A mangrove super-resolution mapping method based on spatio-temporal spectral feature deep extraction

Abstract

The application discloses a mangrove super-resolution mapping method based on spatio-temporal spectral feature deep extraction, and belongs to the technical field of remote sensing image intelligent processing. The method comprises the following steps: S1, acquiring multi-source remote sensing image data in a research area, and preprocessing the multi-source remote sensing image data; S2, performing mangrove classification and extraction on the preprocessed data based on an object-oriented method; S3, extracting a mangrove category map based on an SVR algorithm; S4, constructing a spatio-temporal matching super-resolution mapping model training data set based on a target phase mangrove abundance map; S5, constructing a spatio-temporal super-resolution mapping network framework combined with an encoding-decoding and VSS module based on the spatio-temporal matching super-resolution mapping model training data set; S6, training and verifying the spatio-temporal super-resolution mapping network framework combined with the encoding-decoding and VSS module; and S7, performing batch spatio-temporal super-resolution mapping based on an optimized model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent processing of remote sensing images, in particular to a mangrove super-resolution mapping method based on spatio-temporal spectral feature deep extraction. BACKGROUND

[0002] As an important part of the intertidal zone ecosystem, mangroves have significant ecological, social and economic values. However, due to human activities and climate change, two-thirds of the global mangrove forests have been lost, and the destruction caused by natural factors is still intensifying. Remote sensing technology, as an effective means of dynamic monitoring of mangroves, provides broad coverage and short-period data support. However, traditional methods are easily affected by the mixed pixel problem, especially in the narrow and broken distribution areas of mangroves, where the accuracy is significantly reduced. Although super-resolution mapping technology can alleviate this problem by improving spatial resolution, existing methods are mostly limited to single temporal images, making it difficult to meet the high-precision monitoring needs of complex spatio-temporal evolution of mangroves.

[0003] With the development of remote sensing technology, the research on multi-source data fusion has gradually matured. By fusing remote sensing images with different spatial and temporal resolutions, higher spatio-temporal resolution of mangrove temporal change information can be obtained. However, traditional spatio-temporal fusion methods rely too much on spectral information and fail to fully reflect the spectral feature changes of mangroves under different disturbance conditions. Moreover, they usually depend on the band consistency of high and low resolution images, which is difficult to meet in practical applications, affecting the effect of cross-sensor data fusion. In addition, these methods lack flexibility in dealing with sudden or nonlinear environmental changes and do not consider the complex spatial distribution characteristics of mangroves, resulting in unsatisfactory prediction accuracy in complex areas.

[0004] As an innovative method, spatio-temporal super-resolution mapping aims to overcome the contradiction between spatial and temporal resolution, and can directly generate high spatio-temporal resolution mangrove dynamic change maps to accurately capture the evolution process of mangroves. However, traditional spatio-temporal super-resolution mapping methods mainly rely on statistical models and classical image processing techniques, which are difficult to effectively handle complex nonlinear changes and spatio-temporal dependencies, limiting their application in high-precision mangrove monitoring. In recent years, deep learning methods have significantly improved the performance of spatio-temporal super-resolution mapping by automatically learning complex spatio-temporal relationships. Among them, models such as U-Net can effectively capture spatial details and multi-scale features in images through their encoding-decoding structure, making them suitable for high-resolution mangrove mapping tasks. Multi-source data fusion technology further enhances the model's ability to extract spatial and spectral features of ground objects by combining optical images, radar images, and lidar data, enabling the model to more comprehensively extract spatial and spectral features of ground objects, providing a new solution for high-precision spatio-temporal super-resolution mapping.

[0005] In recent years, the Mamba architecture based on state space model has shown high computational performance and fast inference ability in sequence modeling field, can carry out efficient training and real-time application on large-scale data set, has proved to have excellent multi-scale feature extraction ability in medical image analysis and other fields, can capture multi-level and multi-scale information in image, improve the precision of visual task, and also has superior transfer learning ability, can transfer knowledge between different visual tasks, improve the universality and adaptability of model, the vision state space module (Vision StateSpace Module, referred to as VSS module) of Mamba model can enhance the global feature modeling ability through the introduction of VSS module on the basis of keeping the excellent local feature extraction ability of encoding-decoding architecture, especially suitable for complex mangrove monitoring tasks which need high precision and multi-scale feature processing, therefore, the present application is supported by the spatio-temporal super-resolution mapping technology in computer vision technology, combines different resolution remote sensing data and exploratory combines encoding-decoding and VSS module to realize efficient extraction of mangrove spatio-temporal distribution characteristics, and proposes a mangrove super-resolution mapping method for deep extraction of spatio-temporal spectral features. SUMMARY

[0006] Technical problems to be solved

[0007] The purpose of the present application is to solve the problems of insufficient spatial resolution of existing mangrove remote sensing monitoring methods, leading to loss of broken area information, spectral distortion caused by time difference, and weak dynamic change feature capturing ability, and proposes a mangrove super-resolution mapping method for deep extraction of spatio-temporal spectral features.

