Collaborative learning remote sensing biomass inversion method based on deep feature extraction
By constructing the CoModel model, using deep feature extraction and collaborative learning methods, the problem of insufficient data in complex areas of remote sensing biomass inversion is solved, and a higher precision and more accurate biomass inversion effect is achieved.
Patent Information
- Application Number
- CN202510431831.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-08-08
AI Technical Summary
The existing remote sensing biomass inversion methods cannot achieve accurate biomass inversion in township areas where forest distribution is relatively discrete and tree species is rich in diversity, especially when the data volume is insufficient.
The CoModel model is constructed by using a collaborative learning method based on deep feature extraction, and the data set is expanded by using pseudo-labels, and collaborative learning is carried out in different distance spaces to generate high-dimensional semantic feature vectors, and a biomass distribution map is generated by combining sliding windows and matrix operations.
Improve the accuracy and accuracy of biomass inversion, especially in complex geographical areas, achieving higher biomass prediction accuracy and spatial distribution accuracy.
Smart Images
Figure CN120449116A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision and deep learning, and specifically provides a collaborative learning remote sensing biomass inversion method based on deep feature extraction. Background Art
[0002] The Earth's terrestrial ecosystems absorb approximately 8 Gt of CO2 annually, with forest ecosystems making the primary contribution. Forests play a crucial role in carbon neutrality. Due to the importance of forest dynamics in the carbon cycle, monitoring their biomass is essential. However, existing biomass inversion methods are mostly based on traditional machine learning methods. While these methods have proven effective for inverting forest biomass over large, continuous areas, they still have limitations in rural-urban areas where forests are more discretely distributed and tree species are diverse.
[0003] The remote sensing biomass inversion method based on deep learning technology is a technology that can perform accurate biomass inversion in complex remote sensing scenarios.
[0004] Existing remote sensing biomass inversion methods, due to limited data (requiring manual field collection), generally rely on traditional machine learning methods (such as SVM, RF, KNN, etc.) and their improved methods. This cannot guarantee accurate biomass inversion in complex geographical areas with a combination of towns and villages. How to ensure accurate biomass inversion for forests with diverse, multi-layered, and age-varying characteristics distributed in mixed urban and rural areas, based on the multi-channel data sources provided by multi-source remote sensing, is an urgent problem to be solved. Summary of the Invention
[0005] In order to solve the above problems, the purpose of the present invention is to provide a collaborative learning remote sensing biomass inversion method based on deep feature extraction.
[0006] A collaborative learning remote sensing biomass inversion method based on deep feature extraction, the method comprising the following steps:
[0007] S1: Construct a remote sensing-biomass dataset and then build the basic backbone network CoModel;
[0008] S2: Build a collaborative semi-supervised learning framework. Initialize and train the deep feature extraction part of the CoModel to obtain a set of high-dimensional semantic feature vectors. Use these vectors and the corresponding biomass labels to initialize the KNN regressors in two sets of CoModels using different distance space metrics. Each KNN regressor uses its own independent dataset.
[0009] S3: Performing a semi-supervised training step, i.e., the process of each regressor adding the best pseudo-label to the other's dataset; the process involves initializing the unlabeled data pool, calculating the optimized distance between the pseudo-label and the true value, determining the best pseudo-label based on the optimized distance, and adding the best pseudo-label calculated by each regressor to the other's training dataset;
[0010] S4: After n times of S3, no more than n best pseudo labels are added to the training set of each KNN regressor. The newly added pseudo labels are used together with the original biomass dataset for full training, and two CoModels are obtained for the final biomass prediction. The biomass prediction value is the average sum of the prediction results of the two CoModels.
[0011] S5: By using a sliding window on the multi-feature remote sensing images stacked in sequence, each pixel value used for prediction is obtained in turn. Based on the two CoModels obtained in step S4, predictions are made in sequence according to the pixel arrangement to obtain a complete biomass distribution map. The resolution and geographic coordinate system of the obtained biomass map are equivalent to those of the stacked remote sensing images.
