Mangrove forest distribution diagram generation method based on deep learning
By constructing the SLCNet model, combining GF2 remote sensing satellite optical data and Labelme software annotation, the problem of insufficient processing efficiency and accuracy of mangrove distribution maps in the existing technology is solved, and more efficient and accurate generation of mangrove distribution maps is achieved.
Patent Information
- Application Number
- CN202510102505.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-05-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing deep learning technology is not mature enough in terms of processing and confirmation efficiency and accuracy of mangrove distribution maps, and it is difficult to effectively extract data in images and improve the confirmation efficiency and accuracy of mangrove distribution maps.
Using a deep learning-based mangrove distribution map generation method, mangrove high-precision drawing using GF2 remote sensing satellite optical data, combined with Labelme software for annotation, the SLCNet model is constructed. This model includes HWD hawavelet downsampling and SimConv convolution, which is used for segmentation and feature extraction of mangrove areas.
It significantly improves the confirmation efficiency and accuracy of mangrove distribution maps, reduces the fragmentation of output results, enhances the universality and generalization ability of mapping results, and can adapt to a wider range of regional characteristics and ecological environment conditions.
Smart Images
Figure CN119992332A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deep learning and image processing, and in particular to a method for generating a mangrove distribution map based on deep learning. Background Art
[0002] Mangroves play a vital role in maintaining the safety of coastlines, promoting sediment accumulation, stabilizing beach morphology and maintaining the overall structure of coastlines. They are not only a natural barrier against wind and wave attacks, but also an important part of the blue carbon ecosystem, providing indispensable carbon storage services. According to statistics, the total coverage area of mangroves in the world is about 14.8 million hectares, most of which are distributed in tropical and subtropical coastal areas, especially in South Asia and Southeast Asia. The mangroves in these two regions together account for 80% of the global total area. However, with the rapid increase in population, the acceleration of urbanization and the continuous deepening of resource development, land use in coastal areas has undergone earth-shaking changes, posing an increasingly severe threat to mangroves, a key ecosystem, especially in Asia, Latin America and Africa. A large amount of impermeable surface covers the original mangrove areas, which not only leads to the rapid degradation of the ecosystem, but also further increases the risk of sea level rise, highlighting the urgency of accurately and efficiently mapping the distribution of mangroves to support environmental monitoring and promote sustainable development. Drawing detailed mangrove maps is of great significance for timely identifying threatened areas, formulating effective protection measures and promoting the harmonious coexistence of ecological balance and economic development.
[0003] Existing methods for obtaining mangrove distribution maps often involve neural network deep learning technology, spatial analysis and data processing technology; among them, deep learning technology used in mangrove distribution maps often involves image segmentation, target detection and feature extraction; but the existing methods for processing mangrove distribution maps through deep learning technology are not yet mature, mainly involving how to extract effective data from images for deep learning model training and how to build new deep learning models to improve the confirmation efficiency and accuracy of mangrove distribution maps. Summary of the invention
[0004] The purpose of the present invention is to overcome the above problems existing in the prior art and greatly improve its technical effect on the basis of the original technology; to this end, the present invention provides a method for generating a mangrove distribution map based on deep learning, the method comprising:
[0005] First, we used GF2 remote sensing satellite optical data to carry out high-precision mapping of mangroves and obtained multiple high-precision distribution maps of mangroves;
[0006] The obtained high-precision distribution maps of mangroves were annotated by Labelme software, and the distribution areas of mangroves were annotated by polygons. i In addition, the high-precision images of multiple mangroves were annotated and segmented using annotation software. The segmentation points that were mangroves were marked as 1, and the segmentation points that were not mangroves were marked as 0. All 1s and 0s were combined into a matrix Y i ; Combined with X i Composition data set {(X i , Y i )}, where i represents the i-th high-precision distribution map of mangroves;
[0007] The data set {(X i , Y i )} is divided into a training set, a validation set and a test set. The SLCNet model is trained through the training set. The generalization ability of the SLCNet model during the training process is evaluated through the validation set to ensure that the model will not overfit. The performance of the SLCNet model is finally evaluated through the test set, and it is determined whether the model can perform well on unseen data to obtain a good SLCNet model. If the test effect is good, the final SLCNet model is obtained; the SLCNet model is a similar local convolutional network, and the SLCNet model includes: an SLCNet encoder and an SLCNet decoder;
[0008] If it is necessary to obtain the mangrove distribution map in the image to be processed, the image to be processed is used as the input of the final SLCNet model, and the output is a matrix corresponding to the image to be processed; the elements in the matrix are mapped to the image according to the method of mangrove area segmentation based on multiple high-precision maps of mangroves; the elements in the matrix that are 1 represent that the mapped image area is a mangrove, and the elements that are 0 in the matrix represent that the mapped image area is not a mangrove, that is, the mangrove distribution map of the image to be processed is obtained.