[0008] Technical scheme

[0009] In order to solve the above problems, the present application adopts the following technical scheme:

[0010] A mangrove super-resolution mapping method for deep extraction of spatio-temporal spectral features, comprising the following steps:

[0011] S1, obtaining multi-source remote sensing image data in the research area, preprocessing the multi-source remote sensing image data, and then obtaining preprocessed data;

[0012] S2, classifying and extracting mangroves based on the object-oriented method for the preprocessed data, and then obtaining a high-resolution mangrove class map;

[0013] S3, extracting the mangrove class map based on SVR algorithm, and then obtaining a target time phase mangrove abundance map;

[0014] S4, constructing a spatio-temporal matching super-resolution mapping model training data set based on the target time phase mangrove abundance map;

[0015] S5, constructing the spatio-temporal super-resolution mapping network framework combined with the encoding-decoding and VSS modules based on the spatio-temporal matching super-resolution mapping model training data set, and then obtaining the spatio-temporal super-resolution mapping network framework model combined with the encoding-decoding and VSS modules;

[0016] S6, training and verifying the spatio-temporal super-resolution mapping network framework model combined with the encoding-decoding and VSS modules, then performing model optimization, and finally obtaining an optimized model;

[0017] S7, performing batch spatio-temporal super-resolution mapping based on the optimized model, and then obtaining a high spatio-temporal resolution mangrove distribution map of a target time phase.

[0018] As a preferred scheme of the present application, in step S1:

[0019] The multi-source remote sensing image data includes Landsat multi-spectral data with ten-meter resolution, Sentinel-2 data, and PlanetScope data with meter-level resolution.

[0020] The preprocessing includes format conversion, geographic registration, resampling, and cropping of the multi-source remote sensing image data.

[0021] As a preferred scheme of the present application, the step S2 includes the following steps:

[0022] S201, using the eCognition software to perform mangrove classification and extraction on the preprocessed 3-meter resolution PlanetScope data by using an object-oriented method, and then obtaining a classification result;

[0023] S202, adjusting and improving the classification result in combination with existing global mangrove observation data in the research area, and then obtaining a corresponding high-resolution mangrove category map.

[0024] As a preferred scheme of the present application, the step S3 includes the following steps:

[0025] S301, performing degradation processing on the obtained high-resolution mangrove category map in the spatial scale, estimating the abundance of the category by calculating the mean value of each pixel block, and obtaining a mangrove abundance map of a reference time phase as label data for SVR model training;

[0026] S302, based on the Landsat data and Sentinel-2 data of the target and reference time phases, respectively calculating the real ground reflectance value, and extracting the vegetation index related to mangroves;

[0027] S303, based on the pre-processed low spatial resolution data and the vegetation index related to mangrove, reasonably selecting samples and combining the intertidal zone distribution and the high reflection characteristic of the mangrove in the near-infrared-short wave infrared band, learning the unique spectral characteristics of the mangrove by using the SVR algorithm, extracting the area proportion of the mangrove in the mixed pixel, and then obtaining the target time phase mangrove abundance map;

[0028] S304, based on the trained SVR model, performing batch processing on the target time phase mangrove abundance map, applying the target time phase mangrove abundance map to all data of the target time phase research area, and then generating a low spatial resolution mangrove abundance map of the entire research area.

[0029] As a preferred scheme of the present application, the step S4 comprises the following steps:

[0030] S401, screening representative target and reference time phase data as training and test data, and cutting the data into image blocks of appropriate size, and reserving image blocks without abnormal values;

[0031] S402, based on the training area reference time phase and target time phase high resolution category map, constructing a multi-channel space-time change matrix map by comparing the values of the two time phase category maps on each channel, dynamically representing the change of the mangrove, and taking the space-time change matrix map as the training label of the subsequent super-resolution mapping model and the dynamic change supervision signal of the model;

[0032] S403, according to the space-time change characteristics of the mangrove in the research area, reasonably allocating training and test areas, aligning and standardizing the obtained target time phase low spatial resolution mangrove abundance map, the reference time phase high spatial resolution category map, the low spatial resolution abundance map and the space-time change map of the two time phases, and forming a data pair conforming to the model input.

[0033] As a preferred scheme of the present application, in step S5:

[0034] The space-time super-resolution mapping network framework combined with the encoding-decoding and VSS module comprises an encoder, a connection bridge, a decoder, a skip connection and a VSS module.

