A codec double-focusing based tree height and biomass collaborative inversion method, system, device and medium

Abstract

The application discloses a kind of based on codec double focusing tree height and biomass collaborative inversion method, system, equipment and medium, including the following steps: obtaining the tree height and biomass data of target area, and obtaining the optical and synthetic aperture radar data of corresponding position;The data obtained are preprocessed, and training data set is constructed;Establish multi-task learning model based on encoder and decoder double focusing, train training data set by multi-task learning, generate tree height and biomass collaborative inversion model;Obtain the optical and synthetic aperture radar data in research area, and input tree height and biomass collaborative inversion model to carry out tree height and biomass collaborative inversion, and compare predicted value with true value, assess inversion accuracy.The application improves the collaborative inversion accuracy of strong correlation target.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing data application, and particularly relates to a tree height and biomass collaborative inversion method, system, device and medium based on codec double focusing. BACKGROUND

[0002] The inversion of structural parameters of forests is the basis for ecological monitoring and climate research. Tree height, as a key parameter of forest vertical structure, is directly related to the estimation of above-ground biomass, the evaluation of timber volume, and the monitoring of forest degradation. Biomass, as a direct representation of carbon storage, its accurate inversion is an important prerequisite for quantifying forest carbon sink capacity.

[0003] However, the collaborative inversion of tree height and biomass parameters currently faces the following technical bottlenecks:

[0004] At the data level, a single data source cannot meet the target. Optical remote sensing has high spectral resolution, but is limited by the electromagnetic wave penetration ability, and can only obtain two-dimensional spectral information of the canopy surface, which cannot directly invert the tree height and other forest vertical structures, and is easily disturbed by weather factors such as clouds and aerosols. Synthetic aperture radar can achieve all-weather observation, but the nonlinear relationship between its backscattering signal and forest parameters is complex, and is significantly affected by factors such as terrain undulation and multiple scattering of vegetation layers. When used alone for biomass inversion, the accuracy fluctuates greatly. Spaceborne lidar can penetrate the canopy to obtain true tree height, but the data is distributed in discrete spots, which cannot directly generate continuous surface tree height and biomass grids, and therefore needs to be combined with other remote sensing data to achieve scale expansion. However, data fusion still has technical bottlenecks. Optical, synthetic aperture radar and lidar data have significant differences in spatial resolution, observation mechanism and data form. Traditional fusion methods cannot effectively capture the physical correlation between data, resulting in limited improvement in inversion accuracy.

[0005] At the model level, single-task models lack collaboration. Traditional tree height and biomass inversion methods use independent modeling. Tree height inversion relies on lidar point clouds and optical vegetation indices, while biomass inversion focuses on synthetic aperture radar backscattering coefficients and optical spectral features. However, these methods separate the inherent relationship between tree height and biomass, leading to contradictions in the physical logic of the inversion results. The architecture of multi-task models has limitations. Existing multi-task learning models can reuse features through shared encoders, but the decoders lack explicit interaction between tasks, making it difficult to handle the strong correlation between tree height and biomass. Existing models do not design specific cross-task interaction mechanisms, which cannot achieve information complementarity between tree height and biomass. For example, biomass inversion cannot utilize the representation ability of tree height for vegetation vertical structure, resulting in decreased accuracy in complex forest environments. SUMMARY

[0006] Invention purposes: The first purpose of the present application is to provide a codec double focusing based tree height and biomass collaborative inversion method, which improves the collaborative inversion precision of the strongly correlated target.

[0007] The second purpose of the present application is to provide a codec double focusing based tree height and biomass collaborative inversion system.

[0008] The third purpose of the present application is to provide an electronic device.

[0009] The fourth purpose of the present application is to provide a computer storage medium.

[0010] Technical scheme: In order to achieve the above purposes, the present application provides a codec double focusing based tree height and biomass collaborative inversion method, which comprises the following steps:

[0011] (1) obtaining the tree height and biomass data of the target area by using the spaceborne laser radar, and obtaining the optical and synthetic aperture radar data of the corresponding position;

[0012] (2) preprocessing the obtained data to construct a training data set;

[0013] (3) establishing a multi-task learning model based on the encoder and decoder double focusing, training the training data set through multi-task learning to generate a collaborative inversion model of tree height and biomass; the multi-task learning model comprises an encoder focusing module and a decoder focusing module, wherein the input data of the training data set is input into a residual fusion encoding network in the encoder focusing module, and tree height branch features and biomass branch features are generated respectively, the tree height branch features and the biomass branch features are input into a double-task feature interaction network in the decoder focusing module, and tree height inversion results and biomass inversion results are output;

[0014] The residual fusion encoding network comprises two parallel branch structures of a tree height encoder and a biomass encoder, each branch passes through 5 residual modules in turn, the convolution kernel size of the residual module decreases from large to small, and the channel number doubles per module; after each residual module is processed, a cross-branch interaction module is accessed, the cross-branch interaction module adds the feature maps output by the upper and lower branch residual modules in the channel, and then performs a random shuffling operation on the added feature maps; after 5 times of residual module and cross-branch interaction module processing, the tree height branch features and the biomass branch features are input into pyramid pooling modules for multi-scale pooling operation;

[0015] The double-task feature interaction network comprises two parallel branch structures of a tree height decoder and a biomass decoder, each branch passes through 3 groups of residual modules and up-sampling operations to obtain tree height intermediate features and biomass intermediate features, and then the tree height intermediate features and the biomass intermediate features are input into a feature interaction module for cross-task feature fusion to output the final tree height result and biomass result.

[0016] (4) Obtain the optical and synthetic aperture radar data in the study area, and input into the synergic inversion model of tree height and biomass to perform synergic inversion of tree height and biomass, and compare the predicted value with the true value to evaluate the inversion accuracy.

