Forest aboveground biomass estimation method based on satellite-borne laser radar
By combining DU-Net and Stacking ensemble learning models with sparse lidar and optical imagery, the accuracy and efficiency issues of traditional methods in biomass estimation in complex forest environments are solved, and high-precision biomass inversion is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-04-14
AI Technical Summary
Traditional regression methods struggle to capture complex nonlinear spatial features, while conventional deep learning methods heavily rely on massive amounts of labeled data and lack generalization ability in complex and heterogeneous forest environments. Existing methods are unable to achieve high-precision, high-resolution forest aboveground biomass inversion under conditions of scarce samples.
By combining the DU-Net model with the Stacking ensemble learning model, and utilizing sparse lidar data and high spatiotemporal resolution optical imagery, a high-precision estimation method for forest canopy height and biomass is constructed by pre-training with DINOv2 self-supervised learning technology and fusing features from multiple data sources.
High-precision and high-resolution forest aboveground biomass inversion was achieved under conditions of scarce samples, effectively alleviating the problem of insufficient generalization of single models in complex forest environments and improving the accuracy and efficiency of biomass estimation.
Smart Images
Figure CN121856928A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of quantitative remote sensing and relates to a method for estimating forest aboveground biomass based on spaceborne lidar. Background Technology
[0002] Global climate change has become a major challenge facing humanity. As the main body of the terrestrial biosphere, forest ecosystems play an irreplaceable role in maintaining global climate stability and regulating carbon balance. Traditional methods for measuring forest aboveground biomass mainly involve field quadrat surveys. Due to the wide distribution of forests and complex terrain, traditional ground-based measurement methods require significant manpower, material resources, and time. Remote sensing methods, with their large monitoring range, stable monitoring cycles, and rich information, offer significant advantages in forest aboveground biomass retrieval, being time-saving, labor-saving, and highly efficient.
[0003] Currently, spaceborne LiDAR is in a phase of rapid development. Representative spaceborne LiDAR satellites launched in recent years include the ICESat-2 (Advanced Topographic Laser Altimeter System) launched by NASA for monitoring glacier melting and the GEDI (Global Ecosystem Dynamics Investigation) for monitoring forest parameters, as well as the GF7 satellite launched by my country.
[0004] GEDI provides spatially dense vegetation height data; however, its discrete sampling along and across orbits prevents the direct acquisition of spatially continuous vegetation height products. However, spatially continuous vegetation height mapping can be achieved by fusing spatially continuous remote sensing data with GEDI vegetation height data through machine learning. Establishing the relationship between vegetation structure parameters and biomass using empirical regression models is a common method for estimating biomass based on full-waveform LiDAR and is the standard method for the GEDI04_A product, using GEDI vegetation height as input to estimate biomass. However, different regions have different dominant tree species, and using inapplicable vegetation height parameters cannot effectively estimate biomass. Furthermore, vegetation height has limited explanatory power for biomass, failing to explain the number and spatial distribution of trees within a light spot. Currently, there is still a lack of effective methods for extracting vertical and horizontal vegetation information from large light spot waveforms to estimate biomass. Moreover, although GEDI can achieve spatially continuous forest height and biomass mapping by fusing spatially continuous remote sensing data with GEDI vegetation height through machine learning, extrapolating spatially continuous forest height from optical imagery often results in saturation, thus affecting the accuracy of biomass mapping. Furthermore, GEDI contains a large number of light spots with poor geolocation accuracy, waveforms with poor signal quality, and invalid light spots affected by clouds and other surface conditions. Although GEDI includes data quality judgment parameters such as the quality flag parameter quality_flag and the sensitivity parameter sensitivity, in actual use, it is still necessary to combine factors such as terrain to filter out light spots with poor quality in order to improve the estimation quality of canopy height and biomass. Summary of the Invention
[0005] The technical problem solved by this invention is to overcome the difficulty of traditional regression methods in capturing complex nonlinear spatial features, the heavy reliance of conventional deep learning methods on massive labeled data, and the insufficient generalization ability of single models when facing complex and heterogeneous forest environments. This invention proposes a method that can effectively integrate sparse lidar data with high spatiotemporal resolution optical images to achieve high-precision and high-resolution forest aboveground biomass inversion under the condition of scarce samples.
