A mangrove canopy height inversion method and system based on deep transfer learning

Abstract

This invention discloses a method and system for mangrove canopy height inversion based on deep transfer learning. The method includes: constructing a multi-source monthly time-series feature pre-training sample set; introducing a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; pre-training the missing-aware dual-stream fusion model based on the multi-source monthly time-series feature pre-training sample set to obtain a trained missing-aware dual-stream fusion model; and using the trained missing-aware dual-stream fusion model to invert and map the mangrove canopy height of a target area, obtaining a spatial distribution map of mangrove canopy height. This invention can achieve large-scale mangrove canopy height inversion and improve the inversion accuracy of the true mangrove canopy height. As a method and system for mangrove canopy height inversion based on deep transfer learning, this invention can be widely applied in the field of remote sensing image processing technology.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for retrieving mangrove canopy height based on deep transfer learning. Background Technology

[0002] Mangroves are unique wetland ecosystems growing in the intertidal zones of tropical and subtropical regions, playing irreplaceable ecological functions in coastal protection, biodiversity maintenance, and blue carbon sequestration. Mangrove canopy height is a key parameter for assessing its biomass, carbon storage, and ecosystem health; accurately obtaining the spatial distribution of mangrove canopy height over large areas has significant scientific and practical value. Traditional methods for obtaining mangrove canopy height mainly rely on field measurements. However, mangroves grow in harsh environments characterized by marshes and periodic tidal inundation, with extremely poor accessibility. Furthermore, their dense vegetation and tangled root systems make large-scale ground surveys virtually impossible. Therefore, field measurements can only obtain sporadic sample data, which cannot support the high-resolution mapping requirements at the regional scale.

[0003] Remote sensing technology provides an effective means for large-scale vegetation height inversion. Existing studies mostly utilize UAV-borne lidar to acquire high-precision canopy height data and combine it with satellite remote sensing imagery (such as the Sentinel series satellites) to establish inversion models. However, while UAV-borne lidar offers high accuracy, its coverage is extremely limited, typically only acquiring high-precision height data on the scale of a few square kilometers. Inversion models trained on such small-scale data generally have weak spatial generalization capabilities: when the model is applied to areas with different ecological conditions, tree species composition, or tidal conditions than the training plots, the prediction accuracy often drops sharply, making it difficult to meet the operational needs of large-scale continuous mapping. The Global Ecosystem Dynamics Investigation (GEDI) lidar, a spaceborne full-waveform lidar system carried by the International Space Station, provides large-scale vegetation height measurement data covering an area between approximately 51.6 degrees north and south latitude. GEDI's representative indicator, relative canopy height (RH98) (98% relative height), provides a global-scale canopy height reference, with a coverage far exceeding any airborne remote sensing method. However, due to limitations such as the resolution of spaceborne platform sensors, atmospheric conditions, and waveform interpretation algorithms, the accuracy of GEDI's RH98 data varies across different ecosystems. In particular, in complex mangrove communities, its accuracy and continuity cannot fully meet the requirements for high-precision mangrove height mapping.

[0004] Furthermore, mangroves are mostly distributed in tropical and subtropical cloudy and rainy areas, where optical satellite imagery is often obscured by clouds. Even after monthly time-series composite processing, some months still lack effective observation data due to persistent cloud cover. Traditional machine learning methods (such as random forests and gradient boosting machines) have a certain tolerance for missing values, but they struggle to transfer knowledge from large-scale low-precision data to high-precision small-sample data through transfer learning. While existing deep learning methods possess transfer learning capabilities, most models have limited ability to handle frequently occurring missing values ​​in time-series data. They typically use interpolation to fill in missing data or directly discard missing months. However, the non-random missing information in sensor data and the potential correlation between the missing time points and the target variable (e.g., the lack of effective observations in a certain month often reflects specific climatic characteristics) are not effectively utilized, resulting in compromised inversion accuracy. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention aims to provide a method and system for mangrove canopy height inversion based on deep transfer learning, which can achieve large-scale mangrove canopy height inversion and improve the inversion accuracy of the true value of mangrove canopy height.

[0006] The first technical solution adopted in this invention is: a method for mangrove canopy height inversion based on deep transfer learning, comprising the following steps: Acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time series feature pre-training sample set; A missing-aware convolutional feature extraction module is introduced to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; The missing-aware dual-stream fusion model for mangrove canopy height inversion was pre-trained based on a multi-source monthly time series feature pre-training sample set, resulting in the trained missing-aware dual-stream fusion model. Based on the trained missing-aware dual-stream fusion model, the mangrove canopy height of the target area is inverted and mapped to obtain a spatial distribution map of mangrove canopy height.

[0007] Furthermore, the step of acquiring multi-source time-series remote sensing data of the target area and performing data preprocessing to construct a multi-source monthly time-series feature pre-training sample set specifically includes: Acquire multi-source time-series remote sensing data of the target area, wherein the multi-source time-series remote sensing data includes first remote sensing data and second remote sensing data, wherein the first remote sensing data represents spaceborne synthetic aperture radar data and the second remote sensing data represents spaceborne multispectral optical data; Image data preprocessing is performed on the first remote sensing data to obtain the preprocessed first remote sensing data; Image data preprocessing is performed on the second remote sensing data to obtain preprocessed second remote sensing data; The preprocessed first remote sensing data and the preprocessed second remote sensing data were grouped according to the data of the same month of each year, and the median value of each pixel was taken to synthesize them to obtain the monthly representative observation image of the first remote sensing data and the monthly representative observation image of the second remote sensing data. Several radar index features were extracted from the monthly representative observation images of the first remote sensing data, and several optical vegetation index features were extracted from the monthly representative observation images of the second remote sensing data. By combining several radar index features of the monthly representative observation image of the first remote sensing data and several optical vegetation index features of the monthly representative observation image of the second remote sensing data into a time series by month, a multi-source monthly time series feature is obtained. The RH98 altitude index of the spaceborne lidar GEDI data was obtained and combined with multi-source monthly time series features to construct a multi-source monthly time series feature pre-training sample set.