[0009] Secondly, the method of obtaining a plurality of high-precision distribution maps of mangroves by using GF2 remote sensing satellite optical data for high-precision mapping of mangroves includes: (11) acquiring GF2 remote sensing satellite optical data; (12) preprocessing the acquired GF2 remote sensing satellite optical data; (13) extracting and classifying mangrove features using remote sensing satellite optical data by methods including spectral feature classification and machine learning; (14) generating a plurality of mangrove distribution maps; and (15) mapping the extracted mangrove distribution information by using GIS software to obtain a plurality of high-precision distribution maps of mangroves.
[0010] Secondly, the method of using the labeling software to label and segment the mangrove area of the plurality of high-precision maps of mangroves includes: (21) using the GIMP software to evenly divide each high-precision map of mangroves into a plurality of small squares of equal area in a horizontal and vertical manner; (22) locating the intersections of all the squares and identifying whether the classification at all the intersections is mangroves, marking the mangroves as element 1 of the matrix, and marking the non-mangroves as element 0 of the matrix; (23) taking one intersection of the squares as one element of the matrix, and forming a matrix Y with all 1s and 0s. i , complete the labeling and segmentation of mangrove areas.
[0011] Secondly, the SLCNet encoder includes: HWD Haar wavelet downsampling and SimConv convolution; wherein, HWD Haar wavelet downsampling is a signal downsampling method based on Haar wavelet transform, which is commonly used in signal processing and image processing. HWD Haar wavelet downsampling is used to improve the performance of semantic segmentation models. By incorporating wavelet transform into the downsampling process, the image is decomposed into multiple frequency bands, thereby retaining more edge and detail information; SimConv convolution includes: first, generating weights and local window sizes according to the input image, using the standard kernel K1 for convolution operation, and generating initial feature output; then, calculating the difference between each row and each column of the feature map along the X-axis and Y-axis, and calculating the T ratio by setting the threshold to 0.1. The T ratio refers to treating the result less than the threshold as the correct T , the proportion of correct results; if the T ratio is greater than 0.4, the K2 kernel is selected for convolution; if the T ratio is less than 0.4, the K3 kernel is used; finally, the cosine similarity threshold is set to 0.15, and for two feature vectors generated by the selected convolution kernel and the standard convolution kernel K1 respectively, features are selected according to the difference in cosine similarity between them: if the difference is greater than 0.15 and is positive, the feature generated by the selected convolution kernel is selected and weighted; if the difference is negative, the feature generated by the standard convolution kernel K1 is selected and weighted; if the difference is ≤0.15, the average of the two is taken and weighted; the cosine similarity is a measure of the similarity between two vectors.
