Mangrove forest distribution analysis method based on deep learning

Abstract

The invention provides a mangrove forest distribution analysis method based on deep learning, and relates to the technical field of mangrove forest distribution monitoring, and the method comprises the steps: obtaining original remote sensing data, and converting the original remote sensing data into surface reflectance data; constructing a mangrove forest coarse extraction decision tree model MREDT, and processing the surface reflectance data through the MREDT; carrying out filtering processing on noise and edge information of data generated by MREDT through a morphological algorithm; and training the DACN-M model by using the processed data, and improving the segmentation precision of the mangrove forest region by using a residual superposition method on the basis of the DACN-M model. The method solves the problem of dependence on a large number of annotation samples in a traditional method, and has good generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of mangrove distribution monitoring, and in particular to a mangrove distribution analysis method based on deep learning. Background Art

[0002] Mangroves have important economic and ecological values, but they are threatened by various factors and face the risk of disappearing. Remote sensing and artificial intelligence technologies can help monitor and identify the distribution and changes of mangroves. However, mangrove-covered areas are often difficult to access, which makes field sampling difficult. At the same time, the analysis of mangrove distribution through artificial intelligence technology requires a large number of annotated samples, that is, traditional methods rely on a large amount of manually annotated data, which makes it difficult to identify and monitor mangroves when field sampling is inconvenient. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a mangrove distribution analysis method based on deep learning to solve the problem of traditional methods' dependence on a large number of annotated samples.

[0004] To achieve the above object, the present invention provides a mangrove distribution analysis method based on deep learning, the method comprising:

[0005] Obtaining raw remote sensing data, and converting the raw remote sensing data into surface reflectance data;

[0006] A mangrove rough extraction decision tree model MREDT was constructed and the surface reflectance data were processed through MREDT.

[0007] The noise and edge information of the data generated by MREDT are filtered through morphological algorithms;

[0008] The processed data were used to train the DACN-M model. Based on the DACN-M model, the residual superposition method was used to improve the segmentation accuracy of the mangrove area.

[0009] Furthermore, the original remote sensing data is 1C level pre-processed data with a cloud coverage rate of less than 10%. The original remote sensing data is converted into 2A level surface reflectance data through orthorectification and atmospheric correction. The spatial resolution of the surface reflectance data in the horizontal direction is 20 meters, and the surface reflectance data includes several different spectral bands.

[0010] Furthermore, after obtaining the original remote sensing data, the z-score is used to measure the difference between each original remote sensing data point and the entire original remote sensing data set, and the high-dimensional outliers are removed to filter the abnormal pixels.

[0011] Furthermore, a mangrove rough extraction decision tree model MREDT is constructed, which specifically includes:

[0012] Selecting representative characteristic frequency bands from different frequency bands;

[0013] The linear regression model was used to construct the relevant parameters of the selected characteristic frequency bands, and the mangrove characteristics were analyzed by the following formula:

[0014]

[0015] Among them, ρλ represents the reflectivity at the center of the band, λ 1 , 2 , 3 and λ 4 represents the central wavelength of the selected characteristic frequency band, and f(x) represents the decision feature of mangroves;

[0016] The normalized water vegetation index NDWI is introduced to complete the filtering of water body information, which is expressed as:

[0017]

[0018] Among them, G represents the green band, NIR represents the near-infrared band, f(x) and NDWI together constitute the decision scheme of MREDT, and the strongly correlated pixels of mangroves are collected from remote sensing spectral information through decision conversion.

[0019] Furthermore, the DACN-M model includes a feature dimension increase part and a feature dimension reduction part. The stage containing less channel information is called the low feature stage, and the stage containing more channel information is called the high feature stage. Assume that the input feature matrix is The convolution kernel is The convolution operation is performed by sliding on the two-dimensional data, and its mathematical expression is as follows:

[0020]

[0021] Among them, X*K represents the convolution operation, and (i,j) is the position of the output feature matrix.