[0008] Furthermore, the step of performing image data preprocessing on the first remote sensing data to obtain preprocessed first remote sensing data specifically includes: The first remote sensing data is corrected by performing orbit correction processing on the pre-constructed precise orbit file. Thermal noise removal is performed on the corrected first remote sensing data to obtain the first remote sensing data after thermal noise removal. Radiometric calibration converts the raw digital values ​​of the first remote sensing data after removing thermal noise into linear backscattering coefficients, and then converts the linear backscattering coefficients into decibel values ​​to obtain the preprocessed first remote sensing data.

[0009] Furthermore, the step of performing image data preprocessing on the second remote sensing data to obtain preprocessed second remote sensing data specifically includes: Atmospheric correction is performed on the second remote sensing data to obtain the corrected second remote sensing data. Based on the corrected second remote sensing data, a cloud mask is generated using the Fmask algorithm. Cloud pixels and cloud shadow pixels whose confidence does not meet the preset requirements are removed to obtain the preprocessed second remote sensing data.

[0010] Furthermore, the missing-aware dual-stream fusion model for mangrove canopy height inversion adopts a dual-stream architecture, specifically including a missing-aware convolutional feature extraction module, a seasonally learnable missing token replacement module, a cross-attention fusion module, a mask-aware temporal attention pooling module, and a prediction head module. The missing-aware convolutional feature extraction module, the seasonally learnable missing token replacement module, the cross-attention fusion module, the mask-aware temporal attention pooling module, and the prediction head module are connected sequentially.

[0011] Furthermore, the step of pre-training the missing-aware dual-stream fusion model for mangrove canopy height inversion based on a multi-source monthly time series feature pre-training sample set to obtain the trained missing-aware dual-stream fusion model specifically includes: The pre-trained sample set of multi-source monthly time series features is input into the missing-aware dual-stream fusion model for mangrove canopy height inversion; The missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model performs missing-aware convolutional feature extraction processing on the multi-source monthly time series feature pre-training sample set to obtain multi-source monthly time series missing-aware convolutional features. The seasonally learnable missing token replacement module based on the missing token-aware dual-stream fusion model performs seasonally learnable missing token replacement on the missing token-aware convolutional features of multi-source monthly time series according to the preset monthly missing token, and obtains the single-stream seasonal missing token-aware final features of multi-source monthly time series. The cross-attention fusion module based on the missing-aware dual-stream fusion model performs cross-fusion enhancement processing on the single-stream seasonal missing-aware final features of two multi-source monthly time series to obtain the dual-stream seasonal missing-aware enhanced features of the multi-source monthly time series. The mask-aware temporal attention pooling module based on the missing-aware dual-stream fusion model performs mask-aware temporal attention calculation on the dual-stream seasonal missing-aware enhancement features of multi-source monthly time series to obtain the global feature representation of the multi-source monthly time series. The prediction head module based on the missing-aware dual-stream fusion model predicts the mangrove canopy height value based on the global feature representation of multi-source monthly time series, and obtains the trained missing-aware dual-stream fusion model.

[0012] Furthermore, the missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model performs missing-aware convolutional feature extraction processing on the multi-source monthly time series feature pre-training sample set to obtain multi-source monthly time series missing-aware convolutional features. This step specifically includes: The pre-trained sample set of multi-source monthly time series features is input into the missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model; Define a preset missing value, perform missing value detection on the pre-trained sample set of multi-source monthly time series features, and generate a binary missing mask for the multi-source monthly time series features; Based on the binary missing mask of multi-source monthly time series features, all feature values ​​of time steps marked as missing by the mask are replaced with 0, while the original feature values ​​of non-missing time steps are retained, resulting in the cleaned multi-source monthly time series features. The cleaned multi-source monthly time series features are transposed, the binary missing mask is expanded, and the transposed result is concatenated with the expanded result to obtain the enhanced multi-source monthly time series features. The enhanced multi-source monthly time series features are sequentially convolved and transposed to obtain multi-source monthly time series missing-aware convolutional features.

[0013] Furthermore, the mask-aware temporal attention pooling module based on the missing-aware dual-stream fusion model performs mask-aware temporal attention calculation on the dual-stream seasonal missing-aware enhancement features of the multi-source monthly time series to obtain the global feature representation of the multi-source monthly time series. This step specifically includes: The binary missing mask of multi-source monthly time series features is converted into floating-point type, and the dimension is expanded in the last dimension to obtain the preprocessed binary missing mask; The preprocessed binary missing mask is concatenated with the dual-stream seasonal missing perception enhancement features of the multi-source monthly time series and input into the mask perception time attention pooling module for feature mapping to construct the original attention score matrix. The original attention score matrix is ​​normalized and then weighted and summed with the preset weight values ​​for each month to obtain the global feature representation of the multi-source monthly time series.

[0014] Furthermore, the step of performing mangrove canopy height inversion and mapping on the target area based on the trained missing-aware dual-stream fusion model to obtain a spatial distribution map of mangrove canopy height specifically includes: Obtain multi-source monthly time series features of all valid pixels within the target area; The multi-source monthly time series features of all effective pixels in the target area are input pixel by pixel into the trained missing-aware dual-stream fusion model to perform mangrove canopy height inversion and obtain the predicted mangrove canopy height value for each pixel. The predicted mangrove canopy height of each pixel is converted into a spatial raster map, and combined with the geographic coordinates of each pixel in the target area, a spatial distribution map of mangrove canopy height is generated.

[0015] The second technical solution adopted in this invention is: a mangrove canopy height inversion system based on deep transfer learning, comprising: The first module is used to acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time series feature pre-training sample set. The second module is used to introduce a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion. The third module is used to pre-train the missing-aware dual-stream fusion model for mangrove canopy height inversion based on a multi-source monthly time series feature pre-training sample set, and obtain the trained missing-aware dual-stream fusion model. The fourth module is used to perform mangrove canopy height inversion and mapping in the target area based on the trained missing-aware dual-stream fusion model, and obtain a spatial distribution map of mangrove canopy height.