[0006] The solution of the present invention is: A method for estimating forest aboveground biomass based on spaceborne lidar includes: Preprocessing of GEDI L2A and L4A data was performed to extract forest canopy height and biomass information; Preprocess the Sentinel-1 GRD data; Cloud removal from Sentinel-2 Level-2A images; The preprocessed GEDI data, Sentinel-1 images, and Sentinel-2 images were segmented according to the boundary of the study area, and the data were divided into training set and validation set according to an 8:2 structure. Constructing a DU-Net model to invert forest canopy height; A Stacking ensemble learning model is constructed, which integrates multi-source prediction results through a meta-learner to achieve high-precision inversion of forest aboveground biomass; The constructed DU-Net model and Stacking ensemble learning model were trained and validated using the training and validation sets; The preprocessed real-time satellite data is input into the trained DU-Net model and Stacking ensemble learning model to obtain the forest aboveground biomass result map.
[0007] Preferably, the GEDI L2A data is preprocessed using the following method: GEDI L2A data with the following characteristics were removed: quality_flag ≠ 1, degrade_flag ≠ 0, Sensitivity < 0.95, num_detectedmodes ≥ 15, solar_elevation ≥ 0, rx_assess_flag ≠ 0, |elev_lowestmode –SRTM DEM| > 40, |elev_lowestmode – TanDEM-X| > 50, ground height difference extracted by each algorithm setting > 2m, and slope > 20 degrees. Here, quality_flag is a quality flag (1 indicates acceptable data quality), degrade_flag is a degradation flag (0 indicates normal, 1 indicates degraded data), sensitivity is the probability of the laser beam penetrating the canopy and detecting the ground, num_detectedmodes is the number of detected peaks, solar_elevation is the solar altitude angle, rx_assess_flag is the receiver evaluation flag, and |elev_lowestmode – SRTMDEM| is the height difference between the ground height measured by GEDI and the SRTM elevation data, and |elev_lowestmode - TanDEM-X| is the height difference between the ground height measured by GEDI and the TanDEM-X elevation data. Add back GEDI L2A data that differ from the canopy height in "Global Forest Canopy Height, 2019 | GLAD" by less than 2 meters; For the GEDI L2A data processed above, RH95 is calculated, and after removing the maximum and minimum values of RH95, the average value of the remaining RH95 is calculated. This average value is used as the forest canopy height of the light spot range. Light spots with a canopy height of less than 3m are considered non-forest and removed.
[0008] Preferably, the GEDIL4A data is preprocessed using the following method: For GEDI L4A data, only data with the same spot size as the preprocessed GEDI L2A data are retained, and data with l4_quality_flag ≠ 1 are deleted. Then, the agbd attribute is extracted as the point-scale biomass, where l4_quality_flag is a quality flag bit, and 1 indicates that the data quality is acceptable.
[0009] Preferably, the Sentinel-1 GRD data is preprocessed, including the following steps: orbit correction, radiometric calibration, multi-view processing, Refined Lee filtering, terrain correction, decibel conversion, and cropping.
[0010] Preferably, the method for cloud removal from Sentinel-2 Level-2A images is as follows: Pixels classified as "Clouds Low Probability", "Clouds Medium Probability", "Clouds High Probability", and "CloudShadows" are removed based on the "Scene Classification" band built into Sentinel-2. The Sentinel-2 images from every two months are sorted and stitched in descending order according to the "CLOUDY_PIXEL_PERCENTAGE" attribute.