[0012] According to step S1, constructing a remote sensing-biomass dataset and a basic backbone network includes the following steps:
[0013] S11: Construct a remote sensing-biomass dataset. First, obtain hyperspectral, L-SAR, and their respective calculated remote sensing index data (including NDVI, NDI34, RFDI, RVI, and DPI). The calculation formulas for each index are as follows:
[0014]
[0015] Among them, NIR represents the near-infrared band, RED represents the red band, GREEN represents the green band, and VH and VV represent the dual-polarization bands of the synthetic aperture radar. These data are then stacked in a fixed order in the channel direction and the pixel resolution is defined as 20m×20m. The training set data and the final biomass map will be generated based on this order. Then, based on the measured data of each tree in the sample plot and the binary biomass calculation formula, the total biomass data of each 20m×20m sample plot can be obtained. Finally, according to the recorded sample plot center coordinate values, the 3×3 area corresponding to the coordinate position of the stacked remote sensing image is cropped and used as the input for training the model;
[0016] S12: Constructing the basic backbone network CoModel refers to building a basic model for reasoning operations. The model can be divided into two parts: the remote sensing information deep feature extraction layer and the KNN reasoning layer.
[0017] In step S1, a CoModel is constructed that includes a remote sensing information deep feature extraction layer and a KNN inference layer, including the following steps:
[0018] S121: Constructing a deep feature extraction layer. This involves adding a ResNet 18 model with an intermediate layer before the final fully connected layer. The output of this intermediate layer is extracted and used as a high-dimensional semantic feature vector. During training, this part replaces the output of the final fully connected layer with a 1-dimensional size to obtain the biomass regression results used to calculate the loss. The loss is calculated using the mean squared error (MSE), which is calculated as follows:
[0019]
[0020] Where n is the total number of samples, that is, the number of training batches, y i is the true value of the i-th sample, is the predicted value of the i-th sample;
[0021] S122: Construct a KNN inference layer. By establishing a distance metric space between high-dimensional semantic feature vectors and labels, the three nearest neighbor labels in the distance metric space are used to complete the biomass inference prediction of other remote sensing image coverage areas. At the same time, the three reference labels used in the inference process are saved for subsequent semi-supervised learning steps.
[0022] Constructing the semi-supervised learning framework in step S2 includes the following steps:
[0023] S21: CoModel initialization, using all the data to train the deep feature extraction model in step S1, and then initialize two KNN inference layers using different distance metric spaces;
[0024] S22: Initialize two corresponding training data sets for the two KNNs. In the initialization step, the two training sets are equal, and they will gradually change in the subsequent semi-supervised steps.
[0025] In step S3, semi-supervised training includes the following steps:
[0026] S31: Initialization of the unlabeled data pool: randomly select the center coordinates of 100 pixels in the inversion area, and crop the stacked remote sensing image according to the size of 3×3 pixels as the input data for generating candidate pseudo labels. Initialize two sets of unlabeled data pools to provide potential pseudo labels to the two KNNs.
[0027] S32: Each unlabeled cropped image data in the two unlabeled data pools is input into the deep feature extraction model in the CoModel initialized by step S21 and with frozen parameters, to obtain the high-dimensional semantic features output by the intermediate layer, and then obtain the two biomass prediction values output by the two KNNs and the three labels of the two nearest neighbors used as reference when predicting the biomass;
[0028] S33: Add the two predicted values to the two KNN datasets that predicted the values, and calculate the three labels of the two sets of nearest neighbors. Before and after the two predicted values are added to the training set, the distance between the two sets of values and their corresponding true values decreases. The distance is calculated using Euclidean distance. If it decreases, it means that the biomass value predicted by the current data is a potential pseudo-label that can be added.
[0029] S34: looping steps S32 and S33 until all the neighbor label shortening distances of all the unlabeled data initialized by S31 are calculated, sorting the two groups of 100 shortening distance calculation values, taking the two best unlabeled cropped data and their corresponding biomass pseudo labels from the two groups, and adding them to each other's KNN training set instead of their own training set;
[0030] S35: The process of adding a pair of pseudo labels is now completed, and the above steps are continued to be looped to further add pseudo labels.