[0016] The beneficial effects of the method and system of this invention are as follows: This invention acquires multi-source time-series remote sensing data of the target area and performs data preprocessing to construct a multi-source monthly time-series feature pre-training sample set; it introduces a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; through missing mask injection, seasonally learnable missing token replacement, and mask-aware time attention mechanism, it explicitly utilizes missing state information step by step from the bottom layer to the high-level semantics of the data, enabling the model to learn the climate or phenological patterns implied by the missing time points, significantly improving the prediction robustness under data missing conditions; based on the multi-source monthly time-series feature pre-training sample set, the missing-aware dual-stream fusion model for mangrove canopy height inversion is pre-trained to obtain the trained missing-aware dual-stream fusion model. This study introduces a deep transfer learning strategy. First, it pre-trains the model using GEDI spaceborne lidar data, which has a wide coverage and large data volume, enabling the model to master the general mapping relationship between multi-source remote sensing features and vegetation height. Then, it fine-tunes and calibrates the model using high-precision airborne lidar data from a small area. Based on the trained missing-perception dual-stream fusion model, it inverts and maps the mangrove canopy height in the target area, obtaining a spatial distribution map of mangrove canopy height. This fully leverages the complementary information of multi-source remote sensing data, combining the advantages of synthetic aperture radar data (unaffected by clouds and rain) with the rich spectral information of optical data. Furthermore, a cross-attention mechanism is used to enhance the feature interaction between the two data streams, thereby achieving optimal fusion and efficient utilization of remote sensing data in cloudy and rainy areas and improving the inversion accuracy of the true value of mangrove canopy height. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a mangrove canopy height inversion method based on deep transfer learning according to the present invention. Figure 2 This is a structural block diagram of a mangrove canopy height inversion system based on deep transfer learning according to the present invention; Figure 3This is a flowchart of mangrove canopy height inversion provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the network structure of the missing-aware dual-stream fusion model provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram comparing the accuracy of mangrove canopy height inversion according to a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the spatial distribution of mangrove canopy height inversion results provided in a specific embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0019] First, existing technologies have the following technical problems: the harsh growth environment of mangroves makes field measurements difficult; the traditional UAV-borne radar and satellite data joint modeling method has weak spatial generalization ability; and existing methods cannot simultaneously handle the dual technical challenges of frequent missing data in multi-source remote sensing time series data and effectively utilizing large-scale low-precision data for transfer learning.

[0020] Reference Figure 1 This invention provides a method for mangrove canopy height inversion based on deep transfer learning, the method comprising the following steps: S100. Acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time-series feature pre-training sample set; In this embodiment, the multi-source remote sensing data includes first remote sensing data and second remote sensing data, wherein the first remote sensing data is spaceborne synthetic aperture radar data and the second remote sensing data is spaceborne multispectral optical data; the construction of multi-source monthly time series features specifically includes: preprocessing the first remote sensing data and the second remote sensing data respectively, extracting the remote sensing features of each month, synthesizing the time series by month, and splicing the first remote sensing features and the second remote sensing features of the same month to form a multi-source monthly time series feature containing 12 months.

[0021] S110. Acquire multi-source time-series remote sensing data of the target area. The multi-source time-series remote sensing data includes first remote sensing data and second remote sensing data. The first remote sensing data represents spaceborne synthetic aperture radar data, and the second remote sensing data represents spaceborne multispectral optical data. In this embodiment, the first remote sensing data is a ground distance detection image from Sentinel-1 in wide-swath interferometric mode, with a spatial resolution of 10 meters, and vertical polarization is acquired. ) and vertical-horizontal cross-polarization ( The data is dual-polarization data. It can be downloaded from the European Space Agency's Copernicus Data Open Centre. The second remote sensing data is a Sentinel-2 L2A-class surface reflectance product, spatially resampled to 10 meters. , , , , , , , , , There are ten bands in total. Data can be downloaded from the European Space Agency's Copernicus Data Open Centre. The spaceborne lidar data is from a GEDI L2A-class product, from which extraction... The indicator is the relative height of the vegetation canopy. GEDI data can be obtained from NASA's Earth Data Platform. The airborne lidar data is point cloud data acquired by an UAV-borne LiDAR system. After denoising, ground point classification, and normalization, a canopy height model with a spatial resolution of 1 meter is generated to provide high-precision true values ​​of mangrove canopy height for the fine-tuning stage.

[0022] S120. Perform image data preprocessing on the first remote sensing data to obtain preprocessed first remote sensing data; Specifically, the first remote sensing data is subjected to orbit correction processing using a pre-constructed precise orbit file to obtain the corrected first remote sensing data; thermal noise is removed from the corrected first remote sensing data to obtain the first remote sensing data after removing thermal noise; the original digital values ​​of the first remote sensing data after removing thermal noise are converted into linear backscattering coefficients through radiometric calibration, and the linear backscattering coefficients are converted into decibel values ​​to obtain the preprocessed first remote sensing data.

[0023] In this embodiment, for Sentinel-1 data, the preprocessing steps include: performing orbit correction using a precise orbit file to eliminate orbital errors; removing thermal noise; converting the original digital values ​​into backscattering coefficients through radiometric calibration; performing terrain correction using the SRTM digital elevation model to eliminate the influence of terrain undulations; and converting the linear backscattering coefficients into decibel values. ; S130. Perform image data preprocessing on the second remote sensing data to obtain preprocessed second remote sensing data; Specifically, atmospheric correction is performed on the second remote sensing data to obtain corrected second remote sensing data; based on the corrected second remote sensing data, a cloud mask is generated using the Fmask algorithm, and cloud pixels and cloud shadow pixels whose confidence does not meet the preset requirements are removed to obtain preprocessed second remote sensing data.

[0024] In this embodiment, atmospheric correction is performed on Sentinel-2 data using the ESA Sen2Cor processor, and cloud masks are generated using the Fmask algorithm to remove cloud pixels and cloud shadow pixels with low confidence.

[0025] S140. The preprocessed first remote sensing data and the preprocessed second remote sensing data are grouped according to the data of the same month of each year, and the median value of each pixel is taken for synthesis to obtain the monthly representative observation image of the first remote sensing data and the monthly representative observation image of the second remote sensing data. In this embodiment, to reduce data loss caused by cloud pollution and capture the seasonal dynamics of vegetation growth, the preprocessed Sentinel-1 and Sentinel-2 data are synthesized into a time series by month. Data for the same month of each year are grouped, and the median value for each pixel is used to synthesize a representative observation image for that month, generating a total of 12 monthly images. Pixels with no effective observations for the entire month due to continuous cloud cover are marked as missing values ​​(in this embodiment, the missing value is set to...). ).