[0017] Optionally, the step (1) specifically comprises the following steps:

[0018] (1.1) Download the GEDI data of the target area, the tree height data is derived from the L2 product of the spaceborne lidar, the biomass data is derived from the L4 product of the spaceborne lidar, and the L2A and L4A files of the spaceborne lidar are read to extract the tree height, biomass data, longitude and latitude of the spaceborne lidar spot, original waveform data and elevation information, and the spot is screened by using the screening condition, and the screening condition is:

[0019] 1) When the sensitivity is less than 0.95, the corresponding spot is removed;

[0020] 2) When the elevation difference between the GEDI elevation and the SRTM elevation is greater than 30m, the corresponding spot is removed;

[0021] 3) When the degrade flag is greater than 0, the stale_return_flag is 0, the degrade_flag is 1, and the quality_flag is 0, the corresponding spot is removed;

[0022] 4) When the rx_assess_flag is 1, the corresponding spot is removed;

[0023] (1.2) Add a buffer zone to the longitude and latitude information to generate a vector information file of the corresponding position;

[0024] (1.3) According to the vector information file, download the optical and synthetic aperture radar data of the corresponding position.

[0025] Optionally, the step (2) specifically comprises the following steps:

[0026] (2.1) Optical data preprocessing: according to the time of the tree height and biomass data, the image with a time deviation of less than or equal to 15 days is selected, if there is no data on the same day, the image closest in time is selected; according to the QA_PIXEL waveband, the pixel points of the remote sensing image are masked, and the image cloud removal work is completed;

[0027] (2.2) Synthetic aperture radar data preprocessing: according to the time of tree height and biomass data, the image corresponding to the time is screened; if there is no corresponding time on the same day, the closest time is selected; the radar observation data is converted into backscattering coefficient through radiation calibration, the original data information dB value is converted into intensity information, and the spot noise in the image is removed;

[0028] (2.3) Training dataset making: the tree height, biomass, optical and synthetic aperture radar data are cut into multiple data groups according to the size of 300*300 pixels, and the data groups without effective true value of tree height and biomass are removed; through random sampling, the screened data is divided into training set, test set and validation set according to the ratio of 40%:30%:30%.

[0029] Optionally, the input data in step (3) is taken from the training dataset, and the input data includes optical data and synthetic aperture radar data, wherein the optical data includes 11 band information extracted from the Sentinel 2 image, the synthetic aperture radar data includes radar backscattering coefficients of two polarization modes extracted from the Sentinel 1 image, and the size of the input data is uniformly regularized to 300*300*13.

[0030] Optionally, in the cross-branch interaction module, the features of the two branches are first added in the channel dimension to obtain the fusion features , wherein is the output of the i-th residual module of the tree height branch, i is the output of the i-th residual module of the biomass branch, and wherein represents the number of channels of the output of the i-th residual module; i i , wherein the channel is randomly shuffled, which can be represented as a random rearrangement of the channel index:

[0031] ,

[0032] , wherein is a random permutation function of the channel index.

[0033] Optionally, the pyramid pooling module includes 3*3 pooling kernels, 15*15 pooling kernels, 30*30 pooling kernels and global pooling kernels, and the pyramid pooling module performs multi-scale down-sampling on the input feature to obtain multi-scale pooling features by performing average pooling, which is represented as:

[0034] ,

[0035] ,

[0036] ,​​​

[0037] ,

[0038] wherein, is an input feature, , , , ;

[0039] Channel dimension fusion is performed on the multi-scale pooled features, that is, first, bilinear up-sampling is performed on , and respectively, so that the image size is 100x100, to obtain , and , and then the channel dimension is added to obtain the fused feature , which is expressed as:

[0040] ,

[0041] The final output tree height branch feature and biomass branch feature .

[0042] Optionally, the tree height intermediate feature and the biomass intermediate feature in the feature interaction module are spliced in the channel dimension to construct joint features and , wherein represents a channel dimension splicing operation;

[0043] Then, through a learnable weight matrix , 100 is the number of channels of the joint feature, and 50 is the number of channels of the single task feature, the dependence weight of the tree height task on the biomass task and the dependence weight of the biomass task on the tree height task are calculated, and the calculation formula is:

[0044] ,

[0045] ,

[0046] wherein, softmax is an activation function, and then based on the dependence weight, the tree height and biomass intermediate features are cross-task fused:

[0047] ,

[0048] ,

[0049] wherein a post-fusion feature for tree height, a post-fusion feature for biomass, and finally outputting the tree height and biomass inversion results through a full connection layer.

[0050] Based on the same inventive concept, the application discloses a tree height and biomass collaborative inversion system based on encoder-decoder double focusing, comprising:

[0051] a data acquisition module, configured to acquire tree height and biomass data of a target region by using a spaceborne laser radar, and to acquire optical and synthetic aperture radar data of a corresponding position;

[0052] a data preprocessing module, configured to preprocess the acquired data and construct a training data set;

[0053] a collaborative inversion model construction module, configured to establish a multi-task learning model based on encoder-decoder double focusing, to train the training data set through multi-task learning, and to generate a collaborative inversion model for tree height and biomass; the multi-task learning model comprises an encoder focusing module and a decoder focusing module, wherein the input data of the training data set are input into a residual fusion encoding network in the encoder focusing module, and tree height branch features and biomass branch features are generated respectively, and the tree height branch features and the biomass branch features are input into a double-task feature interaction network in the decoder focusing module, and tree height inversion results and biomass inversion results are output;

[0054] The residual fusion encoding network comprises two parallel branch structures of a tree height encoder and a biomass encoder, each branch sequentially passes through 5 residual modules, the convolution kernel size of the residual modules decreases from large to small, and the number of channels doubles from module to module; after each residual module is processed, a cross-branch interaction module is accessed, the cross-branch interaction module performs channel addition operation on the feature maps output by the upper and lower branch residual modules, and then performs random shuffling operation on the added feature maps; after 5 times of residual module and cross-branch interaction module processing, the tree height branch features and the biomass branch features are input into pyramid pooling modules for multi-scale pooling operation, and finally output.

[0055] The double-task feature interaction network comprises two parallel branch structures of a tree height decoder and a biomass decoder, each branch sequentially passes through 3 groups of residual modules and up-sampling operations to obtain tree height intermediate features and biomass intermediate features, and then the tree height intermediate features and the biomass intermediate features are input into a feature interaction module for cross-task feature fusion to output final tree height results and biomass results.