[0022] Furthermore, the feature dimension increase part includes four convolution groups and a channel attention module. The size of each convolution group is 1×3. The DACN-M model retains the results in the calculation process from the low feature stage to the high feature stage, and superimposes the results on the features of the same size in the feature dimension reduction part. The channel attention module includes a global pool and a multi-layer perceptron. The channel attention module is applied to the high feature stage to emphasize or reduce the importance of different input channel features and enhance the model's understanding of key features. Its mathematical expression is:

[0023] X′=X⊙(σ(W 2 (ReLU(W 1 (Maxpool(X))+b1 ))+b 2 ) T )

[0024] in, represents the weight matrix of the fully connected layer, represents the bias, Maxpool represents the global maximum pooling operation, σ represents the sigmoid activation function, and the channel attention module learns from the input feature matrix X and converts it into a weighted feature matrix X′.

[0025]

[0026] Furthermore, the feature dimension reduction part includes three convolution groups with a scale of 1×3, which are used to gradually reduce the dimension of the feature to obtain the final regression result.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The present invention provides a mangrove distribution analysis method based on deep learning, proposes a mangrove rough extraction decision tree model MREDT based on multi-linear regression, which is used for the automatic collection of mangrove samples, replacing the traditional method that requires a large amount of manual annotation. At the same time, a deep attention convolutional network model DACN-M is designed for spectral multi-feature fusion in mangrove analysis. The method can effectively distinguish mangroves from other vegetation by using the residual results of threshold segmentation superposition, so that the mangrove area can be finely extracted, which significantly improves the classification accuracy of the method. The method solves the problem of dependence on a large number of annotated samples in traditional methods and has good generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 It is a schematic diagram of the overall process of a mangrove distribution analysis method based on deep learning provided in an embodiment of the present invention.

[0031] Figure 2 It is a schematic diagram of sampling of multiple mangrove research areas provided by the experimental examples of the present invention.

[0032] Figure 3 It is a schematic diagram of the classification accuracy evaluation results provided by the experimental example of the present invention. DETAILED DESCRIPTION

[0033] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0034] Example

[0035] Reference Figure 1 , this embodiment provides a mangrove distribution analysis method based on deep learning, the method comprising the following steps:

[0036] S11. Acquire original remote sensing data, and convert the original remote sensing data into surface reflectivity data.

[0037] In this embodiment, the original remote sensing data can be obtained by the MSI (multi-spectral instrument) sensor on the satellite. Exemplarily, the original remote sensing data is 1C-level pre-processed data with a cloud coverage rate of less than 10%, which is processed by orthorectification and atmospheric correction algorithms and converted into 2A-level surface reflectivity data. The spatial resolution of the surface emissivity data in the horizontal direction is 20 meters, and the surface reflectivity data includes several different spectral bands. In some embodiments, the surface reflectivity data includes 10 different spectral bands: blue (B1443nm), green (B2490nm), red (B3560nm), red edge (B4665nm), near infrared (B5705nm), near infrared (B6740nm), near infrared (B7783nm), near near infrared (B8A865nm), short-wave infrared (B11610nm) and short-wave infrared (B12 2190nm). During the field survey, the boundaries of the mangrove area were demarcated and the water bodies and intertidal zones within the mangrove growth area were marked. This approach allowed the identification of representative sample pixels within the remote sensing spectral information.

[0038] The original remote sensing data includes human error, measurement error and spatial variability error. During the sample collection process, these error outliers, such as exposed soil patches or water bodies, are mistakenly classified as mangrove sample pixels. This embodiment measures the difference between each original remote sensing data point and the entire original remote sensing data set through the z-score, removes high-dimensional outliers, and filters abnormal pixels, thereby generating more accurate and reliable spectral information.

[0039] S12. Construct a mangrove rough extraction decision tree model MREDT, and process the surface reflectance data through MREDT.