[0012] Secondly, the SLCNet decoder includes: a CatBlock module and an UpBlock module; the CatBlock module is a CatBlock feature fusion module; in this module, the upsampled feature map is first processed by HWD Ha wavelet downsampling to effectively retain deep and subtle feature details; then, a fully connected operation is applied to enhance the interpretability of the model; finally, a convolution operation is used to maintain the consistency of the output channel and optimize the subsequent feature processing flow; the UpBlock module fuses the feature map output by CatBlock with the shallow feature map to reduce the interference of noise on the model; then, it applies upsampling to ensure the consistency of the final multi-layer feature fusion; the robustness of the model is enhanced and the accuracy of segmentation is improved.
[0013] The beneficial effects of the present invention are:
[0014] The present invention proposes a method for generating a mangrove distribution map based on deep learning; the advantages of the present invention are that (31) the SLCNet model used in the present invention can minimize the fragmentation of the output results and ensure that the generated images or data maintain continuity and integrity in space; at the same time, the universality and generalization ability of the mapping results are significantly enhanced, so that it can adapt to a wider range of regional characteristics and ecological environmental conditions; (32) the present invention uses annotation software to annotate and segment the mangrove areas of multiple high-precision maps of mangroves, extracts data sets of multiple high-precision maps of mangroves, and trains the SLCNet model through the data sets. The final model obtained can directly analyze the image to be processed to obtain the matrix corresponding to the image, and the mangrove distribution map is obtained by mapping the matrix to the image to be processed; (33) combined with (31) and (32), the present invention can more accurately draw a mangrove distribution map, providing more solid technical support for the ecological protection and sustainable utilization of mangroves. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a flow chart of a method for generating a mangrove distribution map based on deep learning of the present invention.
[0016] Figure 2 Schematic diagram of the structure of the SLCNet similar local convolutional network.
[0017] Figure 3 Schematic diagram of the structure of SimConv2d two-dimensional parallel convolution. DETAILED DESCRIPTION
[0018] The specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings; it should be understood that the specific embodiments given here are only used to illustrate and explain the present invention and cannot be used to limit the present invention.
[0019] like Figure 1 FIG. 1 is a flowchart of a method for generating a mangrove distribution map based on deep learning according to an embodiment of the present invention, the flowchart comprising: step S100, first, high-precision mapping of mangroves is performed by using GF2 remote sensing satellite optical data to obtain a plurality of high-precision distribution maps of mangroves; step S200, the obtained plurality of high-precision distribution maps of mangroves are annotated by using Labelme software, the mangrove distribution area is annotated by polygons, and the annotated images are used as X i In addition, the high-precision images of multiple mangroves were annotated and segmented using annotation software. The segmentation points that were mangroves were marked as 1, and the segmentation points that were not mangroves were marked as 0. All 1s and 0s were combined into a matrix Y i ; Combined with X i Composition data set {(X i , Y i )}, where i represents the i-th high-precision distribution map of mangroves; Step S300, the data set {(X i , Y i )} is divided into a training set, a validation set and a test set. The SLCNet model is trained by the training set. The generalization ability of the SLCNet model in the training process is evaluated by the validation set to ensure that the model will not overfit. The performance of the SLCNet model is finally evaluated by the test set, and it is judged whether the model can perform well on unseen data to obtain a good SLCNet model; if the test effect is good, the final SLCNet model is obtained; the SLCNet model is a similar local convolutional network, and the SLCNet model includes: an SLCNet encoder and an SLCNet decoder; step S400, if it is necessary to obtain a mangrove distribution map in the image to be processed, the image to be processed is used as the input of the final SLCNet model, and the output is a matrix corresponding to the image to be processed; according to the method of segmenting the mangrove area with high-precision maps of multiple mangroves, the elements in the matrix are mapped to the image; the elements of the matrix that are 1 represent that the mapped image area is a mangrove, and the elements of the matrix that are 0 represent that the mapped image area is not a mangrove, that is, the mangrove distribution map of the image to be processed is obtained.