[0056] a collaborative inversion prediction module, configured to acquire optical and synthetic aperture radar data in a research region, to input the collaborative inversion model for tree height and biomass into the collaborative inversion model for tree height and biomass collaborative inversion, and to compare the predicted value with the true value to evaluate the inversion accuracy.

[0057] Based on the same inventive concept, an electronic device according to the present application comprises a processor and a storage medium;

[0058] The storage medium is used to store instructions;

[0059] The processor is used to operate according to the instructions to perform the steps of the method as described above.

[0060] Based on the same inventive concept, a computer-readable storage medium according to the present application has a computer program stored thereon, wherein the program is executed by a processor to implement the steps of the method as described above.

[0061] Advantages: Compared with the prior art, the present application has the following significant advantages:

[0062] (1) The encoder focusing module is designed to improve the multi-source data fusion efficiency, enhance the cross-modal feature synergy of optical and synthetic aperture radar data, and exploit the complementarity between data;

[0063] (2) The decoder focusing module is designed to improve the complementary mechanism between strong correlation target inversion tasks;

[0064] (3) The double focusing strategy of the encoder and the decoder is used to form a closed-loop mechanism from feature sharing to task interaction, which improves the cooperative inversion accuracy of the strong correlation target;

[0065] (4) The residual fusion encoding network is designed to realize feature complementarity and enhance the generalization ability among multiple tasks through the cross-branch interaction module, and to balance the local details and global information of the features through the pyramid pooling module, thereby enhancing the utilization rate of spatial information of the model;

[0066] (5) The double-task feature interaction network is designed to realize adaptive fusion through dynamic calculation of task dependency weight, fully utilize complementary information, and retain the core features of each task while avoiding invalid information interference. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 is a flowchart of the present application;

[0068] Figure 2 is a schematic diagram of the multi-task learning network structure in the present application;

[0069] Figure 3 is a schematic diagram of the residual fusion encoding network structure in the present application;

[0070] Figure 4 is a schematic diagram of the residual module structure in the present application;

[0071] Figure 5 Figure 1 is a schematic diagram of the cross-branch interaction mechanism in the present application;

[0072] Figure 6 Figure 2 is a schematic diagram of the pyramid pooling structure in the present application;

[0073] Figure 7 Figure 3 is a schematic diagram of the double-task feature interaction network structure in the present application;

[0074] Figure 8 Figure 4 is a schematic diagram of the feature interaction module structure in the present application. DETAILED DESCRIPTION

[0075] The technical solutions of the present application will be further described below with reference to the accompanying drawings.

[0076] Example 1: As shown in the present application, a tree height and biomass collaborative inversion method based on codec double focusing is disclosed, comprising the following steps: Figure 1

[0077] (1) Obtain the tree height and biomass data of the target area using the spaceborne laser radar, and obtain the optical and synthetic aperture radar data of the corresponding position;

[0078] Step (1) specifically comprises the following steps:

[0079] (1.1) Download the spaceborne laser radar GEDI data of the target area, the tree height data is derived from the 2nd level product of the spaceborne laser radar, and the biomass data is derived from the 4th level product of the spaceborne laser radar, simultaneously read the L2A and L4A files of the spaceborne laser radar, extract the tree height, biomass data, latitude and longitude of the spaceborne laser radar spot, original waveform data and elevation information, and use the screening condition to screen the spot, and retain the good quality spot;

[0080] In order to ensure the quality of the spaceborne laser radar GEDI data, the screening condition is as follows:

[0081] 1) When the sensitivity < 0.95, the corresponding spot is removed;

[0082] 2) When the elevation difference between the GEDI elevation and the SRTM elevation is greater than 30m, the corresponding spot is removed;

[0083] 3) When the degrade > 0, stale_return_flag = 0, degrade_flag = 1, and quality_flag = 0, the corresponding spot is removed;

[0084] 4) When the rx_assess_flag = 1, the corresponding spot is removed; ​

[0085] The screening information is shown in Table 1, and the screening conditions are shown in Table 2.

[0086] Table 1 Screening parameters of spaceborne lidar

[0087]

[0088] Table 2 GEDI screening conditions of spaceborne lidar

[0089]

[0090] (1.2) Add a buffer zone to the latitude and longitude information to generate a vector information file corresponding to the location;

[0091] (1.3) According to the vector information file, download optical and synthetic aperture radar data corresponding to the location; the optical and synthetic aperture radar data are derived from Sentinel 1 and Sentinel 2, as shown in Tables 3 and 4.

[0092] Table 3 Sentinel 1 image extraction parameters

[0093]

[0094] Table 4 Sentinel 2 image extraction parameters

[0095]

[0096] (2) Preprocess the obtained data, construct a training data set, and divide it into a training set, a test set, and a validation set;

[0097] Step (2) specifically includes the following steps:

[0098] (2.1) Optical data preprocessing: according to the time of tree height and biomass data, screen images with a time deviation of ≤15 days, and if there is no data on the same day, select the image closest in time; according to the QA_PIXEL waveband, mask the pixel points of the remote sensing image, and use the Fmask algorithm to complete the image cloud removal work;

[0099] (2.2) Synthetic aperture radar data preprocessing: according to the time of tree height and biomass data, screen images corresponding to the time; if there is no corresponding time on the same day, select the closest time; convert radar observation data into backscatter coefficient through radiometric calibration, and the original data information is dB value, which needs to be converted into intensity information, the formula is as follows:

[0100] ,

[0101] Lee filtering is used to remove speckle noise in the image, which reduces the noise effect while preserving the image details;

[0102] (2.3) Training dataset making:

[0103] The tree height, biomass, optical and synthetic aperture radar data were cut into multiple data sets according to the size of 300*300 pixels, and the data sets without valid true value of tree height and biomass were removed to ensure the quality of the data set for model training; through random sampling, the screened data was divided into training set, test set and validation set according to the ratio of 40%:30%:30%; wherein the tree height and biomass data are the labels of the multi-task learning model, and the optical and synthetic aperture radar data are the input features of the multi-task learning model.