[0035] As a preferred scheme of the present application, the step S6 comprises the following steps:

[0036] S601, fusing the reference time phase high resolution mangrove image M1, the low spatial resolution abundance map F1 and the target time phase low spatial resolution mangrove abundance map F2 in a merging and cascading manner as the multi-channel input data of the network, and simultaneously taking the change matrix map CM as the label for supervising the model training;

[0037] S602, set the loss function of the model training, the expression is:

[0038]

[0039] Where N is the total number of pixels; c∈{0,1,2} represents the category index (0=unchanged, 1=increased, 2=decreased); y c (i) represents the true label of pixel i; p c (i) represents the probability that pixel i belongs to category c predicted by the model; α(i) represents the local proportional dynamic weight, which is used to enhance the attention to the changed area; λ represents the weight coefficient; LCR (i) represents the proportion of change categories in the local window around pixel i;

[0040] S603, during the training process, the multi-scale peak signal-to-noise ratio, the structural similarity, and the change detection accuracy are calculated through the validation set to monitor the model performance in real time and adjust the hyperparameters.

[0041] As a preferred scheme of the present application, the step S7 comprises the following steps:

[0042] S701, in the test program, apply the optimal weight batch saved in the step S6 to generate high-resolution mangrove temporal change images of the target area in batches;

[0043] S702, based on the high-resolution mangrove category map of the reference phase, update the generated high-resolution mangrove temporal change map of the target phase to obtain the high-spatial-temporal resolution mangrove distribution map of the target phase.

[0044] Beneficial effects

[0045] Compared with the prior art, the present application has the advantages of:

[0046] 1、The present application carries out super-resolution reconstruction to low-resolution remote sensing data through a deep learning model, so as to generate a high spatio-temporal resolution mangrove distribution map, aiming at the narrow and broken distribution characteristics of mangroves, on the one hand, we utilize the complementary advantages of different resolution remote sensing data, and combine the intertidal zone distribution of mangroves and its high reflection characteristics in the near-infrared-short wave infrared band, and adopt the SVR algorithm to learn the endmember spectral characteristics of mangroves in low spatial resolution data, so as to obtain the target phase mangrove abundance map, the present application can more accurately extract the spectral information of mangroves, provide high-quality input data for subsequent spatio-temporal super-resolution reconstruction, and also avoid the influence of the change of spectral information caused by the difference of time phase on the monitoring accuracy; on the other hand, we construct a spatio-temporal super-resolution mapping network combining coding-decoding and VSS module, enhance the attention of the model to local details, so that the network can effectively capture the subtle change characteristics when processing high dynamic change of mangrove data; in addition, considering that the traditional cross-entropy loss function may not effectively pay attention to details when processing the boundary and small range change area of mangroves, a suitable loss function is designed to constrain the training process of the model, and the gap between the predicted value and the real label is constantly reduced; finally, the model training weight parameter batch is tested to generate a super-resolution mangrove time series change map using the model with the best verification result, so as to obtain a high spatio-temporal resolution mangrove distribution map, the present application innovatively combines medium spatial resolution Sentinel-2, Landsat and high spatial resolution PlanetScope data, and exploratively combines coding-decoding and VSS module to realize efficient extraction of multi-temporal mangrove distribution characteristics, and generate a high spatial resolution mangrove distribution map, which provides more accurate data support for the protection and sustainable management of mangrove ecosystems. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is a flow chart of the present application;

[0048] Figure 2 is a detailed flowchart of the present application;

[0049] Figure 3 is a spatio-temporal super-resolution mapping network framework combining coding-decoding and VSS module of the present application;

[0050] Figure 4 is a structure diagram of the VSS module in the present application; DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0052] Embodiment 1

[0053] Please refer to Figures 1-4 A mangrove super-resolution mapping method based on spatio-temporal spectral feature deep extraction includes the following steps:

[0054] S1, obtaining multi-source remote sensing image data in the study area, pre-processing the multi-source remote sensing image data, and then obtaining pre-processed data;

[0055] Specifically, the obtained multi-source data in the study area includes ten-meter resolution Landsat multi-spectral data, Sentinel-2 data, and meter-level resolution PlanetScope data; wherein, the Landsat data has the advantages of free access, wide coverage and long time series, and is suitable for large-scale mangrove monitoring; the Sentinel-2 data can also be obtained for free, and its high spatial resolution and red edge band can effectively improve the vegetation recognition ability; Sentinel-2 and Landsat data are selected as low spatial resolution data sources, and through multi-temporal cooperative observation, the time series density can be significantly improved, and the problem of data missing caused by cloud cover or revisit period limitation of single source data can be effectively overcome; the PlanetScope data is used as a reference time-phase high spatial resolution data source, and its meter-level resolution can capture the fine structure and spatial distribution details of mangroves, but the cost of acquisition is higher, and the single scene coverage is small, which is difficult to independently support large-area continuous monitoring;