[0026] S150. Extract several radar index features from the monthly representative observation image of the first remote sensing data and extract several optical vegetation index features from the monthly representative observation image of the second remote sensing data. In this embodiment, for each month's synthesized image, multiple optical vegetation indices are extracted from Sentinel-2 data and multiple radar indices are extracted from Sentinel-1 data.

[0027] This embodiment calculates the following 16 optical vegetation indices based on the surface reflectance of ten bands from Sentinel-2: Normalized Difference Vegetation Index: ; Normalized Difference Vegetation Index 45: ; Red-edged normalized vegetation index: ; Normalized Differential Infrared Index Type I: ; Normalized Differential Infrared Index Type II: ; Surface water index: ; Normalized Building Index: ; Improved Normalized Difference Water Index: ; Wide dynamic range vegetation index: ; Terrestrial chlorophyll index: ; Mangrove vegetation index: ; Ratio vegetation index: ; Anti-red edge chlorophyll index: ; Anthocyanin reflectance index: ; Enhanced vegetation index: ; Combined Mangrove Identification Index: ; Among them, in the above formula , , , , , , , , , These represent the surface reflectance of the corresponding bands of the Sentinel-2 satellite.

[0028] Radar index calculation: First, the decibel-based backscattering coefficient is converted back to a linear value: ; Then calculate the following three radar indices: Normalized difference exponent: ; Radar vegetation index: ; Polarization ratio index: ; S160. Combine several radar index features of the monthly representative observation image of the first remote sensing data and several optical vegetation index features of the monthly representative observation image of the second remote sensing data into a time series by month to obtain multi-source monthly time series features. In this embodiment, for each month, all the extracted optical vegetation indices and radar indices are concatenated to form a multidimensional feature vector for that month. The features of all 12 months are then arranged in chronological order to form a shape... The multi-source monthly time series features, among which The total dimension of the features for each month. The final multi-source monthly time series features will be used as model training inputs in both the pre-training and fine-tuning phases.

[0029] S170. Obtain the RH98 altitude index from the spaceborne lidar GEDI data and combine it with multi-source monthly time series features to construct a multi-source monthly time series feature pre-training sample set.

[0030] In this embodiment, the pre-training sample set is centered on GEDI LiDAR footprint points, and multi-source monthly time series features of corresponding pixels are extracted. The target ground truth value is used. The sample set is randomly divided into a training set and a validation set at a certain ratio (e.g., 8:2). The fine-tuning sample set uses pixels within the coverage area of ​​the UAV-borne LiDAR as samples to extract multi-source monthly sequence features, and uses the height value of the canopy height model as the target ground truth value. It is also divided into a training set and a validation set at a certain ratio.

[0031] S200. Introduce a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; Specifically, the missing-aware dual-stream fusion model for mangrove canopy height inversion adopts a dual-stream architecture, which includes a missing-aware convolutional feature extraction module, a seasonally learnable missing token replacement module, a cross-attention fusion module, a mask-aware temporal attention pooling module, and a prediction head module. The missing-aware convolutional feature extraction module, the seasonally learnable missing token replacement module, the cross-attention fusion module, the mask-aware temporal attention pooling module, and the prediction head module are connected sequentially.

[0032] In this embodiment, the missing-aware convolutional feature extraction module is used to concatenate the missing mask as an additional channel with the original features and then extract features through a one-dimensional convolutional neural network; the seasonal learnable missing token replacement module is used to replace the convolutional output features at the missing position with the learnable parameter vector of the corresponding month; and the mask-aware temporal attention pooling module is used to explicitly input the missing mask into the attention calculation and dynamically allocate the weights of each month according to the data reliability.

[0033] like Figure 4As shown, the missing data-aware dual-stream fusion model adopts a dual-stream architecture, including a structurally symmetrical first remote sensing feature processing stream (corresponding to radar data features) and a second remote sensing feature processing stream (corresponding to optical data features). Each processing stream includes a missing data-aware convolutional feature extraction module, a seasonally learnable missing token replacement module, a cross-attention fusion module, and a mask-aware temporal attention pooling module.

[0034] S300. Based on the multi-source monthly time series feature pre-training sample set, the missing-aware dual-stream fusion model for mangrove canopy height inversion is pre-trained to obtain the trained missing-aware dual-stream fusion model. In this embodiment, for the pre-training stage: using spaceborne lidar data as the first altitude ground truth, and combining it with the multi-source monthly time series features to form a pre-training sample set, the initial model is pre-trained to obtain the pre-trained model; for the fine-tuning stage: using the parameters of the pre-trained model as initial values, using airborne radar data as the second altitude ground truth, and combining it with the multi-source monthly time series features to form a fine-tuning sample set, the pre-trained model is fine-tuned to obtain the final inversion model.

[0035] S310. Input the pre-trained sample set of multi-source monthly time series features into the missing-aware dual-stream fusion model for mangrove canopy height inversion; S320, The missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model performs missing-aware convolutional feature extraction processing on the multi-source monthly time series feature pre-training sample set to obtain multi-source monthly time series missing-aware convolutional features. Specifically, the pre-trained sample set of multi-source monthly time series features is input into the missing-aware convolutional feature extraction module based on a missing-aware dual-stream fusion model; a preset missing value is defined, and missing values ​​are detected on the pre-trained sample set of multi-source monthly time series features to generate a binary missing mask for the multi-source monthly time series features; based on the binary missing mask of the multi-source monthly time series features, all feature values ​​of time steps marked as missing by the mask are replaced with 0, while the original feature values ​​of non-missing time steps are retained, resulting in cleaned multi-source monthly time series features; the cleaned multi-source monthly time series features are transposed, the binary missing mask is expanded, and the transposed result is concatenated with the expanded result to obtain enhanced multi-source monthly time series features; the enhanced multi-source monthly time series features are then subjected to convolution and transpose processing in sequence to obtain multi-source monthly time series missing-aware convolutional features.