[0104] (3) A multi-task learning model based on encoder and decoder double focusing is established, and the training data set is trained through multi-task learning to generate a collaborative inversion model of tree height and biomass;

[0105] As shown in Figure 2 , the multi-task learning model includes an encoder focusing module and a decoder focusing module, wherein the input data of the training data set is input into the residual fusion encoding network in the encoder focusing module to generate tree height branch features and biomass branch features respectively, and the tree height branch features and the biomass branch features are input into the double-task feature interaction network in the decoder focusing module to output the tree height inversion result and the biomass inversion result.

[0106] The input data is taken from the training data set, and the input data includes optical data and synthetic aperture radar data, wherein the optical data is 11 band information extracted from Sentinel-2 in Table 4, and the synthetic aperture radar data is the radar backscattering coefficient of 2 polarization modes of Sentinel-1 image in Table 3, and the input data size is uniformly regularized to 300*300*13.

[0107] As shown in Figure 3 , the residual fusion encoding network includes two parallel branch structures of tree height encoder and biomass encoder, each branch passes through 5 residual modules in turn, and the convolution kernel size of the residual module decreases from large to small, and the channel number doubles per module; after each residual module processing, access to cross-branch interaction module, in the cross-branch interaction module, the feature maps output by the upper and lower branch residual modules are added in channel, and then the added feature maps are randomly shuffled; after 5 times of residual module and cross-branch interaction module processing, the multi-scale pooling operation in the pyramid pooling module is input respectively, and finally the tree height branch features and biomass branch features are output.

[0108] The first residual module uses a 7*7 convolution kernel, and the structure is as shown in Figure 4As shown, the second residual module adopts a 5x5 convolution kernel, the third residual module adopts a 3x3 convolution kernel, and the fourth and fifth residual modules both adopt a 1x1 convolution kernel; the feature image size is unchanged in the five residual modules, and the feature image size is 300x300, and the number of feature channels is doubled in turn, being 26, 52, 104, 208 and 416.

[0109] As shown in FIG. 1, the input feature of the first residual module sequentially passes through a 7x7 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 7x7 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result. Figure 4

[0110] The input feature of the second residual module sequentially passes through a 5x5 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 5x5 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result.

[0111] The input feature of the third residual module sequentially passes through a 3x3 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 3x3 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel, and then passes through a ReLU activation function to output the result.

[0112] The input feature of the fourth and fifth residual modules sequentially passes through a 1x1 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 1x1 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result.

[0113] As shown in FIG. 1, to realize feature fusion between tree height and biomass branches, the cross-branch interaction module first adds the two branch features in the channel dimension to obtain the fused feature Figure 5 This operation directly fuses the two branch features in the channel dimension and shares complementary information; wherein is the output of the first residual module of the tree height branch, is the output of the first residual module of the biomass branch, and wherein i represents the number of channels of the output of the first residual module, i i ​​​​​​

[0114] ,

[0115] wherein, is a random permutation function of channel index, which ensures the dynamic change of channel order each time of training, and improves the model generalization.

[0116] After five residual modules and cross-branch interaction modules are processed, considering that the spot range of spaceborne lidar data is 25 meters, the spatial resolution of Sentinel 1 and Sentinel 2 is 10 meters, and the resolution of spaceborne lidar is about 3 times of optical and synthetic aperture radar data, in order to realize the spatial matching of input features and label information, an input pyramid pooling module is introduced, and a pyramid pooling operation is introduced.

[0117] As shown in Figure 6 , the pyramid pooling module includes a 3x3 pooling kernel, a 15x15 pooling kernel, a 30x30 pooling kernel and a global pooling kernel, and the input features in the pyramid pooling module are down-sampled at multiple scales, and average pooling is performed to obtain multi-scale pooling features:

[0118] ,

[0119] ,

[0120] ,

[0121] ,

[0122] wherein, is the input feature, , , , ;

[0123] In order to keep the number of channels unchanged, the channel dimension fusion is performed on the multi-scale pooling features, first, the bilinear up-sampling is performed on and respectively, so that the image size is 100x100, that is, first, the bilinear up-sampling is performed on , and respectively, so that the image size is 100x100, and , and are obtained, then the channel dimension is added, and the fusion feature is obtained, which is expressed as:

[0124] ,

[0125] The number of channels of the fused feature is still 416, and the size is 100x100, and finally the tree height branch feature Biomass branch features As the input of the decoder focusing module, the feature extraction process of the residual fusion coding network is completed.

[0126] In order to realize the matching of the input features (optical data and synthetic aperture radar data) and the label information (tree height and biomass) in the spatial dimension, a pyramid pooling operation is adopted to reduce the feature image size to 1 / 3 of the original size while keeping the number of feature channels unchanged, and finally output the tree height branch features and the biomass branch features.

[0127] As shown in Figure 7 , the double-task feature interaction network includes two parallel branch structures of tree height decoder and biomass decoder, and each branch structure sequentially passes through three groups of residual modules and up-sampling operations, and then enters the feature interaction module for cross-task fusion to output the final tree height result and biomass result.

[0128] The input data of the double-task feature interaction network is the tree height branch features and the biomass branch features generated by the encoder focusing module, and the two groups of feature data are processed in parallel; the structure of the double-task feature interaction network is as shown in Figure 7 , that is, the input data is the tree height branch features and the biomass branch features generated by the encoder focusing module, and the two groups of data are input into the tree height decoder and the biomass decoder in parallel and start the processing process in parallel.

[0129] Firstly, the data sequentially passes through three groups of residual modules and up-sampling operations, and after the residual module processing, the feature image size remains unchanged, while the number of feature channels decreases. Since the up-sampling operation will gradually increase the feature image size, it effectively extracts features of different scales, and the convolution kernel size used by the residual module is set to increase from small to large. After three up-sampling operations, the feature image size is restored to the original input size.

[0130] Both the tree height decoder and the biomass decoder branch sequentially pass through three groups of residual modules and up-sampling operations. Since the up-sampling operation will increase the feature image size, in order to ensure that the receptive field is approximately equal, the convolution kernel used by the residual module is sequentially increased, which is 3x3, 5x5 and 7x7, respectively. After the residual module, the feature image size remains unchanged, and the feature is sequentially reduced, which is 200, 100 and 50, respectively. After each residual module, the up-sampling operation is followed, and the feature image size is sequentially increased, which is 150x150, 200x200 and 300x300, respectively, restoring to the original input size, and the number of feature channels remains unchanged.