[0056] The pre-processing includes format conversion, geographic registration, resampling, cropping and the like of the obtained multi-source remote sensing image data, so as to ensure the same data format, research range and resolution size;

[0057] S2, based on the object-oriented method, the pre-processed data is classified and extracted for mangrove, and then a high-resolution mangrove category map is obtained, and the specific steps are as follows:

[0058] S201, using eCognition software, the 3-meter resolution PlanetScope data after pre-processing is classified and extracted for mangrove by using the object-oriented method; first, multi-scale segmentation is used to generate image objects with spatial, texture and spectral information, and then classification is performed in combination with spectral, texture and shape features (such as NDVI, NDWI, mean value, etc.);

[0059] S202, adjust and improve the classification results by combining the existing global mangrove observation data of the study area, thereby obtaining the corresponding high-resolution mangrove category map;

[0060] S3, based on the SVR algorithm, the mangrove category map is extracted, and then the target phase mangrove abundance map is obtained, and the specific steps are as follows:

[0061] S301, the high-resolution mangrove category map obtained is degraded in spatial scale, and the abundance of the category is estimated by calculating the mean value of each pixel block to obtain the mangrove abundance map of the reference phase, which is used as the label data for SVR model training;

[0062] S302, based on the target and reference phase Landsat and Sentinel-2 data, the true ground reflectance value is calculated, and the mangrove related vegetation index (NDVI, NDWI, MVI, etc.) is extracted, among which the mangrove vegetation index (MVI) can effectively enhance the spectral characteristics of mangrove and suppress background interference by combining near-infrared, short-wave infrared and green band; Finally, do well in outlier processing and align and standardize all data to eliminate the differences between different bands and data sources, and improve the applicability and stability of the algorithm;

[0063] S303, based on the pre-processed low spatial resolution data and mangrove related index, reasonable samples are selected and combined with the intertidal zone distribution unique to mangrove and its high reflection characteristics in near-infrared-short-wave infrared band, the SVR algorithm is used to learn the unique spectral characteristics of mangrove, and the area proportion of mangrove in mixed pixels is extracted, thereby obtaining the target phase mangrove abundance map; The input features of the model include blue, green, red, near-infrared, short-wave infrared band data and NDVI, NDWI, MVI and other indexes to capture the rich information of mangrove spectrum;

[0064] S304, use the trained SVR model for batch processing, and apply it to all data of the target phase study area to generate a low spatial resolution mangrove abundance map of the entire study area;

[0065] S4, based on the target phase mangrove abundance map, a spatio-temporal matching super-resolution mapping model training data set is constructed, and the specific steps are as follows:

[0066] S401, select representative target and reference phase data as training and testing data to ensure that they are representative in time and can capture the growth or decline process of mangrove, and crop the image blocks to an appropriate size and retain the non-outlier image blocks;

[0067] S402, based on the training area reference phase and the target phase high resolution category map, the multi-channel space-time change matrix map is constructed by comparing the values of two phase category maps in each channel pixel by pixel, which dynamically represents the change of mangrove and is used as the training label of subsequent super-resolution mapping model; the space-time change matrix map is used as the dynamic change supervision signal of the model, which not only helps to learn the spatial distribution characteristics of mangrove, but also effectively captures the time evolution law of mangrove;

[0068] S403, according to the space-time change characteristics of mangrove in the study area, the training and test areas are reasonably allocated; finally, the target phase low spatial resolution mangrove abundance map is obtained, and the reference phase high spatial resolution category map, low spatial resolution abundance map and space-time change map of two phases are aligned and standardized to form a data pair conforming to the model input;

[0069] Further, the final network input is: the low spatial resolution mangrove abundance map F1 of the reference phase and the corresponding high spatial resolution category map M1, the low spatial resolution mangrove abundance map F2 of the target phase, and the mangrove change matrix CM of two phases as the label;

[0070] S5, the space-time matching super-resolution mapping model training data set is constructed, and the space-time super-resolution mapping network framework combined with encoding-decoding and VSS module is obtained, and the space-time super-resolution mapping network framework combined with encoding-decoding and VSS module mainly includes five parts of encoder, connection bridge, decoder, jump connection and VSS module;

[0071] Further, the specific steps of constructing the space-time super-resolution mapping network framework combined with encoding-decoding and VSS module are as follows:

[0072] S501, the encoder part, the input data is first extracted by double convolution layer to extract the preliminary features of mangrove; then, through the combination structure of four down-sampling and VSS module, the multi-level and multi-scale semantic feature of mangrove is gradually extracted; among them, the down-sampling layer combines the maximum pooling and convolution operation, which gradually reduces the resolution and increases the channel number, so as to retain more details of mangrove; the introduction of VSS module can enhance the sensitivity of the model to small local space-time change, and its unique multi-scale attention mechanism further enhances the space-time information extraction ability, especially improves the accurate capture ability of mangrove boundary, which provides rich feature support for mapping task;