[0020] Among them, in step S100, the steps of obtaining multiple high-precision distribution maps of mangroves by using GF2 remote sensing satellite optical data for high-precision mapping of mangroves are: (11) obtaining GF2 remote sensing satellite optical data; (12) preprocessing the obtained GF2 remote sensing satellite optical data; (13) extracting and classifying mangrove features using remote sensing satellite optical data through methods including: spectral feature classification and machine learning; (14) generating multiple mangrove distribution maps; (15) mapping the extracted mangrove distribution information through GIS software to obtain multiple high-precision distribution maps of mangroves.
[0021] Among them, in step S200, the steps of annotating and segmenting the mangrove area of the obtained multiple high-precision maps of mangroves by using the annotation software are as follows: (21) using GIMP software to evenly divide each high-precision map of mangroves into multiple small squares of equal area in a horizontal and vertical manner; (22) locating the intersections of all the squares, and identifying whether the classification at all the intersections is mangroves, marking the mangroves as element 1 of the matrix, and marking the non-mangroves as element 0 of the matrix; (23) taking one intersection of the squares as one element of the matrix, all 1s and 0s are combined into a matrix Y i , complete the labeling and segmentation of mangrove areas.
[0022] Among them, in step S300, the SLCNet encoder includes: HWD Haar wavelet downsampling and SimConv convolution; HWD Haar wavelet downsampling is a signal downsampling method based on Haar wavelet transform, which is commonly used in signal processing and image processing. HWD Haar wavelet downsampling is used to improve the performance of semantic segmentation models. By incorporating wavelet transform into the downsampling process, the image is decomposed into multiple frequency bands, thereby retaining more edge and detail information; SimConv convolution includes: first, generating weights and local window sizes according to the input image, using the standard kernel K1 for convolution operation, and generating initial feature output; then, calculating the difference of each row and column of the feature map along the X-axis and Y-axis, and calculating T by setting the threshold to 0.1. Ratio, the T ratio refers to the proportion of correct results that are less than the threshold value as correct T; if the T ratio is greater than 0.4, the K2 kernel is selected for convolution; if the T ratio is less than 0.4, the K3 kernel is used; finally, the cosine similarity threshold is set to 0.15, and for two feature vectors generated by the selected convolution kernel and the standard convolution kernel K1 respectively, features are selected according to the difference in cosine similarity between them: if the difference is greater than 0.15 and is positive, the feature generated by the selected convolution kernel is selected and weighted; if the difference is negative, the feature generated by the standard convolution kernel K1 is selected and weighted; if the difference is ≤0.15, the average of the two is taken and weighted; the cosine similarity is a measurement method for measuring the similarity between two vectors.
[0023] In the above embodiment, the SLCNet decoder includes: a CatBlock module and an UpBlock module; the CatBlock module is a CatBlock feature fusion module; in this module, the upsampled feature map is first processed by HWD Ha wavelet downsampling to effectively retain deep and subtle feature details; then, a fully connected operation is applied to enhance the interpretability of the model; finally, a convolution operation is used to maintain the consistency of the output channel and optimize the subsequent feature processing flow; the UpBlock module fuses the feature map output by CatBlock with the shallow feature map, thereby reducing the interference of noise on the model; then, it applies upsampling to ensure the consistency of the final multi-layer feature fusion; the robustness of the model is enhanced and the accuracy of segmentation is improved.
[0024] Among them, in step S400, the image to be processed is used as the input of the final SLCNet model, and the output is a matrix corresponding to the image to be processed, and the matrix segmentation method is the same as the image segmentation method in step S200; therefore, according to the image segmentation method of step S200, the elements in the matrix are mapped to the image; at this time, the elements of 1 in the matrix represent that the mapped image area is a mangrove, and the elements of 0 in the matrix represent that the mapped image area is not a mangrove, and the targets in the image are marked with 1 and 0, that is, the mangrove distribution map of the image to be processed is obtained; it should be pointed out that, by the above segmentation method, the more small squares are divided, the more accurate the mangrove distribution map obtained by the present invention will be.