[0031] In step S4, the obtained pseudo labels are used to train the deep feature extraction model in CoModel, including the following steps:
[0032] S41: Initialize the deep feature extraction model and KNN of the two CoModels, and also use the same distance metric space as step S21 to initialize KNN;
[0033] S42: Read two sets of pseudo-labeled datasets that have been summarized, and use the two sets of datasets to train two deep feature extraction models respectively. At the same time, during the training process, high-dimensional semantic features and their corresponding labels or pseudo-labels are continuously added to the KNN dataset;
[0034] S43: When predicting the final biomass result, the input cropped remote sensing image is passed through the two CoModels trained in step S42 to obtain the high-dimensional semantic features of the middle layer. The features are input into the KNN model in the CoModel to obtain two prediction results based on different distance metric spaces. The average of these two results is taken as the final prediction result.
[0035] Generating a biomass spatial distribution map in step S5 includes the following steps:
[0036] S51: define a sliding window of 3×3 pixels and slide it on the stacked remote sensing image;
[0037] S52: Slide row by row and set the pixel values outside the edge to 0, take the cropped image within each window, and obtain the biomass prediction value of the central pixel of the window according to step S43. The prediction values are spliced together to form a one-dimensional sequence;
[0038] S53: Through matrix operations, the sequence is restored into a two-dimensional matrix row by row, and a geographic coordinate system is generated based on the stacked remote sensing images to obtain a biomass spatial distribution map.
[0039] The advantages and beneficial effects of the present invention are as follows:
[0040] The present invention designs a CoModel model. The model first obtains a high-dimensional semantic feature vector used for KNN regression through a deep feature extraction method. Based on this vector, KNN can obtain more effective neighbor discrimination compared to when the original input features are used as a training set. Starting from the problem of too few biomass data sets, based on the neighbor characteristics of KNN, a semi-supervised learning method based on neighbor criterion is constructed. The number of data sets is further expanded through pseudo-labels, and a large amount of unlabeled data is fully utilized, so that the inversion of biomass is more accurate. Based on the collaborative learning method, a dual model is constructed to predict biomass in different distance spaces, thereby paying attention to the features of different perspectives. By taking the average sum of the predicted values of the two models, the final biomass prediction result is obtained. Compared with the traditional RF biomass inversion method, the accuracy of the present invention is higher. Finally, the production of the biomass spatial distribution map is completed based on matrix operations and geographic coordinate system replication. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is a flow chart of a collaborative learning remote sensing biomass inversion method based on deep feature extraction according to the present invention;
[0042] Figure 2 The basic backbone network CoModel described in the present invention;
[0043] Figure 3 This is the semi-supervised learning framework structure diagram of the present invention, where CoModel refers to Figure 2 said network;
[0044] Figure 4 The left figure is the result of the random forest test, and the right figure is the result of the test of the method proposed in this invention;
[0045] Figure 5 The generated result of the biomass spatial distribution map of the present invention;
[0046] Figure 6 This is a flowchart of the steps of a collaborative learning remote sensing biomass inversion method based on deep feature extraction described in the present invention. DETAILED DESCRIPTION
[0047] The collaborative learning remote sensing biomass inversion method based on deep feature extraction provided by the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] See also Figure 1 As shown, a collaborative learning remote sensing biomass inversion method based on deep feature extraction includes the following steps:
[0049] S1: Construct a remote sensing-biomass dataset and then build the basic backbone network CoModel;
[0050] Specifically, constructing the remote sensing-biomass dataset includes the following steps:
[0051] S111: Obtain hyperspectral, L-SAR, and their calculated remote sensing index data (including NDVI, NDI34, RFDI, RVI, and DPI). The calculation formulas for each index are as follows:
[0052]
[0053] Among them, NIR represents the near infrared band, RED represents the red light band, GREEN represents the green light band, and VH and VV represent the dual-polarization bands of synthetic aperture radar.
[0054] S112: These data are then stacked in a fixed order in the channel direction and the pixel resolution is defined as 20m×20m, which is consistent with the measured size of the sample plot. The measured data of the sample plot record the tree height and diameter at breast height of each tree in the sample plot, and are calculated as biomass using the binary biomass calculation formula. The total biomass of a sample plot is obtained by accumulation;
[0055] S113: Based on the total biomass data of each plot, according to the recorded coordinate values of the plot center, a 3×3 area of the corresponding coordinate position of the stacked remote sensing image is cropped and used as input for training the model.