[0073] S502, the connection bridge part is responsible for transmitting the mangrove deep features extracted by the encoder to the decoder, and through multi-scale feature alignment and spatio-temporal information enhancement, it ensures that the mangrove boundary, texture and other features are accurately connected to the decoder for recovery, ensuring that the spatio-temporal information of the mangrove flows in the network and is reasonably converted;

[0074] S503, the decoder part is composed of four up-sampling modules, each of which is embedded with a VSS module to strengthen the reconstruction ability of the spatio-temporal features of the mangrove; wherein the up-sampling operation restores the resolution through bilinear interpolation or transposed convolution, and then further refines the features through double convolution layers; the introduction of the VSS module not only improves the sensitivity to changes in the mangrove, especially in handling narrow and broken areas, but also effectively restores the lost detail information in the down-sampling process; at the same time, this module promotes the fusion of the shallow features of the encoder and the deep features after up-sampling, gradually reconstructing high-resolution feature maps, and improving the reconstruction accuracy of the narrow structure and complex boundary of the mangrove;

[0075] S504, the skip connection part, which directly transmits the shallow features of the encoder to the corresponding layer of the decoder, ensures that the key information of the mangrove edge is not lost when restoring details; not only ensures the fine recovery of details, but also avoids the loss of information caused by excessive down-sampling, and enhances the recovery ability of local details of the mangrove;

[0076] S505, the VSS module part, the VSS module combines lightweight depth separable convolution and two-dimensional convolution, not only can extract local and global features, through the introduction of layer normalization to balance model efficiency and training robustness; and through the merging of the feature flow of the two paths, both local details and global information are retained, enhancing the representation ability of broken mangrove areas; embedding this module after the down-sampling layer, using multi-scale attention mechanism to enhance the extraction ability of spatio-temporal information, improve the sensitivity to spatio-temporal changes of the mangrove, especially in the recognition of changes in broken areas; the introduction of the VSS module ensures that the decoder effectively restores the spatio-temporal information in the up-sampling process, avoids the loss of details when restoring low-resolution images, and thus accurately reconstructs the spatial distribution and temporal changes of the mangrove, improves the adaptability and reconstruction accuracy of the model to the special distribution characteristics of the mangrove;

[0077] S6, train and verify the spatio-temporal super-resolution mapping network framework model combined with the encoding-decoding and VSS module, then optimize the model, and finally obtain the optimized model, the specific steps are as follows:

[0078] S601, merge, cascade and fuse the reference phase mangrove high-resolution image M1, the low spatial resolution abundance map F1 and the target phase low spatial resolution mangrove abundance map F2 as multi-channel input data of the network, and simultaneously use the change matrix map CM as a label for supervised model training;

[0079] S602, set the loss function of model training, in order to improve the recognition accuracy of the spatio-temporal change extraction model in the small range change area, an adaptive cross-entropy loss function is designed to reduce the gap between the predicted value and the true value, and its expression is:

[0080]

[0081] where N is the total number of pixels; c e {0, 1, 2} represents the class index (0 = no change, 1 = increase, 2 = decrease); y c (i) represents the true label of pixel i; p c (i) represents the probability that the model predicts that pixel i belongs to class c; a (i) represents the local proportional dynamic weight, which is used to enhance the attention to the change area; λ represents the weight coefficient; LCR (i) represents the proportion of change classes in the local window around pixel i;

[0082] S603, during the training process, calculate the multi-scale peak signal-to-noise ratio (PSNR), structural similarity (SSIM), change detection accuracy (Precision, Recall, F1-score) and other indicators through the validation set, real-time monitor the model performance and adjust the hyperparameters; when the loss value decreases to a certain extent and almost does not change, and the validation index value also increases to a certain extent and almost does not change, it can be considered that the model effect is better at this time, and the model weight is saved at this time;

[0083] S7, batch spatio-temporal super-resolution mapping based on the optimized model, and then obtaining the high spatio-temporal resolution mangrove distribution map of the target phase, which specifically includes the following steps:

[0084] In the test program, the optimal weight saved in step six is applied to batch generate the high-resolution mangrove time series change image of the target area; finally, based on the high-resolution mangrove category map of the reference phase, the generated high-resolution target phase mangrove time series change image is applied to update it, thereby obtaining the high spatio-temporal resolution mangrove distribution map of the target phase;

[0085] Example 2:

[0086] The embodiment proposes a mangrove super-resolution mapping method based on spatio-temporal spectral feature deep extraction. Taking ten-meter resolution Landsat satellite remote sensing image data, Sentinel-2 optical remote sensing image data and PlanetScope meter-level resolution remote sensing image data as examples, the method is used to monitor the spatio-temporal distribution of mangroves in some areas along the southeast coast of China. The specific implementation process can be divided into the following seven stages:

