index, texture feature and pressure height index; the feature quantities included in the band information are coastal band, blue band, green band, red band, near infrared wave, short infrared wave 1 and short infrared wave 2; the feature quantities included in the vegetation index are normalized vegetation index NDVI, ratio vegetation index RVI, difference vegetation index DVI, soil adjusted vegetation index SAVI, enhanced vegetation index EVI and green band normalized difference vegetation index GNDVI; the feature quantities included in the texture feature are variance VAR, homogeneity HOM, contrast CON, heterogeneity DIS, entropy ENT, angular second moment SM, correlation COR and mean ME; the pressure height index is terrain factor TOF.

[0037] The random window feature disturbance sets the window length, randomly selects the number of features corresponding to the window length in each original feature quantity data, and uses Gaussian noise to perform feature masking on the selected features. The expression is:

[0038] , , where P is the randomly selected position information of the disturbance window, It means randomly selecting an integer from a to b, L is the total number of input features, W is the window length, represents the generated auxiliary data, represents Gaussian noise, x\left [ {a,b} \right ] Represents the features of segment a to b of the original feature data, Indicates feature disturbance;

[0039] The expression of data splicing is: ,in, Represents the residual data after splicing, Represents a data concatenation operation.

[0040] After inputting the residual data, the multi-view convolution module first uses small-scale convolution to obtain small-view features using different convolution step sizes. The expression is: ,in, represents the i-th step output of the small-scale convolution, Represents a small-scale convolution with a step size of i;

[0041] Based on the small field of view features, medium-scale convolution is used to obtain the medium field of view features of different branches using different convolution step sizes. The expression is: ,in, represents the i-th step output of the medium-scale convolution, represents a medium-scale convolution with a step size of i;

[0042] Finally, the small field of view features and the medium field of view features are obtained. Based on the cascade convolution mode, large-scale convolution is used to obtain the large field of view features under different convolution step sizes. The expression is: ,in, represents the i-th step output of the large-scale convolution, Represents a large-scale convolution with a step size of i; Represents feature concatenation operation;

[0043] The feature concatenation operation is used to process the large-view features to obtain the deep features output by a single multi-view convolution module, which can be expressed as ,in, Represents a single MVCCM module output feature.

[0044] The multi-view convolutional neural network is constructed by cascading multiple multi-view convolutional modules and pooling operations; the output feature expression of the multi-view convolutional neural network is: ,in, represents the output features of the multi-view convolutional neural network, It represents the cascade operation of n modules, and Pool represents the pooling operation.

[0045] The expression of the input of the simplified decoder is ,in, To simplify the input features of the decoder after concatenation, represents the feature tiling operation; the expression of the simplified decoder is ,in, represents the output features of the decoder; the learning task model sets a fully connected layer to process the output features of the decoder to obtain the reconstructed feature data of the mangrove forest, which is expressed as ,in, Represents the reconstructed feature data, Represents the fully connected layer processing the output features of the decoder.

[0046] The calculation formula of the model loss function is: ,in, represents the stage 1 loss, and They represent the reconstructed value and true value of the ith feature of the jth mangrove sample, respectively, and M is the total number of samples.

[0047] The expression of the biomass identifier is: ,in, represents the estimated mangrove biomass, represents the i-th fully connected layer; the calculation formula for the estimated biomass loss is ,in, represents the estimated biomass loss in phase 2, and represent the estimated value and true value of the biomass of the jth sample respectively.

[0048] Through Landsat8 remote sensing images and DEM elevation data, 22 original feature quantities of band information, vegetation index, texture characteristics and pressure height indicators are extracted; then, for the original feature quantities, some feature quantities are randomly selected to perturb the order and add interference, and the remaining feature quantities are used to guide the reconstruction of the feature quantities before disturbance, so that the characteristics of mangroves are deeply extracted from a small amount of data; then a learning task model is established, and a multi-view convolutional neural network is constructed based on the deep stacking of multiple multi-view cascade convolution modules to extract the deep features of mangroves; finally, a biomass estimator is constructed and the model parameters are fine-tuned to obtain the final model and obtain the accurate estimation value of mangrove biomass under small samples, forming a complete self-supervised perturbation feature reconstruction network; the estimation method based on the self-supervised perturbation feature reconstruction network composed of the above technical scheme is realized, which can circumvent the original data sample size limitation when extracting the deep features of mangroves, solve the problem of data scarcity, improve the generalization of the model, and effectively improve the accuracy of mangrove biomass estimation under small sample conditions.

[0049] like Figure 4 As shown in the figure, by extracting 22 characteristic quantities of band information, vegetation index, texture characteristics and pressure height index, the small sample mangrove biomass data is obtained, making the final model widely applicable to a variety of different plants and having good generalization; Figure 5 As shown in the figure, the data is divided into 8:2, the sample size is small, and the data is scarce, which is in line with the technical problem to be solved; Figure 6 In, through the corresponding Figure 5 The network training process of the sample. The upper and lower sub-figures respectively represent the self-supervised learning loss and the mangrove biomass estimation loss. In the figure, the self-supervised learning loss and the mangrove biomass estimation loss of the model gradually decrease with the increase of the number of iterations.

[0050] like Figure 7 As shown, the original feature quantity data, the feature perturbation data and the reconstructed feature data. Under the influence of the perturbation window, the perturbed feature is significantly disrupted compared to some feature segment data of the original feature quantity; after the feature extraction and decoding operation of step s2, the reconstructed feature data can restore the initial feature quantity to a large extent, and the surface network can effectively learn the deep characteristics of mangroves when data is scarce. The estimation method based on the self-supervised perturbation feature reconstruction network can effectively improve the model feature extraction performance when data is scarce; and according to the mangrove biomass corresponding to the ordinate, the biomass estimation value corresponding to the reconstructed feature can be obtained, which is relatively close to the true value.