[0011] Preferably, the DU-Net model is constructed to invert forest canopy height, as follows: The Sentinel-1 and Sentinel-2 images were segmented into 256x256 pixel images, and DINOv2 self-supervised learning was used to obtain an image encoder specifically for vegetation images. The image encoder is divided into two versions: a teacher model and a student model. The teacher model receives a global view, while the student model receives both local and global views. The teacher model and the student model are jointly trained to output similar feature representations, thereby obtaining a pre-trained image encoder; The pre-trained teacher model weights are transferred to the encoder part of the U-Net model and frozen to construct a trainable decoder part. Using preprocessed GEDI data as labels and Sentinel images as input, the decoder of the DU-Net model is trained under supervision to generate a forest canopy height map.
[0012] Preferably, forest aboveground biomass is retrieved using a Stacking ensemble learning model, as follows: Features were extracted from the preprocessed Sentinel-1 and Sentinel-2 images, including band features (red, green, blue, near-infrared, VH, VV), vegetation index features (GNDVI, NDVI, DVI, RVI, SAVI), feature variables (VH+VV, VH-VV, VH / VV), and texture features for each band extracted using windows of 3x3, 5x5, 7x7, and 9x9. Select the features most useful for prediction from all the features and feature variables mentioned above; Using the selected feature variables and the forest canopy height obtained by inversion from the DU-Net model as input features, multiple differentiated base learners are trained, and a linear regression model is used as a meta-learner to fuse the prediction results of the base learners, thereby constructing a Stacking ensemble learning model to predict forest aboveground biomass.
[0013] The preferred formulas for calculating vegetation indices GNDVI, NDVI, DVI, RVI, and SAVI are as follows:
[0014]
[0015]
[0016]
[0017]
[0018] NIR represents the near-infrared band, Green represents the green light band, and Red represents the red light band.
[0019] Preferably, the texture features include:
[0020]
[0021]
[0022]
[0023]
[0024]
[0025] in, This represents the mean. The number of elements in the normalized gray-level co-occurrence matrix represents the number of elements in the normalized gray-level co- line, number The element values of the column, This represents the number of quantization levels for the image's grayscale. and This represents the grayscale index of a pixel in the image, and its value range is... arrive .
[0026] Preferably, the features most useful for prediction are selected, as follows: The Boruta algorithm is used to evaluate which original feature variables are more informative than image features by repeatedly training the random forest model, thereby selecting the feature variables most useful for prediction.
[0027] The advantages of this invention compared to the prior art are: This invention introduces DINOv2 self-supervised learning technology, which makes full use of massive unlabeled satellite imagery for pre-training. This enables the model to extract robust vegetation spatial features even under GEDI sparse sample conditions, which is significantly better than traditional deep learning models trained from scratch.
[0028] This invention innovatively integrates "optical / radar features + canopy height" to predict forest aboveground biomass. By first inverting tree height using DU-Net, vertical structure information is explicitly introduced into the biomass estimation model, effectively alleviating the "signal saturation" problem in high biomass areas caused by relying solely on optical images.
[0029] This invention employs a stacking ensemble learning strategy, organically integrating the advantages of different base learners. Compared to a single model, this method can more effectively capture nonlinear relationships in the data, reduce model variance and overfitting risk, and improve prediction accuracy in complex and heterogeneous forests. This approach enables the efficient acquisition of large-scale and high-resolution forest aboveground biomass data, bringing significant achievements to fields such as climate change monitoring. Attached Figure Description
[0030] Figure 1 This is a flowchart illustrating a method for estimating forest aboveground biomass based on spaceborne lidar. Detailed Implementation
[0031] The present invention will be further described below with reference to the embodiments.
[0032] This invention discloses a method for estimating forest aboveground biomass based on spaceborne lidar, comprising the following steps: Using the relevant parameters built into GEDI and making full use of existing forest canopy height products, abnormal GEDI spots were identified and filtered out. A high-quality set of true forest canopy height values was constructed using the RH95 index. The Sentinel-1 / 2 data was preprocessed.