[0131] After three groups of residual modules and up-sampling operations, the intermediate features of the tree height decoder and the biomass decoder are obtained, which are the tree height intermediate features and the biomass intermediate features , into the feature interaction module, and the cross-task dependency relationship between the tree height and the biomass task is learned.

[0132] As shown in Figure 8 , the tree height intermediate feature and the biomass intermediate feature are spliced in the channel dimension in the feature interaction module to construct joint features and , wherein represents a channel dimension splicing operation.

[0133] The dependency weight of the tree height task on the biomass task and the dependency weight of the biomass task on the tree height task are calculated through a learnable weight matrix , wherein 100 is the number of joint feature channels, and 50 is the number of single-task feature channels.

[0134] ,

[0135] ,

[0136] The softmax function ensures that the weights are normalized, so that reflects the dependency ratio of the tree height task on the biomass task feature, reflects the dependency ratio of the biomass task on the tree height task feature.

[0137] Based on the dependency weight, the cross-task fusion of the tree height and biomass task features is performed:

[0138] ,

[0139] ,

[0140] wherein and are the intermediate features of the tree height task decoder and the biomass task decoder, which form joint features by channel splicing in the feature interaction module, and are learnable weight matrices, and are the feature vectors of the tree height task and the biomass task after cross-task interaction processing; and finally, the output of the tree height and biomass inversion results is completed through a fully connected layer.

[0141] In the present application, the encoder and decoder double focusing strategy is optimized by using a multi-task joint loss function, wherein the loss function is composed of: using mean square error (MSE) as the basic loss, and dynamically adjusting the tree height loss and biomass loss The total loss is weighted by the weight of the biomass loss L The calculation formula is:

[0142] ,

[0143] Wherein, is a tree height task uncertainty parameter, is a biomass task uncertainty parameter, which is automatically learned by the network; is an L2 regularization term of the encoder parameter, is a regularization coefficient.

[0144] Optimization strategy: adopt Adam optimizer, set the initial learning rate to 1e-4, and decay by 0.8 times after every 50 training rounds; the training batch size is set to 32, and the iteration is 200 rounds until the loss function converges, and the convergence threshold is the loss fluctuation amplitude.

[0145] (4) Obtain optical and synthetic aperture radar data in the research area, and input the tree height and biomass collaborative inversion model to perform tree height and biomass collaborative inversion, and compare the predicted value with the true value to evaluate the inversion accuracy;

[0146] The root mean square error RMSE is used as the accuracy evaluation index, and the formula is as follows:

[0147] ,

[0148] In the formula, n is the total number of samples, is the model predicted value, is the true value, and the range of the root mean square error is When the predicted value is completely consistent with the true value, it is equal to 0, that is, perfect prediction; the worse the prediction effect is, the larger the root mean square error value is.

[0149] Embodiment 2: The application discloses a kind of tree height and biomass collaborative inversion systems based on codec double focusing, comprising:

[0150] Data acquisition module is used to obtain the tree height and biomass data of target area by using spaceborne lidar, and obtains the optical and synthetic aperture radar data of corresponding position;Download the spaceborne lidar GEDI data of target area, the tree height data is derived from the 2nd product of spaceborne lidar, the biomass data is derived from the 4th product of spaceborne lidar, simultaneously reads the L2A and L4A files of spaceborne lidar, extracts the longitude and latitude of tree height, biomass data, spaceborne lidar light spot, original waveform data and elevation information, filters light spot using screening condition, and retains good quality light spot;

[0151] In order to ensure the quality of spaceborne lidar GEDI data, the screening condition is as follows:

[0152] 1) When sensitivity < 0.95, the corresponding light spot is removed;

[0153] 2) When the elevation difference between GEDI elevation and SRTM elevation is greater than 30m, the corresponding light spot is removed;

[0154] 3) When the degradation flag degrade > 0, the stale flag stale_return_flag = 0, the degradation flag degrade_flag = 1, and the quality flag quality_flag = 0, the corresponding light spot is removed;

[0155] 4) When the error condition flag rx_assess_flag = 1, the corresponding light spot is removed; the screening information is shown in Table 1, and the screening condition is shown in Table 2;

[0156] A buffer zone is added to the latitude and longitude information to generate a vector information file of the corresponding position;

[0157] According to the vector information file, optical and synthetic aperture radar data of the corresponding position are downloaded; the optical and synthetic aperture radar data are derived from Sentinel 1 and Sentinel 2, as shown in Tables 3 and 4.

[0158] The data preprocessing module is used for preprocessing the acquired data and constructing a training data set; optical data preprocessing: according to the time of tree height and biomass data, the image with a time deviation of ≤15 days is screened, and if there is no data on the same day, the image closest in time is selected; according to the QA_PIXEL wave band, the pixel points of the remote sensing image are masked, and the Fmask algorithm is used to complete the image cloud removal work;

[0159] Synthetic aperture radar data preprocessing: according to the time of tree height and biomass data, the image corresponding to the time is screened; if there is no corresponding time on the same day, the closest time is selected; the radar observation data is converted into backscattering coefficient through radiation calibration, the original data information is dB value, and the dB value needs to be converted into intensity information, the formula is as follows:

[0160] ,

[0161] Lee filtering is used to process and remove spot noise in the image, while retaining image details and reducing noise influence;

[0162] Training dataset making: the tree height, biomass, optical and synthetic aperture radar data are cut into multiple data sets according to the size of 300*300 pixels, and the data sets without valid true value of tree height and biomass are removed to ensure the quality of the data set for model training; through random sampling, the screened data is divided into training set, test set and validation set according to the ratio of 40%:30%:30%; wherein the tree height and biomass data are the labels of the multi-task learning model, and the optical and synthetic aperture radar data are the input features of the multi-task learning model.