[0087] The first stage: obtaining and preprocessing multi-source remote sensing image data:

[0088] Step 1-1: Obtain multi-source remote sensing image data with long time span from October to December without cloud in some areas along the southeast coast of China from the U.S. Geological Survey, Copernicus Open Access Center and the official platform of PlanetLabs, including ten-meter resolution Landsat data, Sentinel-2 data and meter-level resolution PlanetScope data;

[0089] Step 1-2: Use SNAP software to preprocess Sentinel-2 data, mainly including using quality band to remove cloud and cloud shadow and filling invalid values, using nearest neighbor method to resample to 30 meters, and cropping to the study area;

[0090] Step 1-3: Check the data quality of Landsat data, use the quality evaluation band to generate a cloud mask to remove clouds, cloud shadows and saturated pixels, etc. and use the nearest neighbor method to fill, and finally crop to the study area;

[0091] The second stage: high spatial resolution mangrove category map extraction:

[0092] Step 2-1: Use the object-oriented method in eCognition software to classify and extract mangroves from two-time-phase PlanetScope high-resolution images. First, generate image objects with spatial, texture and spectral information through multi-scale segmentation, and then combine spectral, texture and shape features (such as NDVI, NDWI, mean and standard deviation, etc.) for classification;

[0093] Step 2-2: Combine the global mangrove observation data (Global Mangrove Watch, GMW) to manually adjust and improve the classification results, thereby obtaining the reference time phase high-resolution mangrove category map M1;

[0094] The third stage: low spatial resolution mangrove abundance map extraction:

[0095] Step 3-1: Degradation processing on the acquired reference phase high-resolution mangrove category map in spatial scale, estimate the abundance of the category by calculating the mean value of the block where each pixel is located, to obtain the reference abundance map F1 as the label for SVR model training;

[0096] Step 3-2: Based on the target and reference phase Landsat and Sentinel-2 data, the true land surface reflectance value is calculated respectively, and the mangrove related vegetation index (including NDVI, NDWI, MVI, etc.) is extracted, among which the mangrove vegetation index (MVI) can effectively enhance the spectral characteristics of mangrove and suppress background interference by combining near-infrared, short-wave infrared and green bands; Finally, do a good job in outlier processing, and all data will be aligned and standardized;

[0097] Step 3-3: Based on the processed low spatial resolution data and mangrove related index, reasonably select samples and combine the intertidal zone distribution unique to mangrove and its high reflection characteristics in near-infrared-short-wave infrared band, use SVR algorithm to learn its unique spectral characteristics, extract the area proportion of mangrove in mixed pixels, and obtain the target phase low spatial resolution mangrove abundance map; The input features of the model include blue, green, red, near-infrared, short-wave infrared band data and NDVI, NDWI, MVI and other indexes, and the abundance label F1 is used as the training target of SVR to train the model, to capture the rich information of mangrove spectrum;

[0098] Step 3-4: Use the trained SVR model for batch processing, apply it to all data of the target phase study area to generate the mangrove abundance map F2 of the entire study area;

[0099] The fourth stage: Constructing the training data set of spatio-temporal matching super-resolution mapping model:

[0100] Step 4-1: Select a representative multiple regional target and reference phase data as training and testing data, ensure that it is representative in time, and there is a certain change in the area of mangrove, to ensure that the dynamic change of mangrove can be captured;

[0101] Step 4-2: Uniformly pre-process the selected target and reference phase high and low spatial resolution data, remove outliers and fill in null values, and respectively crop to appropriate size, retaining no outlier image blocks;

[0102] Step 4-3: Based on the high-resolution category maps (M1, M2) of the training area reference time phase T1 and the target time phase T2, a multi-channel spatio-temporal change matrix CM is constructed by comparing the values of the two time phase category maps in each channel pixel by pixel, which dynamically represents the change of mangrove; wherein the first channel of CM represents the area without change, the second channel represents the area where mangrove increases (i.e. non-mangrove to mangrove), and the third channel represents the area where mangrove decreases (i.e. mangrove to non-mangrove);

[0103] Step 4-4: Reasonable allocation of training set and test set, the selected study area is allocated training and test data according to 5:3, forming a data pair (F1-F2-M1, CM) that meets the input requirements of the model;

[0104] The fifth stage: constructing a spatio-temporal super-resolution network model combining encoding-decoding and VSS module:

[0105] The spatio-temporal super-resolution network framework mainly includes five parts: encoder, connection bridge, decoder, skip connection and VSS module, as shown in Figure 2 ;