[0056] Specifically, see Figure 2 ,Constructing the basic backbone network CoModel includes the following steps:
[0057] S121: Constructs a deep feature extraction layer for remote sensing information. This involves adding a ResNet 18 model with an intermediate layer before the final fully connected layer. The output of this intermediate layer is extracted and used as a high-dimensional semantic feature vector. During training, this part replaces the output of the final fully connected layer with a 1-dimensional size to obtain the biomass regression results used to calculate loss. Loss is calculated using the mean squared error (MSE), which is calculated as follows:
[0058]
[0059] Where n is the total number of samples, that is, the number of training batches, y i is the true value of the i-th sample, is the predicted value of the i-th sample;
[0060] S122: Construct a KNN inference layer. By establishing a distance metric space between high-dimensional semantic feature vectors and labels, the three nearest neighbor labels in the distance metric space are used to complete the biomass inference prediction of other remote sensing image coverage areas. At the same time, the three reference labels used in the inference process are saved for subsequent semi-supervised learning steps.
[0061] S2: Constructing a collaborative semi-supervised learning framework, see Figure 3 ,By initializing and training CoModel, we obtain a set of high-dimensional semantic feature vectors, and ,using these vectors and the corresponding biomass labels, and ,at the same time, initialize two sets of KNN regressors using different distance ,space metrics schemes, and each KNN regressor uses its own ,independent dataset;
[0062] Specifically, building a semi-supervised learning framework involves initializing two CoModels to create a framework for subsequent collaborative training. Collaborative learning involves adding the best pseudo-labels calculated by two models based on different distance metric spaces based on neighboring reference labels to each other's training set.
[0063] S3: Performing a semi-supervised training step, i.e., the process of each regressor adding the best pseudo-label to the other's dataset; the process involves initializing the unlabeled data pool, calculating the optimized distance between the pseudo-label and the true value, determining the best pseudo-label based on the optimized distance, and adding the best pseudo-label calculated by each regressor to the other's training dataset;
[0064] Specifically, the semi-supervised training process includes the following steps:
[0065] S31: Initialization of the unlabeled data pool: randomly select the center coordinates of 100 pixels in the inversion area, and crop the stacked remote sensing image according to the size of 3×3 pixels as the input data for generating candidate pseudo labels. Initialize two sets of unlabeled data pools to provide potential pseudo labels to the two KNNs.
[0066] S32: Each unlabeled cropped image data in the two unlabeled data pools is input into the deep feature extraction model in the CoModel initialized by step S21 and with frozen parameters, to obtain the high-dimensional semantic features output by the intermediate layer, and then obtain the two biomass prediction values output by the two KNNs and the three labels of the two nearest neighbors used as reference when predicting the biomass;
[0067] S33: Add the two predicted values to the two KNN datasets that predicted the values, and calculate the three labels of the two sets of nearest neighbors. Before and after the two predicted values are added to the training set, the distance between the two sets of values and their corresponding true values decreases. The distance is calculated using Euclidean distance. If it decreases, it means that the biomass value predicted by the current data is a potential pseudo-label that can be added.
[0068] S34: looping steps S32 and S33 until all the neighbor label shortening distances of all the unlabeled data initialized by S31 are calculated, sorting the two groups of 100 shortening distance calculation values, taking the two best unlabeled cropped data and their corresponding biomass pseudo labels from the two groups, and adding them to each other's KNN training set instead of their own training set;
[0069] S35: The process of adding a pair of pseudo labels is now completed, and the above steps are continued to be looped to further add pseudo labels.
[0070] S4: After n times of S3, no more than n best pseudo labels are added to the training set of each regressor. The newly added pseudo labels are used together with the original biomass dataset for full training. Two CoModels are obtained for the final biomass prediction. The biomass prediction value is the average sum of the prediction results of the two models.