[0036] In this embodiment, the processing steps of the missing-aware convolutional feature extraction module include: for the input multi-source monthly time series features, generating a missing mask by detecting a preset missing marker value; after replacing the feature value at the missing position with zero, expanding the missing mask to the same spatial dimension as the original feature, and concatenating them in the feature channel dimension to obtain the enhanced input; and feeding the enhanced input into a multi-layer one-dimensional convolutional network for feature extraction.

[0037] More specifically, let the input multi-source monthly time series features be... ,in This represents the batch size, and 12 represents the number of months. This represents the feature dimension of the remote sensing data stream.

[0038] First, a missing mask is generated by detecting a preset missing marker value (-99 in this embodiment). : ; Replace the feature values ​​at the missing locations with zeros to obtain the cleaned data. : ; After expanding the missing mask to the same spatial shape as the input data, it is concatenated with the original features as an extra channel along the feature dimension to obtain the enhanced input. : ; in, Indicates will Transpose The shape, The shape was broadcast as .

[0039] The enhanced input is fed into a two-layer one-dimensional convolutional neural network for feature extraction. The first convolutional layer increases the number of input channels. The first convolution maps to 64 dimensions, and the second convolution maps 64 dimensions to 32 dimensions. The kernel size is 5, with padding of 2 to maintain the sequence length. Each convolution is followed by batch normalization and a ReLU activation function. The transposed convolution output has the following shape: .

[0040] S330. A seasonally learnable missing token replacement module based on a missing-aware dual-stream fusion model replaces the missing tokens of multi-source monthly time series with seasonally learnable missing tokens according to the preset monthly missing tokens, so as to obtain the single-stream seasonal missing-aware final features of multi-source monthly time series. In this embodiment, the processing steps of the seasonal learnable missing token replacement module include: creating an independent learnable parameter vector for each month, with the dimension of each parameter vector being consistent with the dimension of the convolutional output feature; for each sample, when there is missing data in a certain month, replacing the convolutional output feature of that month with the learnable parameter vector of the corresponding month; when the data of a certain month is valid, retaining the convolutional output feature of that month.

[0041] More specifically, construct a set of learnable parameter vectors The tokens represent the missing tokens for each of the 12 months, initialized as random values ​​with a mean of 0 and a standard deviation of 0.02. During model training, the token parameters are optimized along with other network parameters. Let the convolution output be... The seasonal token replacement formula is: ; This mechanism makes the representation of missing locations seasonally specific, meaning that missing data in dry months and missing data in rainy months will be assigned different learnable feature vectors, which the model can use to capture the implicit relationship between missing data and climate and phenological conditions.

[0042] S340. The cross-attention fusion module based on the missing perception dual-stream fusion model performs cross-fusion enhancement processing on the single-stream seasonal missing perception final features of two multi-source monthly time series to obtain the dual-stream seasonal missing perception enhanced features of the multi-source monthly time series. In this embodiment, a cross-attention fusion module is positioned between the first and second remote sensing feature processing streams to interactively enhance the intermediate features of the two streams; and a prediction head module is used to concatenate the features of the two streams after temporal attention pooling and output the predicted value of mangrove canopy height through a fully connected network.

[0043] More specifically, the first remote sensing feature processing stream and the second remote sensing feature processing stream each obtain Then, the two-stream information is interacted through a cross-attention fusion module. Let the characteristics of the second remote sensing stream be... The characteristics of the first remote sensing stream are The cross-attention calculation process is as follows: ; ; In this setup, the number of multi-head attention heads is set to 2, and the Dropout ratio is set to 0.4. The second remote sensing stream uses itself as the query and the first remote sensing stream as the key and value to obtain complementary information from the radar data stream. Similarly, the first remote sensing stream also uses itself as the query and the second remote sensing stream as the key and value for cross-attention fusion, resulting in... .

[0044] S350. A mask-aware temporal attention pooling module based on a missing-aware dual-stream fusion model performs mask-aware temporal attention calculation on the dual-stream seasonal missing-aware enhancement features of multi-source monthly time series to obtain the global feature representation of the multi-source monthly time series. Specifically, the binary missing mask of the multi-source monthly time series features is converted to floating-point type, and the dimension is expanded in the last dimension to obtain the preprocessed binary missing mask. The preprocessed binary missing mask is concatenated with the dual-stream seasonal missing perception enhancement features of the multi-source monthly time series and input into the mask perception time attention pooling module for feature mapping to construct the original attention score matrix. The original attention score matrix is ​​normalized and weighted and summed with the preset weight values ​​of each month to obtain the global feature representation of the multi-source monthly time series.

[0045] In this embodiment, the processing steps of the mask-aware temporal attention pooling module include: converting the missing mask into a floating-point type and expanding its dimensions, concatenating it with the feature after the missing token replacement on the feature dimension, inputting it into the attention computing network to obtain the attention weights for each month; and performing weighted pooling on the features of multiple months based on the attention weights to obtain a fixed-dimensional feature representation of the processing stream.

[0046] More specifically, the features after cross-attention fusion enter the mask-aware temporal attention pooling module. Taking the second remote sensing stream as an example, its input is... and missing mask .

[0047] First, convert the missing mask to floating-point type and expand its dimensions: ； ; Then, it is concatenated with the features and input into an attention network to calculate the attention score for each month: ; in, Map the input dimension 33 (32+1) to 16. Map 16 to 1.

[0048] Attention weights are obtained through softmax normalization: ; Finally, the features from the 12 months are weighted and pooled: ; The first remote sensing feature processing stream uses the same method to obtain... .

[0049] Because the missing mask is explicitly input into the attention network, the model can dynamically adjust the importance weights of each month based on whether the data for that month is valid, automatically reducing the contribution weight of missing months, thereby improving the prediction robustness under high missing rate conditions.

[0050] S360, a prediction head module based on a missing-aware dual-stream fusion model, predicts mangrove canopy height values ​​based on the global feature representation of multi-source monthly time series, and obtains the trained missing-aware dual-stream fusion model.