[0163] The collaborative inversion model construction module is used to establish a multi-task learning model based on an encoder and a decoder double focus, train the training data set through multi-task learning, and generate a collaborative inversion model of tree height and biomass; the multi-task learning model includes an encoder focus module and a decoder focus module, wherein the input data of the training data set is input into a residual fusion encoding network in the encoder focus module to generate tree height branch features and biomass branch features respectively, and the tree height branch features and the biomass branch features are input into a double-task feature interaction network in the decoder focus module to output tree height inversion results and biomass inversion results.

[0164] The residual fusion encoding network includes two parallel branch structures of a tree height encoder and a biomass encoder, each branch passes through 5 residual modules in turn, the convolution kernel size of the residual module decreases from large to small, and the channel number doubles per module; after each residual module is processed, a cross-branch interaction module is accessed, the cross-branch interaction module adds the channel of the feature maps output by the upper and lower branch residual modules, and then performs a random shuffling operation on the added feature maps; after 5 times of residual module and cross-branch interaction module processing, the tree height branch features and the biomass branch features are output by inputting into a pyramid pooling module for multi-scale pooling operation.

[0165] The double-task feature interaction network includes two parallel branch structures of a tree height decoder and a biomass decoder, each branch passes through 3 groups of residual modules and up-sampling operations to obtain tree height intermediate features and biomass intermediate features, and then the tree height intermediate features and the biomass intermediate features are input into a feature interaction module for cross-task feature fusion to output the final tree height result and biomass result.

[0166] The input data is taken from the training data set, and the input data includes optical data and synthetic aperture radar data, wherein the optical data is 11 band information extracted from Sentinel-2 in Table 4, and the synthetic aperture radar data is the radar backscattering coefficient of 2 polarization modes of Sentinel-1 image in Table 3, and the input data size is uniformly regularized to 300*300*13.

[0167] The first residual module uses a 7*7 convolution kernel, and the structure is as follows Figure 4As shown, the second residual module adopts a 5x5 convolution kernel, the third residual module adopts a 3x3 convolution kernel, and the fourth and fifth residual modules both adopt a 1x1 convolution kernel; the feature image size is unchanged in the five residual modules, and the feature image size is 300x300, and the number of feature channels is doubled in turn, being 26, 52, 104, 208 and 416.

[0168] As shown in FIG. 1, the input feature of the first residual module sequentially passes through a 7x7 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 7x7 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result. Figure 4

[0169] The input feature of the second residual module sequentially passes through a 5x5 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 5x5 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result.

[0170] The input feature of the third residual module sequentially passes through a 3x3 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 3x3 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel, and then passes through a ReLU activation function to output the result.

[0171] The input feature of the fourth and fifth residual modules sequentially passes through a 1x1 convolution kernel Conv, a normalization layer BN, a ReLU activation function, a 1x1 convolution kernel Conv and a normalization layer BN, is summed with the input feature passing through only a 1x1 convolution kernel Conv, and then passes through a ReLU activation function to output the result.

[0172] As shown in FIG. 1, to realize feature fusion between tree height and biomass branches, the cross-branch interaction module first adds the two branch features in the channel dimension to obtain the fused feature Figure 5 This operation directly fuses the two branch features in the channel dimension and shares complementary information; wherein is the output of the first residual module of the tree height branch, is the output of the first residual module of the biomass branch, wherein i represents the number of channels of the output of the first residual module; and i i ​​​​​​

[0173] ,

[0174] wherein, is a random permutation function of channel index, which ensures the dynamic change of channel order each time of training, and improves the model generalization.

[0175] After five residual modules and cross-branch interaction modules are processed, considering that the spot range of spaceborne lidar data is 25 meters, the spatial resolution of Sentinel 1 and Sentinel 2 is 10 meters, and the resolution of spaceborne lidar is about 3 times of optical and synthetic aperture radar data, in order to realize the spatial matching of input features and label information, an input pyramid pooling module is introduced, and a pyramid pooling operation is introduced.

[0176] As shown in Figure 6 , the pyramid pooling module includes a 3x3 pooling kernel, a 15x15 pooling kernel, a 30x30 pooling kernel and a global pooling kernel, and the pyramid pooling module performs multi-scale down-sampling on the input feature , and performs average pooling to obtain multi-scale pooling features:

[0177] ,

[0178] ,

[0179] ,

[0180] ,

[0181] wherein, is the input feature, , , , ;

[0182] In order to keep the number of channels unchanged, the multi-scale pooling features are fused in the channel dimension. First, the bilinear up-sampling is performed on , and respectively, so that the image size is 100x100, and , and are obtained, then the channel dimension is added, and the fused feature is obtained, which is expressed as:

[0183] ,

[0184] The number of channels of the fused feature is still 416, and the size is 100x100. Finally, the tree height branch feature and the biomass branch feature As the input of the decoder focusing module, the feature extraction process of the residual fusion coding network is completed.

[0185] In order to realize the matching of input features (optical data and synthetic aperture radar data) and label information (tree height and biomass) in the spatial dimension, a pyramid pooling operation is adopted to reduce the feature image size to 1 / 3 of the original size while keeping the number of feature channels unchanged, and finally output the tree height branch feature and the biomass branch feature.

[0186] The input data of the double-task feature interaction network is the tree height branch feature generated by the encoder focusing module and the biomass branch feature , and the two groups of feature data are processed in parallel; the structure of the double-task feature interaction network is shown in Figure 7 , that is, the input data is the tree height branch feature and the biomass branch feature generated by the encoder focusing module, and the two groups of data are input into the tree height decoder and the biomass decoder two parallel branches, and the processing process is started in parallel.

[0187] Firstly, the data sequentially pass through three groups of residual modules and up-sampling operations. After the residual module processing, the feature image size remains unchanged, and the number of feature channels decreases. Since the up-sampling operation will gradually increase the feature image size, different scale features are effectively extracted, and the convolution kernel size used by the residual module is set to increase from small to large. After three up-sampling operations, the feature image size is restored to the original input size.

[0188] Both the tree height decoder and the biomass decoder branch sequentially pass through three groups of residual modules and up-sampling operations. Since the up-sampling operation will increase the feature image size, in order to ensure that the receptive field is approximately equal, the convolution kernel used by the residual module is increased in turn, which is 3×3, 5×5 and 7×7 in turn. After the residual module, the feature image size remains unchanged, and the feature decreases in turn, which is 200, 100 and 50 in turn. After each residual module, the up-sampling operation is followed, and the feature image size increases in turn, which is 150×150, 200×200 and 300×300 in turn, restoring to the original input size, and the number of feature channels remains unchanged.