[0040] In the construction process of MREDT, the scope of mangroves was first determined on site, and the characteristics of mangroves and non-mangroves were extracted from the original remote sensing data. Then, by analyzing the spectral characteristics of mangroves and non-mangroves, stable and strongly correlated spectral feature information was selected, and the MREDT model was constructed through linear regression analysis.

[0041] As a preferred example, a mangrove rough extraction decision tree model MREDT is constructed, which specifically includes:

[0042] S21. Select a representative characteristic frequency band from different frequency bands.

[0043] The representative characteristic bands selected in this embodiment are B2, B8A, B11 and B12. In order to mine more characteristic information, the residual operator is introduced to obtain the characteristic results of the interaction between spectra. After comparison, it is found that the reflectivity of mangroves is higher in the B8A band, and lower in the low-wavelength band and high-wavelength band. This characteristic is common in vegetation. However, due to the influence of the intertidal zone and the mangroves themselves, the reflectivity value of the B8A mangrove canopy is often higher, and the reflectivity in the high band is lower. This feature can highlight the characteristic information of mangroves. Therefore, we focus on the residual values ​​calculated by B8A and B2 and the residual values ​​obtained by calculating B12 and B8A. And use the natural index conversion of the result to obtain more obvious mangrove characteristics.

[0044] S22. Use the linear regression model to construct the relevant parameters of the selected characteristic frequency bands, and analyze the characteristics of mangroves through the following formula:

[0045]

[0046] Among them, ρλ represents the reflectivity at the center of the band, λ 1 , 2 , 3 and λ 4 represents the central wavelength of the selected characteristic frequency band, and f(x) represents the decision feature of mangroves. Compared with general vegetation indices, this method is more sensitive to the spectral characteristics of mangroves.

[0047] S23, introduce the normalized water vegetation index NDWI to complete the filtering of water body information, expressed as:

[0048]

[0049] Among them, G represents the green band, NIR represents the near infrared band, f(x) and NDWI together constitute the decision scheme of MREDT, and the strongly correlated pixels of mangroves are collected from remote sensing spectral information through decision conversion. The above method is based on vegetation feature analysis and can run stably under different lighting and backgrounds in the study area, thereby ensuring that downstream tasks obtain stable mangrove samples.

[0050] S13. Filter the noise and edge information of the data generated by MREDT using a morphological algorithm.

[0051] In the field of computer image processing, morphological erosion is used to process structural elements in digital images, thereby refining target features in remote sensing images. Morphological erosion helps to separate and distinguish adjacent objects or elements in the image, which helps to analyze and identify them. Defined by set theory, where A represents the set in the binary image and B represents the structural element (also a set), morphological erosion can be defined as the result of A being eroded by B, expressed as

[0052]

[0053] Where E represents the Euclidean space, z is the unknown pixel in E, (B) z is the translation of the structure pixel B relative to the pixel position z, and Represents a subset, and when the structure element B is completely contained in the set A, the pixel position z will be included in the erosion result.

[0054] S14. Use the processed data to train the DACN-M model. Based on the DACN-M model, use the residual superposition method to improve the segmentation accuracy of the mangrove area.

[0055] In this embodiment, the DACN-M model is a fully convolutional neural network that combines the channel attention mechanism and MREDT-based data processing to achieve deep mining of mangrove feature information. The main architecture of the DACN-M model is divided into two core parts: feature dimension increase and feature dimension reduction. We call the stage containing less channel information the low-feature stage (the step close to the network input and output), and the step containing more channel information the high-feature stage.

[0056] Assume that the input feature matrix is The convolution kernel is The convolution operation is performed by sliding on the two-dimensional data, and its mathematical expression is as follows:

[0057]

[0058] Where X*K represents the convolution operation and (i,j) is the position of the output feature matrix. It is assumed that appropriate padding has been applied to the input feature matrix X.