[0106] Step 5-1: The encoder part extracts spatio-temporal feature information by combining multiple convolution layers with VSS modules; VSS modules are added after each down-sampling layer to effectively enhance the ability to extract small local spatio-temporal change information of mangrove;

[0107] Step 5-2: The connection bridge part is responsible for converting the features output by the encoder and the input of the decoder to improve the decoding performance of the decoder;

[0108] Step 5-3: The decoder part gradually restores the image resolution through transposed convolution or bilinear interpolation, and adds VSS modules after each up-sampling operation and before the double convolution layer, which enhances the recovery ability of mangrove spatio-temporal features and effectively reduces the loss of detail information at the broken or boundary, and gradually fuses the shallow features of the encoder with the deep features after up-sampling;

[0109] Step 5-4: The skip connection part directly transmits the shallow features of the encoder to the corresponding layer of the decoder, and fuses them with the deep features output by the decoder, which enhances the detail recovery ability;

[0110] Step 5-5: as Figure 3As shown, in the VSS module, the input features are first mapped to a high-dimensional space by a linear embedding layer (Linear) to better capture complex feature representations, and then split into two paths; one path goes through a lightweight deep separable convolution (DWCNN) and a SiLU activation function, then enters a two-dimensional convolution (SS2D) module, and passes through layer normalization (LN); the other path directly goes through a SiLU activation function, preserving the global information of the original features; finally, the feature streams of the two paths are merged, which not only preserves local details but also integrates global information, enhancing the representation ability of the broken mangrove area; the combination of local spatial features extracted by DWCNN and global spectral-spatial features extracted by SS2D balances the model efficiency and training robustness through the introduction of LN; therefore, adding a VSS module in each layer of the encoder and decoder not only strengthens the encoder's capture of spatiotemporal information of mangroves, but also further improves the decoder's spatiotemporal feature reconstruction ability when restoring details; this design enables the model to more accurately restore and reconstruct detailed information when processing complex mangrove areas;

[0111] Sixth stage: model training and optimization:

[0112] Step 6-1: In the training sample set, in order to fully utilize the feature information and change information of mangroves in each time phase, multi-temporal data needs to be combined for model input; the specific combination method is: first subtract F1 from F2, then superimpose it with M1 in spatial scale, to get the input layer (2 layers (F2-F1), M1);

[0113] Step 6-2: Set the loss function for model training; in order to improve the recognition accuracy of the mangrove spatiotemporal change extraction model in small-scale change areas, an adaptive cross-entropy loss function is designed to reduce the gap between the predicted value and the true value; this loss function dynamically adjusts the weight of each pixel according to the local proportion of mangrove changes to enhance attention to detail areas such as boundaries and broken areas; its expression is:

[0114]

[0115]

[0116] where N is the total number of pixels; c e {0, 1, 2} represents the class index (0 = unchanged, 1 = increased, 2 = decreased); y c (i) represents the true label of pixel i; p c (i) represents the probability that pixel i belongs to class c predicted by the model; a(i) represents the local proportion dynamic weight, which is used to enhance attention to change areas; λ represents the weight coefficient; LCR (i) represents the proportion of change classes within the local window around pixel i;

[0117] Step 6-3: During the training process, the multi-scale peak signal-to-noise ratio (PSNR), structural similarity (SSIM), change detection precision (Precision, Recall, F1-score), etc. are calculated by the validation set to monitor the model performance in real time and adjust the hyperparameters; when the loss value decreases to a certain extent and almost does not change, and the validation index value also increases to a certain extent and almost does not change, it can be considered that the model effect is better at this time, and the model weight is saved at this time;

[0118] Seventh stage: model application and batch mapping:

[0119] Step 7-1: Save the model weight when the validation effect in the model training process reaches the optimal;

[0120] Step 7-2: Apply the optimal model weight in the test stage to batch generate the mangrove temporal change image of the target area;

[0121] Step 7-3: Based on the high-resolution mangrove distribution map of the reference phase, the generated high-resolution mangrove temporal change image of the target phase is applied to update it, so as to obtain the high-resolution mangrove category map of the target phase;