[0189] After three groups of residual modules and up-sampling operations, the intermediate features of the tree height decoder and the biomass decoder are obtained as and , respectively, which enter the feature interaction module to learn the cross-task dependency relationship between the tree height and the biomass task.

[0190] As shown in Figure 8 , in the feature interaction module, the tree height intermediate feature and the biomass intermediate feature are spliced according to the channel dimension to construct the joint feature and wherein represents a channel dimension concatenation operation;

[0191] and then the learnable weight matrix is used to calculate the dependency weight of the tree height task on the biomass task and the dependency weight of the biomass task on the tree height task The specific calculation formula is as follows:

[0192] ,

[0193] ,

[0194] wherein the softmax function ensures the normalization of the weight, so that reflects the dependency proportion of the tree height task on the biomass task feature, reflects the dependency proportion of the biomass task on the tree height task feature;

[0195] Based on the dependency weight, the tree height and biomass task features are cross-task fused:

[0196] ,

[0197] ,

[0198] wherein and are the intermediate features of the tree height task decoder and the biomass task decoder, and the channel concatenation is used to form the joint feature in the feature interaction module, and are learnable weight matrices, and are the feature vectors of the tree height task and the biomass task after cross-task interaction processing; finally, the full connection layer is used to complete the output of the tree height and biomass inversion results.

[0199] In the present application, the encoder and decoder double focusing strategy is optimized by using a multi-task joint loss function, wherein the loss function is composed of: using mean square error (MSE) as the basic loss, combining with the task uncertainty weighting mechanism to dynamically adjust the weight of the tree height loss and the biomass loss The total loss L is calculated by the formula:

[0200] ,

[0201] wherein, is the tree height task uncertainty parameter, The biomass task uncertainty parameter is automatically learned through the network; The L2 regularization term is a parameter of the encoder, The regularization coefficient is a parameter of the encoder.

[0202] Optimization strategy: Adam optimizer is used, the initial learning rate is set to 1e-4, and it is decayed by 0.8 times after every 50 training rounds; the training batch size is set to 32, and the iteration is performed for 200 rounds until the loss function converges, and the convergence threshold is the loss fluctuation amplitude.

[0203] The collaborative inversion prediction module is used to obtain optical and synthetic aperture radar data in the research area, and input the collaborative inversion model of tree height and biomass to perform collaborative inversion of tree height and biomass, and compare the predicted value with the true value to evaluate the inversion accuracy.

[0204] The root mean square error (RMSE) is used as the accuracy evaluation index, and the formula is as follows:

[0205] ,

[0206] In the formula, n is the total number of samples, is the model predicted value, is the true value, and the range of the root mean square error is When the predicted value is completely consistent with the true value, it is equal to 0, i.e. perfect prediction; the worse the prediction effect is, the larger the root mean square error value is.

[0207] Embodiment 3: An electronic device in this embodiment includes a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the steps of the method as described above.

[0208] Embodiment 4: A computer-readable storage medium in this embodiment has a computer program stored thereon, which is executed by a processor to implement the steps of the method as described above.

Claims

1. A method for simultaneous inversion of tree height and biomass based on codec dual-focusing, characterized in that, The method comprises the following steps: (1) obtaining tree height and biomass data of a target area by using a spaceborne lidar, and obtaining optical and synthetic aperture radar data of a corresponding position; (2) preprocessing the obtained data to construct a training data set; (3) establishing a multi-task learning model based on an encoder and a decoder double focus, training the training data set through multi-task learning, and generating a collaborative inversion model of tree height and biomass; The multi-task learning model comprises an encoder focus module and a decoder focus module, wherein the input data of the training data set are input into a residual fusion encoding network in the encoder focus module to generate tree height branch features and biomass branch features respectively, and the tree height branch features and the biomass branch features are input into a double-task feature interaction network in the decoder focus module to output tree height inversion results and biomass inversion results; The residual fusion encoding network comprises two parallel branch structures of a tree height encoder and a biomass encoder, the two parallel branch structures of the tree height encoder and the biomass encoder sequentially pass through five residual modules, the convolution kernel size of the residual modules decreases from large to small, and the channel number doubles per module; after each residual module is processed, a cross-branch interaction module is accessed, the feature maps output by the residual modules of the two parallel branch structures of the tree height encoder and the biomass encoder in the cross-branch interaction module are added in the channel, and then the added feature maps are randomly shuffled; after being processed by the five residual modules and the cross-branch interaction module, the tree height branch features and the biomass branch features are output by inputting into pyramid pooling modules for multi-scale pooling operations; The double-task feature interaction network comprises two parallel branch structures of a tree height decoder and a biomass decoder, the two parallel branch structures of the tree height decoder and the biomass decoder sequentially pass through three groups of residual modules and up-sampling operations to obtain tree height intermediate features and biomass intermediate features, and then the tree height intermediate features and the biomass intermediate features are input into a feature interaction module for cross-task feature fusion to output final tree height results and biomass results; (4) obtaining optical and synthetic aperture radar data in a research area, inputting the data into the collaborative inversion model of tree height and biomass for collaborative inversion of tree height and biomass, comparing the predicted values with the true values, and evaluating the inversion accuracy.

2. The method according to claim 1, wherein the method is characterized by: The step (1) specifically comprises the following steps: (1.1) downloading spaceborne lidar GEDI data of a target area, obtaining tree height data from a second-level product of the spaceborne lidar, obtaining biomass data from a fourth-level product of the spaceborne lidar, reading L2A and L4A files of the spaceborne lidar, extracting tree height, biomass data, longitude and latitude of a spot of the spaceborne lidar, original waveform data and elevation information, and screening the spot by using a screening condition, wherein the screening condition is: 1) when the sensitivity is less than 0.95, the corresponding spot is removed; 2) when the elevation difference between the GEDI elevation and the SRTM elevation is greater than 30 m, the corresponding spot is removed; 3) When the degradation identifier degrade>0, the obsolescence identifier stale_return_flag=0, the degradation identifier degrade_flag=1, and the quality identifier quality_flag=0, the corresponding light spot is removed; 4) When the error condition identifier rx_assess_flag=1, the corresponding light spot is removed; (1.2) A buffer zone is added to the latitude and longitude information to generate a vector information file of the corresponding position; (1.3) According to the vector information file, optical and synthetic aperture radar data of the corresponding position are downloaded.