[0071] Specifically, the steps for complete training using the newly added pseudo labels are:
[0072] S41: Initialize two CoModels and use the same distance metric space as step S21 to initialize the KNN in the CoModel;
[0073] S42: Read two sets of pseudo-labeled datasets that have been summarized, and use them to train the deep feature extraction models in the two CoModels respectively. At the same time, during the training process, high-dimensional semantic features and their corresponding labels or pseudo-labels are continuously added to the KNN dataset;
[0074] S43: When predicting the final biomass result, the input cropped remote sensing image is subjected to the deep feature extraction model in the two CoModels trained in step S42 to obtain the high-dimensional semantic features of the middle layer. The features are input into the KNN model in the CoModel to obtain two prediction results based on different distance metric spaces. The average of these two results is taken as the final prediction result.
[0075] S5: By using a sliding window on the multi-feature remote sensing images stacked in sequence, each pixel value used for prediction is obtained in turn. Based on the two collaborative models obtained in step S4, predictions are made in sequence according to the pixel arrangement to obtain a complete biomass distribution map. The resolution and geographic coordinate system of the obtained biomass map are equivalent to those of the original remote sensing image.
[0076] Specifically, based on the sliding window and the trained model, the steps to generate the biomass spatial distribution map are as follows:
[0077] S51: define a sliding window of 3×3 pixels and slide it on the stacked remote sensing image;
[0078] S52: Slide row by row and set the pixel values outside the edge to 0, take the cropped image within each window, and obtain the biomass prediction value of the central pixel of the window according to step S43. The prediction values are spliced together to form a one-dimensional sequence;
[0079] S53: Through matrix operations, the sequence is restored into a two-dimensional matrix row by row, and a geographic coordinate system is generated based on the stacked remote sensing images to obtain a biomass spatial distribution map.
Claims
1. A collaborative learning remote sensing biomass inversion method based on deep feature extraction, characterized by: The following steps are involved: S1: Construct a remote sensing-biomass dataset and then build the basic backbone network CoModel; S2: Build a collaborative semi-supervised learning framework. Initialize and train the CoModel to obtain a set of high-dimensional semantic feature vectors. Use these vectors and the corresponding biomass labels to initialize two sets of KNN regressors using different distance space metrics. Each KNN regressor uses its own independent dataset. S3: Performing a semi-supervised training step, i.e., the process of each regressor adding the best pseudo-label to the other's dataset; the process involves initializing the unlabeled data pool, calculating the optimized distance between the pseudo-label and the true value, determining the best pseudo-label based on the optimized distance, and adding the best pseudo-label calculated by each regressor to the other's training dataset; S4: After n times of S3, no more than n best pseudo labels are added to the training set of each regressor. The newly added pseudo labels are used together with the original biomass dataset for full training. Two CoModels are obtained for the final biomass prediction. The biomass prediction value is the average sum of the prediction results of the two models. S5: By using a sliding window on the multi-feature remote sensing images stacked in sequence, each pixel value used for prediction is obtained in turn. Based on the two CoModels obtained in step S4, predictions are made in sequence according to the pixel arrangement to obtain a complete biomass distribution map. The resolution and geographic coordinate system of the obtained biomass map are equivalent to those of the original remote sensing image.
2. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: In step S1, constructing the remote sensing-biomass dataset and building the basic backbone network includes the following steps: S11: Construct a remote sensing-biomass dataset and stack it in a fixed order along the channel direction, defining a pixel resolution of 20m×20m. This order will be used to generate the training set data and the final biomass map. Based on the measured tree-per-tree measurement data and the binary biomass calculation formula, the total biomass data for each 20m×20m plot can be obtained. Finally, based on the recorded plot center coordinates, a 3×3 area corresponding to the coordinate position of the stacked remote sensing image is cropped and used as input for the training model. S12: Building a basic backbone network refers to building a basic model for reasoning operations - CoModel. This model can be divided into two parts: the remote sensing information deep feature extraction layer and the KNN reasoning layer.
3. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: In step S1, a CoModel is constructed that includes a remote sensing information deep feature extraction layer and a KNN inference layer, including the following steps: S121: Constructing a deep feature extraction layer. This involves adding an intermediate layer of ResNet18 model before the final fully connected layer. The output of this intermediate layer is extracted and used as a high-dimensional semantic feature vector. During training, this part replaces the output of the final fully connected layer with a 1-dimensional size to obtain the biomass regression results used to calculate the loss. The loss is calculated using the mean squared error (MSE). The MSE calculation formula is as follows: Where n is the total number of samples, that is, the number of training batches, y i is the true value of the i-th sample, is the predicted value of the i-th sample; S122: Construct a KNN inference layer, that is, use the K-nearest neighbor (KNN) algorithm to establish a distance metric space between high-dimensional semantic feature vectors and labels, and complete the biomass inference prediction of other remote sensing image coverage areas based on the k nearest neighbor labels in the distance metric space. At the same time, save the k reference labels used in the inference process for subsequent semi-supervised learning steps.
4. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: Constructing the semi-supervised learning framework in step S2 includes the following steps: S21: Initialize the deep feature extraction model, using all the data to train the CoModel in step S1, and then initialize two KNN inference layers in the CoModel using different distance metric spaces; S22: Initialize two corresponding training data sets for the two KNNs. In the initialization step, the two training sets are equal, and they will gradually change in the subsequent semi-supervised steps.
5. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: The semi-supervised training in step S3, i.e., the process of adding pseudo labels to the two datasets initialized in step S22, includes the following steps: S31: Initialization of the unlabeled data pool: randomly select the center coordinates of 100 pixels in the inversion area, and crop the stacked remote sensing image according to the size of 3×3 pixels as the input data for generating candidate pseudo labels. Initialize two sets of unlabeled data pools to provide potential pseudo labels to the two KNNs. S32: Each unlabeled cropped image data in the two unlabeled data pools is input into the CoModel initialized by step S21 and whose parameters have been frozen, to obtain the high-dimensional semantic features output by the intermediate layer, and then obtain the two biomass prediction values output by the two KNNs and the three labels of the two nearest neighbors used as reference when predicting the biomass; S33: Add the two predicted values and their corresponding high-dimensional semantic features to the two KNN datasets that predicted the values. Calculate the three labels of the two sets of nearest neighbors. Check whether the distance between the two sets of values and their corresponding true values decreases before and after the two predicted values are added to the training set. The distance is calculated using Euclidean distance. If it decreases, it means that the biomass value predicted by the current data is a potential pseudo-label that can be added. S34: looping steps S32 and S33 until all the neighbor label shortening distances of all the unlabeled data initialized by S31 are calculated, sorting the two groups of 100 shortening distance calculation values, taking the two best unlabeled cropped data and their corresponding biomass pseudo labels from the two groups, and adding them to each other's KNN training set instead of their own training set; S35: The process of adding a pair of pseudo labels is now completed, and the above steps are continued to be looped to further add pseudo labels.
6. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: Using the pseudo labels obtained in step S4, the CoModel is trained, including the following steps: S41: Initialize two CoModels and use the same distance metric space as step S21 to initialize KNN; S42: Read two sets of pseudo-labeled datasets that have been summarized, and use them to train the deep feature extraction parts of the two CoModels respectively. At the same time, during the training process, high-dimensional semantic features and their corresponding labels or pseudo-labels are continuously added to the KNN dataset in the CoModel; S43: When predicting the final biomass result, the input cropped remote sensing image is passed through the two CoModels trained in step S42 to obtain the high-dimensional semantic features of the middle layer. The features are input into the KNN model in the CoModel to obtain two prediction results based on different distance metric spaces. The average of these two results is taken as the final prediction result.
7. The collaborative learning remote sensing biomass inversion method based on deep feature extraction according to claim 1 is characterized by: Generating a biomass spatial distribution map in step S5 includes the following steps: S51: define a sliding window of 3×3 pixels and slide it on the stacked remote sensing image; S52: Slide row by row and set the pixel values outside the edge to 0, take the cropped image within each window, and obtain the biomass prediction value of the central pixel of the window according to step S43. The prediction values are spliced together to form a one-dimensional sequence; S53: Through matrix operations, the sequence is restored into a two-dimensional matrix row by row, and a geographic coordinate system is generated based on the stacked remote sensing images to obtain a biomass spatial distribution map.