[0122] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical solution and the improvement concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction, characterized in that, The method comprises the following steps: S1, acquiring multi-source remote sensing image data in a study area, preprocessing the multi-source remote sensing image data, and then acquiring preprocessed data; S2, classifying and extracting mangrove forests from the preprocessed data based on an object-oriented method, and then acquiring a high-resolution mangrove forest category map; S3, extracting the mangrove forest category map based on an SVR algorithm, and then acquiring a target time-phase mangrove forest abundance map; S4, constructing a spatio-temporal matching super-resolution mapping model training data set based on the target time-phase mangrove forest abundance map; S5, constructing a spatio-temporal super-resolution mapping network framework combining coding-decoding and VSS modules based on the spatio-temporal matching super-resolution mapping model training data set, and then acquiring a spatio-temporal super-resolution mapping network framework model combining coding-decoding and VSS modules; S6, training and verifying the spatio-temporal super-resolution mapping network framework model combining coding-decoding and VSS modules, then optimizing the model, and finally acquiring an optimized model; S7, performing batch spatio-temporal super-resolution mapping based on the optimized model, and then acquiring a target time-phase high-spatio-temporal resolution mangrove forest distribution map; The step S6 comprises the following steps: S601, fusing the reference time-phase high-resolution mangrove forest image M1, the low-spatial-resolution abundance map F1, and the target time-phase low-spatial-resolution mangrove forest abundance map F2 in a merging and cascading manner as multi-channel input data of the network, and simultaneously taking the change matrix map CM as a label for supervising model training; S602, setting a loss function for model training, and the expression is: where N is the total number of pixels; c e {0, 1, 2} represents the class index, where: 0 = no change, 1 = increase, 2 = decrease; y c (i) represents the true label of pixel i; p c (i) represents the probability that the model predicts pixel i belongs to class c; a(i) represents the local scale dynamic weight, which is used to enhance the attention to the changed region; λ represents the weight coefficient; LCR(i) represents the proportion of the changed class in the local window around pixel i; S603, calculating the multi-scale peak signal-to-noise ratio, structural similarity, and change detection accuracy through a verification set during the training process, monitoring the model performance in real time, and adjusting the hyperparameters.

2. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 1, characterized in that, In step S1: The multi-source remote sensing image data comprises ten-meter resolution Landsat multi-spectral data, Sentinel-2 data, and meter-level resolution PlanetScope data; The preprocessing comprises: format conversion, geographic registration, resampling, and cropping of the multi-source remote sensing image data.

3. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 2, characterized in that, The step S2 comprises the following steps: S201, using the eCognition software to classify and extract mangrove forests from the preprocessed 3-meter resolution PlanetScope data based on an object-oriented method, and then acquiring a classification result; S202, adjusting and improving the classification result in combination with existing global mangrove observation data in the study area, and then obtaining a corresponding high-resolution mangrove forest category map.

4. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 3, characterized in that, The step S3 comprises the following steps: S301, performing degradation processing on the acquired high-resolution mangrove forest category map in the spatial scale, estimating the abundance of the category by calculating the mean value of each pixel block, to acquire a reference time-phase mangrove forest abundance map as label data for SVR model training; S302, calculating real ground reflectance values based on target and reference time-phase Landsat data and Sentinel-2 data, and extracting mangrove-related vegetation indices; S303, based on the pre-processed low spatial resolution data and the mangrove-related vegetation index, reasonably selecting samples and combining the intertidal zone distribution unique to mangroves and the high reflectivity of mangroves in the near-infrared-short-wave infrared band, using the SVR algorithm to learn the unique spectral characteristics of mangroves, extracting the area proportion of mangroves in the mixed pixel, and then obtaining the target time phase mangrove abundance map; S304, based on the trained SVR model, batch processing the target time phase mangrove abundance map, applying it to all data of the target time phase study area, and then generating a low spatial resolution mangrove abundance map of the entire study area.

5. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 4, characterized in that, The step S4 comprises the following steps: S401, screening representative target and reference time phase data as training and test data, and cutting them into image blocks of appropriate size, and retaining non-anomalous value image blocks; S402, based on the training area reference time phase and target time phase high-resolution category map, constructing a multi-channel spatio-temporal change matrix by comparing the values of the two time phase category maps in each channel, dynamically representing the change of mangroves, and using the spatio-temporal change matrix as the training label of the subsequent super-resolution mapping model, and using the spatio-temporal change matrix as the dynamic change supervision signal of the model; S403, according to the spatio-temporal change characteristics of the mangroves in the study area, reasonably allocating the training and test areas, aligning and standardizing the obtained target time phase low spatial resolution mangrove abundance map with the reference time phase high spatial resolution category map, low spatial resolution abundance map, and spatio-temporal change map of the two time phases, to form a data pair conforming to the model input.

6. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 5, characterized in that, In step S5: The spatio-temporal super-resolution mapping network framework combined with the encoding-decoding and VSS module comprises an encoder, a connection bridge, a decoder, a skip connection, and a VSS module.

7. The mangrove super-resolution mapping method of spatio-temporal spectral feature depth extraction according to claim 6, characterized in that, The step S7 comprises the following steps: S701, applying the optimal weight saved in the step S6 to batch generate the high-resolution mangrove time series change image of the target area in the test program; S702, based on the high-resolution mangrove category map of the reference time phase, applying the generated high-resolution target time phase mangrove time series change map to update it, and then obtaining the high-spatio-temporal resolution mangrove distribution map of the target time phase.

Images