[0059] In this embodiment, the feature dimension increase part uses convolution operations to promote the interaction between input channels, so as to mine more channel feature expression patterns. The feature dimension increase part includes four convolution groups and a channel attention module. The scale of each convolution group is 1×3. The DACN-M model retains the results during the calculation process from the low feature stage to the high feature stage, and superimposes the results on the features of the same size in the feature dimension reduction part. The channel attention module includes a global pool and a multi-layer perceptron. The channel attention module is applied to the high feature stage to emphasize or reduce the importance of different input channel features, enhance the model's understanding of key features, and improve the stability and accuracy of the model. Its mathematical expression is:

[0060] X′=X⊙(σ(W 2 (ReLU(W 1 (Maxpool(X))+b 1 ))+b 2 ) T )

[0061] in, represents the weight matrix of the fully connected layer, represents the bias, Maxpool represents the global maximum pooling operation, σ represents the sigmoid activation function, and the channel attention module learns from the input feature matrix X and converts it into a weighted feature matrix X′.

[0062]

[0063] The feature dimension reduction part includes three convolution groups with a scale of 1×3, which are used to gradually reduce the dimension of the feature to obtain the final regression result. In the process from the high feature stage to the low feature stage, the feature dimension increase part provides the semantics of the original feature, ensuring the stability of the feature regression in the final stage of the DACN-M model regression.

[0064] As an optional implementation, the calculation results of DACN-M at different stages are activated by the rectified linear unit (ReLU) nonlinear mapping method. Its calculation method is simple, but has gradient stability and enhances the nonlinear fitting effect of the model. Its mathematical expression is:

[0065]

[0066] As an optional implementation, the DACN-M model uses the strategy of superimposed residuals to further enhance the crown characteristics of mangroves. Usually, model fitting requires a certain number of negative samples to enhance the generalization ability of the model (such as semi-mangrove windbreaks, which are easily confused with mangrove plants). However, this will further increase the workload, and sample selection will directly affect the regression performance of the model. The "superimposed residual results using threshold segmentation" method effectively solves this problem. When the DACN-M model is fitted to each spectral feature of mangroves, oscillations and errors of other targets may occur. However, the errors of the spectral features of mangroves are relatively small. Therefore, in the superimposed residuals, the errors of other targets will be further amplified. By iteratively finding the optimal threshold for binary segmentation, the model can obtain accurate mangrove classification results.

[0067] Experimental example

[0068] This experiment uses NVIDIA RTX 4090 GPU and PyTorch (GPU version) to train the deep learning model. The initial learning rate is set to 0.001, and it is dynamically adjusted based on this. The mean square error (MSE) is used to calculate the fitting loss of DACN-M.

[0069] Reference Figure 2 ,In order to compare the effectiveness of mangrove monitoring methods, an equidistant sampling method based on image analysis was used. For each study area, a validation point set was created by extracting a sample point every three pixels. The mangrove areas were manually labeled through visual interpretation. The visual interpretation was based on high spatial resolution Google phase images and attribute labeling combined with research results and field surveys of related studies.

[0070] The set of red points represents non-mangrove crowns, while the set of white points represents mangrove crowns. This sampling method effectively ensures that the class distribution of the verification points is consistent with the original image. It also ensures the randomness and independence of the verification points, which is conducive to evaluating the accuracy of different monitoring methods in mangrove identification. In all research areas, we compared the prediction performance of KNN, RF, SVM and DACN-M, and validated DACN-M. In addition, KNN, RF and SVM require labeled samples for model training, while DACN-M benefits from MREDT, which replaces the need for data collection. Therefore, we start from Figure 2 500 positive samples and 1000 negative samples were extracted from the annotated data in for KNN, RF and SVM model training.