Claims

1. A small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network, characterized by: The following steps are included: s1, use Landsat8 remote sensing images and DEM elevation data to process small sample mangrove data, extract original feature quantities, use random window feature perturbation and noise masking on the original feature quantities, and obtain auxiliary data The features of the unmasked part are concatenated to obtain the residual data x res And record the position information P of the disturbance window; s2, build a learning task model, use multiple multi-view convolution modules MVCCM to build a multi-view convolutional neural network MVCNN, and extract deep features of mangroves from residual data O MVCCM Output; then use multiple one-dimensional convolution and pooling layers to form a simplified decoder, flatten the deep features, consider the position information of the disturbance window, and splice it with the auxiliary data as the input of the simplified decoder to obtain the reconstructed feature data of the mangrove And the model loss function representing the reconstruction error of the disturbed features in stage 1, the convolution weights and biases of the multi-view convolution module are adjusted according to the loss value of the reconstructed features; s3, a mangrove biomass identifier is constructed through two fully connected layers, and the deep features of the mangrove are input to obtain the estimated biomass of the mangrove; the estimated biomass value is compared with the actual biomass value, the estimated biomass loss in stage 2 is calculated, and the model parameters are fine-tuned based on the estimated loss value to obtain the final biomass estimation model; The random window feature disturbance sets the window length, randomly selects the number of features corresponding to the window length in each original feature quantity data, and uses Gaussian noise to perform feature masking on the selected features. The expression is: Among them, P is the randomly selected position information of the disturbance window, It means randomly selecting an integer from a to b, L is the total number of input features, W is the window length, represents the generated auxiliary data, represents Gaussian noise, x[a, b] represents the feature from segment a to b of the original feature data, and χ() represents feature disturbance; The expression of data splicing is: Among them, x res Represents the residual data after splicing, Indicates data concatenation operation; After inputting the residual data, the multi-view convolution module first uses small-scale convolution to obtain small-view features using different convolution step sizes. The expression is: in, represents the i-th step output of the small-scale convolution, Represents a small-scale convolution with a step size of i; Based on the small field of view features, medium-scale convolution is used to obtain the medium field of view features of different branches using different convolution step sizes. The expression is: in, represents the i-th step output of the medium-scale convolution, represents a medium-scale convolution with a step size of i; Finally, the small field of view features and the medium field of view features are obtained. Based on the cascade convolution mode, large-scale convolution is used to obtain the large field of view features under different convolution step sizes. The expression is: in, represents the i-th step output of the large-scale convolution, represents a large-scale convolution with a step size of i; Θ{} represents a feature concatenation operation; The feature concatenation operation is used to process the large-view features to obtain the deep features output by a single multi-view convolution module, which can be expressed as Among them, O MVCCM Represents the output characteristics of a single MVCCM module; The multi-view convolutional neural network is constructed by cascading multiple multi-view convolutional modules and pooling operations; the output feature expression of the multi-view convolutional neural network is: MVCNN =Φ n <Pool(MVCCM)> , where O MVCNN represents the output features of the multi-view convolutional neural network, Φ n <> indicates the cascade operation of n modules, and Pool indicates the pooling operation.

2. According to claim 1, a small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network is characterized by: The original feature quantities extracted in step s1 include four types: band information, vegetation index, texture feature and pressure height index; the feature quantities included in the band information are coastal band, blue band, green band, red band, near infrared wave, short infrared wave 1 and short infrared wave 2; the feature quantities included in the vegetation index are normalized vegetation index NDVI, ratio vegetation index RVI, difference vegetation index DVI, soil adjusted vegetation index SAVI, enhanced vegetation index EVI and green band normalized difference vegetation index GNDVI; the feature quantities included in the texture feature are variance VAR, homogeneity HOM, contrast CON, heterogeneity DIS, entropy ENT, angular second moment SM, correlation COR and mean ME; the pressure height index is terrain factor TOF.

3. According to claim 1, a small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network is characterized by: The expression of the input of the simplified decoder is Among them, x de is the input feature of the simplified decoder after splicing, Flat() represents the feature flattening operation; the expression of the simplified decoder is de =Φ n <Pool(Conv(x de ))>, where O de represents the output features of the decoder; the learning task model sets a fully connected layer to process the output features of the decoder to obtain the reconstructed feature data of the mangrove forest, which is expressed as in, Represents the reconstructed feature data, and FC() represents the output features of the decoder processed by the fully connected layer.

4. According to claim 3, a small sample mangrove biomass estimation method based on a self-supervised perturbation feature reconstruction network is characterized by: The calculation formula of the model loss function is: Among them, L s1 represents the stage 1 loss, With x j (i′) represents the reconstructed value and true value of the i′th feature of the jth mangrove sample, respectively, and M is the total number of samples.

5. According to claim 1, a small sample mangrove biomass estimation method based on self-supervised perturbation feature reconstruction network is characterized by: The expression of the biomass identifier is: in, represents the estimated biomass of mangroves, FC i" represents the i"th fully connected layer; the calculation formula for the estimated biomass loss is Among them, L s2 represents the estimated biomass loss in phase 2, With y j represent the estimated value and true value of the biomass of the jth sample respectively, and M is the total number of samples.

Images