[0051] More specifically, the outputs of the two processes that have passed through temporal attention pooling are concatenated: ; The concatenated features are input into a fully connected prediction head network. The first layer maps 64 dimensions to 32 dimensions, with intermediate normalization, ReLU activation, and Dropout (scale 0.5). The second layer directly maps the 32 dimensions to a 1-dimensional output. ; in, This represents the mangrove canopy height predicted by the model.

[0052] Furthermore, during the pre-training phase, the missing-aware dual-stream fusion model is trained using a pre-training sample set. The loss function is the mean squared error. ; in, These are the model's predicted values. For GEDI High truth value, This represents the number of samples in the batch.

[0053] The optimizer used is Adam, and the initial learning rate is set to... The batch size is 128, and the number of training epochs is set to 200. An early stopping strategy is adopted: training stops when the validation set loss does not decrease for 10 consecutive epochs, and the parameters of the model that performs best on the validation set are saved as the pre-trained model.

[0054] Through this stage of training, the model learns a general mapping relationship from multi-source monthly time series features to vegetation height.

[0055] During the fine-tuning phase, all parameters of the pre-trained model are loaded as initial values, and training is performed using the fine-tuning sample set. The loss function also uses mean squared error. ; in, This is a high-precision true value of mangrove canopy height obtained by an airborne LiDAR on a drone.

[0056] The optimizer for the fine-tuning stage still uses Adam, but the initial learning rate is appropriately reduced to 1. The batch size was set to 64, and the number of training rounds was set to 100. The early stopping strategy was also used to save the optimal model on the validation set. The fine-tuned model is the final mangrove canopy height inversion model.

[0057] S400, based on the trained missing-aware dual-stream fusion model, performs mangrove canopy height inversion and mapping on the target area, and obtains a spatial distribution map of mangrove canopy height.

[0058] Specifically, the multi-source monthly time series features of all valid pixels within the target area are obtained; the multi-source monthly time series features of all valid pixels within the target area are input pixel by pixel into the trained missing-aware dual-stream fusion model to perform mangrove canopy height inversion, obtaining the predicted mangrove canopy height value for each pixel; the predicted mangrove canopy height value for each pixel is converted into a spatial raster map, and combined with the geographic coordinate information of each pixel within the target area, a spatial distribution map of mangrove canopy height is generated.

[0059] The multi-source monthly time-series features of all valid pixels within the target study area are input pixel-by-pixel into a fine-tuned missing-aware dual-stream fusion model to obtain the predicted mangrove canopy height for each pixel. Combining the geographic coordinates of each pixel, the prediction results are output as a spatial raster map to generate a spatial distribution map of mangrove canopy height, such as... Figure 6 As shown.

[0060] Through the above steps, this invention achieves fully automated inversion of mangrove canopy height spatial distribution from multi-source satellite remote sensing time-series data. For example... Figure 5 As shown, the performance comparison results of our method with various models such as CatBoost, HistGBM, LightGBM, RandomForest, and XGBoost in the mangrove canopy height inversion task are presented. Under the same sample size (N=767), the prediction results of our method have higher consistency with the measured values, and its coefficient of determination is higher. The accuracy reached 0.73, with a root mean square error (RMSE) of 1.75 m and a mean absolute error (MAE) of 1.18 m, which is generally better than other comparative models (comparative models). The scatter plots show that the predicted points of the method in this invention are more concentrated near the 1:1 reference line, with lower dispersion, indicating higher accuracy and better stability in the inversion results. Especially in the mid-to-high canopy height range, the method significantly reduces systematic underestimation compared to other models. Furthermore, the scatter plot density distribution shows that the method maintains a smaller bias in high-density sample areas, demonstrating stronger generalization ability and robustness. Therefore, the method in this invention has superior accuracy and application potential in mangrove canopy height inversion.

[0061] Therefore, the embodiments of this invention firstly address the difficulty of on-site measurement in mangrove areas. This invention is entirely based on satellite remote sensing data and airborne / spaceborne lidar data, eliminating the need for large-scale ground-based manual surveys and enabling automated inversion of mangrove canopy height over a wide area. Secondly, to balance large-scale mapping capabilities with high-precision inversion requirements, this invention introduces a deep transfer learning strategy. It first pre-trains the model using GEDI spaceborne lidar data, which has a wide coverage area and large data volume, enabling the model to master the universal mapping relationship between multi-source remote sensing features and vegetation height. Then, it combines this with small-scale, high-precision airborne lidar data for fine-tuning and calibration, thereby overcoming the bottleneck of insufficient spatial generalization ability in traditional UAV methods. Furthermore, it addresses the challenges of tropical... To address the frequent missing data in remote sensing time series data in cloud-prone areas, this invention constructs a missing data-aware dual-stream fusion model. Through missing mask injection, seasonally learnable missing token replacement, and a mask-aware temporal attention mechanism, it explicitly utilizes missing state information hierarchically from the data's lower to higher semantic levels. This enables the model to learn the implicit climate or phenological patterns at missing time points, significantly improving prediction robustness under data-missing conditions. Finally, this invention fully leverages the complementary information from multi-source remote sensing data, combining the advantages of synthetic aperture radar data being unaffected by clouds and rain with the rich spectral information of optical data. A cross-attention mechanism is used to enhance the feature interaction between the two data streams, thereby achieving optimal fusion and efficient utilization of remote sensing data in cloudy and rainy areas.

[0062] Finally, as Figure 3 As shown, in the multi-source feature extraction stage of this embodiment of the invention, synthetic aperture radar data from Sentinel-1 and multispectral optical data from Sentinel-2 are acquired respectively. The two types of data are preprocessed, monthly synthesized, and feature extracted respectively. The extracted first remote sensing features and second remote sensing features are merged to form a unified multi-source monthly time series feature.

[0063] During the model pre-training phase, multi-source monthly time series features and spaceborne lidar GEDI data are combined. By combining height indicators, a pre-training sample set is constructed to pre-train the missing-sensor dual-stream fusion model, resulting in a pre-trained model. This stage leverages the wide coverage of GEDI data to enable the model to learn a general mapping relationship from multi-source remote sensing features to vegetation height.