[0025] like Figure 2 As shown in FIG. 1 , it is a schematic diagram of the structure of the SLCNet similar local convolutional network of the present invention; in the schematic diagram, X is first i The input is processed by the SimBlock synchronization block. After the synchronization is fast processed, the processing result is passed to the CatBlock splicing block for processing. After the splicing block is processed, the processed information is passed to the UpBlock upsampling block. After the upsampling block performs the Upsample upsampling operation on the information, the processed information is concentrated and passed to the OutBlock output block. The output block generates the output Y i , get the distribution map of mangroves; the specific operation process of CatBlock splicing block, UpBlock upsampling block and Out Block output block Figure 2 There are specific processes in , Input is input, Upsample is upsampling, H WD represents Ha wavelet downsampling, represents a fully connected operation, conv+SILU represents a combination of a convolution operation and an activation function, and BN represents batch normalization.
[0026] like Figure 3As shown in the figure, it is a structural schematic diagram of the SimConv2d two-dimensional parallel convolution of the present invention; in the schematic diagram, the image with a width and height of W*H is first segmented into X and Y, and the weight weight, convolution kernel K1 and the selected convolution kernel are respectively operated on the segmented image at the same time; the operation of the selected convolution kernel is to select k2 or k3 according to the setting of the condition to perform the convolution operation on the segmented image; by setting the threshold threshold to 0.1, the T ratio is calculated, and the T ratio refers to treating the result less than the threshold as the correct T, and treating the result not less than the threshold as the wrong F, and the ratio of T to T+F is the T ratio; if the T ratio is greater than the ratio Ratio (set to 0.4), the K2 kernel is selected for convolution; if the T ratio is less than 0. 4, then use the K3 kernel for convolution; finally, calculate the cosine similarity of the two feature vectors obtained by operating the segmented image with the standard convolution kernel k1 and the selected convolution kernel, and perform subsequent operations according to the set rules; the set rules are: set the cosine similarity threshold to 0.15, and select features based on the difference in cosine similarity between the two feature vectors: if the difference is greater than 0.15 and is positive, select the feature generated by the selected convolution kernel and weight it with the output generated by the weight weight; if the difference is negative, select the feature generated by the standard convolution kernel K1 and weight it with the output generated by the weight weight; if the difference is ≤0.15, take the average of the two and weight it with the output generated by the weight weight.
Claims
1. A method for generating a mangrove distribution map based on deep learning, characterized in that: The method comprises the following steps: First, we used GF2 remote sensing satellite optical data to carry out high-precision mapping of mangroves and obtained multiple high-precision distribution maps of mangroves; The obtained high-precision distribution maps of mangroves were annotated by Labelme software, and the distribution areas of mangroves were annotated by polygons. i In addition, the high-precision images of multiple mangroves were annotated and segmented using annotation software. The segmentation points that were mangroves were marked as 1, and the segmentation points that were not mangroves were marked as 0. All 1s and 0s were combined into a matrix Y i ; Combined with X i Composition data set {(X i , Y i )}, where i represents the i-th high-precision distribution map of mangroves; The data set {(X i , Y i )} is divided into a training set, a validation set and a test set. The SLCNet model is trained through the training set. The generalization ability of the SLCNet model during the training process is evaluated through the validation set to ensure that the model will not overfit. The performance of the SLCNet model is finally evaluated through the test set, and it is determined whether the model can perform well on unseen data to obtain a good SLCNet model. If the test effect is good, the final SLCNet model is obtained; the SLCNet model is a similar local convolutional network, and the SLCNet model includes: an SLCNet encoder and an SLCNet decoder; If it is necessary to obtain the mangrove distribution map in the image to be processed, the image to be processed is used as the input of the final SLCNet model, and the output is a matrix corresponding to the image to be processed; the elements in the matrix are mapped to the image according to the method of mangrove area segmentation based on multiple high-precision maps of mangroves; the elements in the matrix that are 1 represent that the mapped image area is a mangrove, and the elements that are 0 in the matrix represent that the mapped image area is not a mangrove, that is, the mangrove distribution map of the image to be processed is obtained.