[0071] Figure 3The results in the paper show that DACN-M can reveal the semantic differences between mangrove crowns and non-mangrove crowns through multi-level feature extraction and fusion of contextual information. DACN-M shows absolute advantages and achieves excellent results in terms of precision, recall, F1 score and IoU evaluation indicators. The precision values ​​of all categories are greater than 85% and the IoU exceeds 70%, indicating a high consistency between the annotated ground truth and the detection results obtained by DACN-M.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A mangrove distribution analysis method based on deep learning, characterized in that: The method comprises: Obtaining raw remote sensing data, and converting the raw remote sensing data into surface reflectance data; A mangrove rough extraction decision tree model MREDT was constructed and the surface reflectance data were processed through MREDT. The noise and edge information of the data generated by MREDT are filtered through morphological algorithms; The processed data were used to train the DACN-M model. Based on the DACN-M model, the residual superposition method was used to improve the segmentation accuracy of the mangrove area.

2. A mangrove distribution analysis method based on deep learning according to claim 1, characterized in that: The original remote sensing data is 1C level pre-processed data with cloud coverage less than 10%. The original remote sensing data is converted into 2A level surface reflectance data through orthorectification and atmospheric correction. The spatial resolution of the surface reflectance data in the horizontal direction is 20 meters, and the surface reflectance data includes several different spectral bands.

3. The method for analyzing mangrove distribution based on deep learning according to claim 1, characterized in that: After obtaining the original remote sensing data, the z-score is used to measure the difference between each original remote sensing data point and the entire original remote sensing data set, and the high-dimensional outliers are removed and the abnormal pixels are filtered.

4. The method for analyzing mangrove distribution based on deep learning according to claim 1, characterized in that: Construct a mangrove rough extraction decision tree model MREDT, which includes: Selecting representative characteristic frequency bands from different frequency bands; The linear regression model was used to construct the relevant parameters of the selected characteristic frequency bands, and the mangrove characteristics were analyzed by the following formula: Among them, ρλ represents the central reflectivity of the band, λ1, λ2, λ3 and λ4 represent the central wavelengths of the selected characteristic bands, and f(x) represents the decision feature of mangroves; The normalized water vegetation index NDWI is introduced to complete the filtering of water body information, which is expressed as: Among them, G represents the green band, NIR represents the near-infrared band, f(x) and NDWI together constitute the decision scheme of MREDT, and the strongly correlated pixels of mangroves are collected from remote sensing spectral information through decision conversion.

5. The method for analyzing mangrove distribution based on deep learning according to claim 1, characterized in that: The DACN-M model includes a feature dimension increase part and a feature dimension reduction part. The stage containing less channel information is called the low feature stage, and the stage containing more channel information is called the high feature stage. Assume that the input feature matrix is The convolution kernel is The convolution operation is performed by sliding on the two-dimensional data, and its mathematical expression is as follows: Among them, X*K represents the convolution operation, and (i,j) is the position of the output feature matrix.

6. A method for analyzing mangrove distribution based on deep learning according to claim 5, characterized in that: The feature dimension increase part includes four convolution groups and a channel attention module. The scale of each convolution group is 1×3. The DACN-M model retains the results in the calculation process from the low feature stage to the high feature stage, and superimposes the results on the features of the same size in the feature dimension reduction part. The channel attention module includes a global pool and a multi-layer perceptron. The channel attention module is applied to the high feature stage to emphasize or reduce the importance of different input channel features and enhance the model's understanding of key features. Its mathematical expression is: X′=X⊙(σ(W2·(ReLU(W1·(Maxpool(X))+b1))+b2) T ) in, represents the weight matrix of the fully connected layer, represents the bias, Maxpool represents the global maximum pooling operation, σ represents the sigmoid activation function, and the channel attention module learns from the input feature matrix X and converts it into a weighted feature matrix X′.

7. The method for analyzing mangrove distribution based on deep learning according to claim 5, characterized in that: The feature dimension reduction part includes three convolution groups with a scale of 1×3, which are used to gradually reduce the dimension of the feature to obtain the final regression result.

Images