[0064] In the model fine-tuning and output stage, the parameters of the pre-trained model are used as initial values. High-precision mangrove canopy height data acquired by UAV-borne LiDAR and multi-source monthly time series features are combined to form a fine-tuning sample set to fine-tune and calibrate the model, obtaining the optimal inversion model. Finally, the multi-source monthly time series features of the target area are input into the optimal model for pixel-by-pixel prediction to generate a spatial distribution map of mangrove canopy height.

[0065] In summary, this invention achieves feature extraction for large-area missing features in remote sensing time series data of cloudy regions by acquiring monthly time series features from multi-source remote sensing data; inputting these features into a pre-constructed missing-aware dual-stream fusion model; and dynamically replacing missing features by injecting a missing mask as an extra channel into the convolutional network and introducing seasonally learnable tokens at the missing locations. The model training employs a deep transfer learning strategy, first pre-training with RH98 data from the GEDI spaceborne lidar to learn general mapping relationships, and then fine-tuning with high-precision airborne radar data to calibrate the model. Finally, the fine-tuned model is used to invert and obtain the mangrove canopy height map. This invention effectively solves the dual technical challenges of the special growth environment of mangroves leading to difficulties in field measurement and the high-frequency missing features in remote sensing data of tropical cloudy regions, achieving effective fusion of multi-source remote sensing time series features and large-scale, high-precision mangrove height mapping.

[0066] Therefore, the embodiments of the present invention have the following advantages compared with the prior art: 1) Effectively solves the problem of difficulty in field measurement of large-scale mangrove forests. This invention is based entirely on satellite remote sensing data and airborne / spaceborne lidar data, and can realize the automated inversion of mangrove canopy height over a large area without the need for large-scale ground manual surveys.

[0067] 2) Balancing the needs for large-scale mapping with high-precision inversion. This invention employs a deep transfer learning strategy, first using GEDI spaceborne lidar data with wide coverage and large data volume for pre-training, enabling the model to learn the general mapping relationship from multi-source remote sensing features to vegetation height; then using high-precision airborne lidar data with a small area for fine-tuning and calibration, overcoming the bottleneck of weak spatial generalization ability of traditional UAV methods.

[0068] 3) Effectively addressing the problem of frequent missing remote sensing time-series data in tropical cloudy regions. The missing data-aware dual-stream fusion model constructed in this invention utilizes missing mask injection, seasonally learnable missing token replacement, and mask-aware temporal attention mechanism to explicitly utilize missing state information step by step from the data's lower-level semantics to the higher-level semantics. This enables the model to learn the climate or phenological patterns inherent in the missing time points themselves, significantly improving the prediction robustness under data missing conditions.

[0069] 4) Fully leverage the complementary information from multi-source remote sensing data. This invention simultaneously utilizes the characteristic that synthetic aperture radar data is unaffected by clouds and rain, and the characteristic that optical data has rich spectral information. By using a cross-attention mechanism to enhance the features of the two data streams, it achieves optimal fusion and utilization of remote sensing data in cloudy and rainy areas.

[0070] Reference Figure 2 A mangrove canopy height inversion system based on deep transfer learning includes: The first module 201 is used to acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time series feature pre-training sample set; The second module 202 is used to introduce a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; The third module 203 is used to pre-train the missing-aware dual-stream fusion model for mangrove canopy height inversion based on a multi-source monthly time series feature pre-training sample set, and obtain the trained missing-aware dual-stream fusion model. The fourth module 204 is used to perform mangrove canopy height inversion and mapping in the target area based on the trained missing-aware dual-stream fusion model, and obtain a spatial distribution map of mangrove canopy height.

[0071] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0072] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A method for mangrove canopy height inversion based on deep transfer learning, characterized in that, Includes the following steps: Acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time series feature pre-training sample set; A missing-aware convolutional feature extraction module is introduced to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion; The missing-aware dual-stream fusion model for mangrove canopy height inversion was pre-trained based on a multi-source monthly time series feature pre-training sample set, resulting in the trained missing-aware dual-stream fusion model. Based on the trained missing-aware dual-stream fusion model, the mangrove canopy height of the target area is inverted and mapped to obtain a spatial distribution map of mangrove canopy height.

2. The method for mangrove canopy height inversion based on deep transfer learning according to claim 1, characterized in that, The step of acquiring multi-source time-series remote sensing data of the target area and performing data preprocessing to construct a multi-source monthly time-series feature pre-training sample set specifically includes: Acquire multi-source time-series remote sensing data of the target area, wherein the multi-source time-series remote sensing data includes first remote sensing data and second remote sensing data, wherein the first remote sensing data represents spaceborne synthetic aperture radar data and the second remote sensing data represents spaceborne multispectral optical data; Image data preprocessing is performed on the first remote sensing data to obtain the preprocessed first remote sensing data; Image data preprocessing is performed on the second remote sensing data to obtain preprocessed second remote sensing data; The preprocessed first remote sensing data and the preprocessed second remote sensing data were grouped according to the data of the same month of each year, and the median value of each pixel was taken to synthesize them to obtain the monthly representative observation image of the first remote sensing data and the monthly representative observation image of the second remote sensing data. Several radar index features were extracted from the monthly representative observation images of the first remote sensing data, and several optical vegetation index features were extracted from the monthly representative observation images of the second remote sensing data. By combining several radar index features of the monthly representative observation image of the first remote sensing data and several optical vegetation index features of the monthly representative observation image of the second remote sensing data into a time series by month, a multi-source monthly time series feature is obtained. The RH98 altitude index of the spaceborne lidar GEDI data was obtained and combined with multi-source monthly time series features to construct a multi-source monthly time series feature pre-training sample set.

3. The method for mangrove canopy height inversion based on deep transfer learning according to claim 2, characterized in that, The step of performing image data preprocessing on the first remote sensing data to obtain preprocessed first remote sensing data specifically includes: The first remote sensing data is corrected by performing orbit correction processing on the pre-constructed precise orbit file. Thermal noise removal is performed on the corrected first remote sensing data to obtain the first remote sensing data after thermal noise removal. Radiometric calibration converts the raw digital values ​​of the first remote sensing data after removing thermal noise into linear backscattering coefficients, and then converts the linear backscattering coefficients into decibel values ​​to obtain the preprocessed first remote sensing data.