2. A method for generating a mangrove distribution map based on deep learning according to claim 1, characterized in that: The method of obtaining a plurality of high-precision distribution maps of mangroves by using GF2 remote sensing satellite optical data for high-precision mapping of mangroves comprises: (11) acquiring GF2 remote sensing satellite optical data; (12) preprocessing the acquired GF2 remote sensing satellite optical data; (13) extracting and classifying mangrove features using remote sensing satellite optical data by means of spectral feature classification and machine learning methods; (14) generating a plurality of mangrove distribution maps; and (15) mapping the extracted mangrove distribution information by means of GIS software to obtain a plurality of high-precision distribution maps of mangroves.
3. A method for generating a mangrove distribution map based on deep learning according to claim 2, characterized in that: The method of labeling and segmenting the mangrove area of the plurality of high-precision maps of mangroves obtained by using labeling software includes: (21) using GIMP software to evenly divide each high-precision map of mangroves into a plurality of small squares of equal area in a horizontal and vertical manner; (22) locating the intersections of all the squares and identifying whether the classification at all the intersections is mangroves, marking the mangroves as elements 1 of the matrix, and marking the non-mangroves as elements 0 of the matrix; (23) taking one intersection of the squares as one element of the matrix, and forming a matrix Y with all 1s and 0s. i , complete the labeling and segmentation of mangrove areas.
4. The method for generating a mangrove distribution map based on deep learning according to claim 1, characterized in that: The SLCNet encoder includes: HWD Haar wavelet downsampling and SimConv convolution; wherein, HWD Haar wavelet downsampling is a signal downsampling method based on Haar wavelet transform, which is commonly used in the fields of signal processing and image processing. HWD Haar wavelet downsampling is used to improve the performance of semantic segmentation models. By incorporating wavelet transform into the downsampling process, the image is decomposed into multiple frequency bands, thereby retaining more edge and detail information; SimConv convolution includes: firstly, generating weights and local window sizes according to the input image, using the standard kernel K1 for convolution operation, and generating initial feature output; then, calculating the difference of each row and column of the feature map along the X-axis and the Y-axis, and calculating the T ratio by setting the threshold to 0.1, The T ratio refers to the proportion of correct results that are less than the threshold value and are considered as correct T; if the T ratio is greater than 0.4, the K2 kernel is selected for convolution; if the T ratio is less than 0.4, the K3 kernel is used; finally, the cosine similarity threshold is set to 0.15, and for two feature vectors generated by the selected convolution kernel and the standard convolution kernel K1 respectively, features are selected according to the difference in cosine similarity between them: if the difference is greater than 0.15 and is positive, the feature generated by the selected convolution kernel is selected and weighted; if the difference is negative, the feature generated by the standard convolution kernel K1 is selected and weighted; if the difference is ≤0.15, the average of the two is taken and weighted; the cosine similarity is a measurement method for measuring the similarity between two vectors.
5. The method for generating a mangrove distribution map based on deep learning according to claim 1, characterized in that: The SLCNet decoder includes: a CatBlock module and an UpBlock module; the CatBlock module is a CatBlock feature fusion module; in this module, the upsampled feature map is first processed by HWD Ha wavelet downsampling to effectively retain deep and subtle feature details; then, a fully connected operation is applied to enhance the interpretability of the model; finally, a convolution operation is used to maintain the consistency of the output channel and optimize the subsequent feature processing flow; the UpBlock module fuses the feature map output by CatBlock with the shallow feature map to reduce the interference of noise on the model; then, it applies upsampling to ensure the consistency of the final multi-layer feature fusion; the robustness of the model is enhanced and the accuracy of segmentation is improved.