3. The method according to claim 1, wherein: The step (2) specifically comprises the following steps: (2.1) Optical data preprocessing: according to the time of tree height and biomass data, images with a time deviation of less than or equal to 15 days are selected, and if there is no data on the same day, the image closest in time is selected; according to the QA_PIXEL waveband, the pixel points of the remote sensing image are masked, and then the image cloud removal work is completed; (2.2) Synthetic aperture radar data preprocessing: according to the time of tree height and biomass data, images corresponding to the time are selected; If there is no corresponding time on the same day, the closest time is selected; The radar observation data is converted into backscatter coefficient through radiation calibration, the original data information dB value is converted into intensity information, and the spot noise in the image is removed; (2.3) Training dataset making: the tree height, biomass, optical and synthetic aperture radar data are cut into multiple data groups according to the size of 300x300 pixels, and the data groups without valid true value of tree height and biomass are removed; through random sampling, the screened data is divided into training set, test set and validation set according to the ratio of 40:30:

30.

4. The method according to claim 1, wherein: The input data in the step (3) is taken from the training dataset, and the input data includes optical data and synthetic aperture radar data, wherein the optical data includes 11 waveband information extracted from the Sentinel 2 image, and the synthetic aperture radar data includes radar backscatter coefficients of 2 polarization modes extracted from the Sentinel 1 image, and the input data size is uniformly regularized to 300x300x13.

5. The method of claim 1, wherein: The cross-branch interaction module first adds the two branch features in the channel dimension to obtain the fused feature wherein is the output of the first i residual module of the tree height branch, is the output of the first i residual module of the biomass branch, wherein represents the number of channels of the output of the first i residual module. Again Randomly shuffling channels can be represented as randomly rearranging the channel indices: ， wherein, is a random permutation function of channel indices.

6. The method according to claim 1, wherein: The pyramid pooling module includes a 3*3 pooling kernel, a 15*15 pooling kernel, a 30*30 pooling kernel, and a global pooling kernel, and the pyramid pooling module performs multi-scale down-sampling on the input features to obtain multi-scale pooling features by performing average pooling, and is represented as: ， ， ， ， wherein is an input feature, , , , ; Channel dimension fusion is performed on the multi-scale pooling features again, that is, first, bilinear up-sampling is performed on , and respectively, so that the image size is 100*100, to obtain , and , and then channel dimension addition is performed to obtain the fused features , which is expressed as: ， Final output tree height branch feature With biomass branch feature .

7. The method according to claim 1, wherein: The feature interaction module in the tree middle feature and the biomass middle feature Concatenate in the channel dimension to construct the joint feature and wherein represents a channel dimension concatenation operation; Through the learnable weight matrix , 100 is the number of joint feature channels, 50 is the number of single task feature channels, the dependency weight of the tree height task on the biomass task is calculated And the dependency weight of the biomass task on the tree height task The calculation formula is: ， ， Wherein, softmax is an activation function, and based on the dependence weight, the tree height and biomass intermediate features are cross-task fused: ， ， wherein is the tree height post-fusion feature, is the biomass post-fusion feature, and finally the output of the tree height and biomass inversion results is completed through the fully connected layer.

8. A codec dual-focusing based tree height and biomass co-inversion system, characterized in that, Including: A data acquisition module is used to acquire tree height and biomass data of a target area by using a spaceborne laser radar, and optical and synthetic aperture radar data of a corresponding position are acquired; A data preprocessing module is used to preprocess the acquired data and construct a training dataset; A collaborative inversion model construction module is used to establish a multi-task learning model based on an encoder and a decoder double focus, train the training dataset through multi-task learning, and generate a collaborative inversion model of tree height and biomass; The multi-task learning model includes an encoder focusing module and a decoder focusing module, wherein the input data of the training dataset is input into a residual fusion encoding network in the encoder focusing module to generate tree height branch features and biomass branch features respectively, and the tree height branch features and the biomass branch features are input into a double-task feature interaction network in the decoder focusing module to output tree height inversion results and biomass inversion results; The residual fusion coding network comprises two parallel branch structures of a tree height encoder and a biomass encoder, the two parallel branch structures of the tree height encoder and the biomass encoder sequentially pass through five residual modules, the convolution kernel size of the residual modules decreases from large to small, and the number of channels doubles every module; after processing of each residual module, a cross-branch interaction module is accessed, the cross-branch interaction module performs channel addition operation on the feature maps output by the residual modules of the two parallel branch structures of the tree height encoder and the biomass encoder, and then performs random shuffling operation on the added feature maps; after processing of the five residual modules and the cross-branch interaction module, the tree height intermediate features and the biomass intermediate features are obtained by inputting the tree height intermediate features and the biomass intermediate features into pyramid pooling modules for multi-scale pooling operation, and finally the tree height branch features and the biomass branch features are output; The dual-task feature interaction network comprises two parallel branch structures of a tree height decoder and a biomass decoder, the two parallel branch structures of the tree height decoder and the biomass decoder sequentially pass through three groups of residual modules and upsampling operations to obtain tree height intermediate features and biomass intermediate features, the tree height intermediate features and the biomass intermediate features are input into a feature interaction module for cross-task feature fusion, and finally the tree height result and the biomass result are output; The collaborative inversion prediction module is configured to acquire optical and synthetic aperture radar data in a research area, input the data into a collaborative inversion model of tree height and biomass, perform collaborative inversion of the tree height and the biomass, compare predicted values with true values, and evaluate inversion accuracy.

9. An electronic device, comprising: The device comprises a processor and a storage medium. The storage medium is configured to store instructions. The processor is configured to operate according to the instructions to perform steps of the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is configured to be executed by the processor to perform steps of the method of any one of claims 1-7.

Images