4. The method for mangrove canopy height inversion based on deep transfer learning according to claim 3, characterized in that, The step of performing image data preprocessing on the second remote sensing data to obtain preprocessed second remote sensing data specifically includes: Atmospheric correction is performed on the second remote sensing data to obtain the corrected second remote sensing data. Based on the corrected second remote sensing data, a cloud mask is generated using the Fmask algorithm. Cloud pixels and cloud shadow pixels whose confidence does not meet the preset requirements are removed to obtain the preprocessed second remote sensing data.

5. The method for mangrove canopy height inversion based on deep transfer learning according to claim 4, characterized in that, The missing-aware dual-stream fusion model for mangrove canopy height inversion adopts a dual-stream architecture, specifically including a missing-aware convolutional feature extraction module, a seasonally learnable missing token replacement module, a cross-attention fusion module, a mask-aware temporal attention pooling module, and a prediction head module. The missing-aware convolutional feature extraction module, the seasonally learnable missing token replacement module, the cross-attention fusion module, the mask-aware temporal attention pooling module, and the prediction head module are connected sequentially.

6. The method for mangrove canopy height inversion based on deep transfer learning according to claim 5, characterized in that, The step of pre-training the missing-aware dual-stream fusion model for mangrove canopy height inversion based on a multi-source monthly time series feature pre-training sample set to obtain the trained missing-aware dual-stream fusion model specifically includes: The pre-trained sample set of multi-source monthly time series features is input into the missing-aware dual-stream fusion model for mangrove canopy height inversion; The missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model performs missing-aware convolutional feature extraction processing on the multi-source monthly time series feature pre-training sample set to obtain multi-source monthly time series missing-aware convolutional features. The seasonally learnable missing token replacement module based on the missing token-aware dual-stream fusion model performs seasonally learnable missing token replacement on the missing token-aware convolutional features of multi-source monthly time series according to the preset monthly missing token, and obtains the single-stream seasonal missing token-aware final features of multi-source monthly time series. The cross-attention fusion module based on the missing-aware dual-stream fusion model performs cross-fusion enhancement processing on the single-stream seasonal missing-aware final features of two multi-source monthly time series to obtain the dual-stream seasonal missing-aware enhanced features of the multi-source monthly time series. The mask-aware temporal attention pooling module based on the missing-aware dual-stream fusion model performs mask-aware temporal attention calculation on the dual-stream seasonal missing-aware enhancement features of multi-source monthly time series to obtain the global feature representation of the multi-source monthly time series. The prediction head module based on the missing-aware dual-stream fusion model predicts the mangrove canopy height value based on the global feature representation of multi-source monthly time series, and obtains the trained missing-aware dual-stream fusion model.

7. The method for mangrove canopy height inversion based on deep transfer learning according to claim 6, characterized in that, The missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model performs missing-aware convolutional feature extraction processing on the multi-source monthly time series feature pre-training sample set to obtain multi-source monthly time series missing-aware convolutional features. This step specifically includes: The pre-trained sample set of multi-source monthly time series features is input into the missing-aware convolutional feature extraction module based on the missing-aware dual-stream fusion model; Define a preset missing value, perform missing value detection on the pre-trained sample set of multi-source monthly time series features, and generate a binary missing mask for the multi-source monthly time series features; Based on the binary missing mask of multi-source monthly time series features, all feature values ​​of time steps marked as missing by the mask are replaced with 0, while the original feature values ​​of non-missing time steps are retained, resulting in the cleaned multi-source monthly time series features. The cleaned multi-source monthly time series features are transposed, the binary missing mask is expanded, and the transposed result is concatenated with the expanded result to obtain the enhanced multi-source monthly time series features. The enhanced multi-source monthly time series features are sequentially convolved and transposed to obtain multi-source monthly time series missing-aware convolutional features.

8. The method for mangrove canopy height inversion based on deep transfer learning according to claim 7, characterized in that, The mask-aware temporal attention pooling module based on the missing-aware dual-stream fusion model performs mask-aware temporal attention calculation on the dual-stream seasonal missing-aware enhancement features of multi-source monthly time series to obtain the global feature representation of the multi-source monthly time series. This step specifically includes: The binary missing mask of multi-source monthly time series features is converted into floating-point type, and the dimension is expanded in the last dimension to obtain the preprocessed binary missing mask; The preprocessed binary missing mask is concatenated with the dual-stream seasonal missing perception enhancement features of the multi-source monthly time series and input into the mask perception time attention pooling module for feature mapping to construct the original attention score matrix. The original attention score matrix is ​​normalized and then weighted and summed with the preset weight values ​​for each month to obtain the global feature representation of the multi-source monthly time series.

9. The method for mangrove canopy height inversion based on deep transfer learning according to claim 8, characterized in that, The step of inverting and mapping the mangrove canopy height of the target area based on the trained missing-aware dual-stream fusion model to obtain a spatial distribution map of mangrove canopy height specifically includes: Obtain multi-source monthly time series features of all valid pixels within the target area; The multi-source monthly time series features of all effective pixels in the target area are input pixel by pixel into the trained missing-aware dual-stream fusion model to perform mangrove canopy height inversion and obtain the predicted mangrove canopy height value for each pixel. The predicted mangrove canopy height of each pixel is converted into a spatial raster map, and combined with the geographic coordinates of each pixel in the target area, a spatial distribution map of mangrove canopy height is generated.

10. A mangrove canopy height inversion system based on deep transfer learning, characterized in that, Includes the following modules: The first module is used to acquire multi-source time-series remote sensing data of the target area and perform data preprocessing to construct a multi-source monthly time series feature pre-training sample set. The second module is used to introduce a missing-aware convolutional feature extraction module to construct a missing-aware dual-stream fusion model for mangrove canopy height inversion. The third module is used to pre-train the missing-aware dual-stream fusion model for mangrove canopy height inversion based on a multi-source monthly time series feature pre-training sample set, and obtain the trained missing-aware dual-stream fusion model. The fourth module is used to perform mangrove canopy height inversion and mapping in the target area based on the trained missing-aware dual-stream fusion model, and obtain a spatial distribution map of mangrove canopy height.

Images