[0033] Using massive amounts of unlabeled Sentinel imagery, a teacher-student model was trained based on the DINOv2 self-supervised learning framework to obtain a vegetation image encoder with strong generalization ability. The encoder weights were frozen and transferred to the U-Net model. The decoder was then fine-tuned using preprocessed GEDI spot data to construct the DU-Net model to generate continuous forest canopy height maps.
[0034] Features were extracted from the preprocessed Sentinel-1 and Sentinel-2 images, including band features (red, green, blue, near-infrared, VH, VV), vegetation index features (GNDVI, NDVI, DVI, RVI, SAVI), feature variables (VH+VV, VH-VV, VH / VV), as well as the minimum, maximum, median, mean, and standard deviation of the band features, vegetation index features, and feature variables in the time series, and texture features for each band extracted using 3x3, 5x5, 7x7, and 9x9 windows. Using the canopy height map retrieved from DU-Net and selected multidimensional remote sensing features as input, and GEDI L4A biomass data as labels, a Stacking ensemble learning model containing multiple differential base learners is constructed. Linear regression is used as the meta-learner to fuse the prediction results of the base learners, ultimately achieving high-precision estimation of forest aboveground biomass.
[0035] Example: like Figure 1 As shown, the forest aboveground biomass estimation method based on spaceborne lidar includes: Step S101: Preprocess the GEDI L2A and L4A data to extract forest canopy height and biomass. Because GEDI contains many low-quality light spots with inaccurate height measurements, we first use the relevant parameters built into GEDI and make full use of existing forest canopy height products to identify and filter out abnormal GEDI light spots. In GEDI Level 2A data, data with the following characteristics were removed: quality_flag ≠ 1, degrade_flag ≠ 0, Sensitivity < 0.95, num_detectedmodes ≥ 15, solar_elevation ≥ 0, rx_assess_flag ≠ 0, |elev_lowestmode –SRTM DEM| > 40, |elev_lowestmode – TanDEM-X| > 50, ground height difference extracted by each algorithm setting > 2m, and slope > 20 degrees. Here, quality_flag is a quality flag (1 indicates acceptable data quality), degrade_flag is a degradation flag (0 indicates normal, 1 indicates degraded data), sensitivity is the probability of the laser beam penetrating the canopy and detecting the ground, num_detectedmodes is the number of detected peaks, solar_elevation is the solar altitude angle, rx_assess_flag is the receiver evaluation flag, and |elev_lowestmode – SRTMDEM| is the height difference between the ground height measured by GEDI and the SRTM elevation data, and |elev_lowestmode - TanDEM-X| is the height difference between the ground height measured by GEDI and the TanDEM-X elevation data.
[0036] Next, GEDI data with canopy heights within 2 meters of those in *Global Forest Canopy Height, 2019 | GLAD* were added back. For the filtered GEDI data, the maximum and minimum values were removed using the RH95 calculated by its six algorithm settings, and the average value was calculated as the forest canopy height within the spot range. Spots with canopy heights less than 3 meters were considered non-forest and removed. For GEDI Level 4A data, only Level 4A data with spots identical to those in the previously filtered Level 2A data were retained, and data with l4_quality_flag ≠ 1 were deleted. Then, the agbd attribute was extracted as the point-scale biomass, where l4_quality_flag is a quality flag, with 1 indicating acceptable data quality.
[0037] Step S102, preprocess the Sentinel-1 GRD data: The Sentinel-1 GRD imagery was preprocessed using SNAP and ArcGIS software. The main processing steps included: orbit correction, radiometric calibration, multi-view processing, Refined Lee filtering, terrain correction, decibel conversion, and cropping.
[0038] Step S103: Remove clouds from the Sentinel-2 MSI Level-2A image: Pixels classified as "Clouds Low Probability," "Clouds Medium Probability," "Clouds High Probability," and "Cloud Shadows" were removed based on the "Scene Classification" bands built into Sentinel-2. To fill the resulting gaps, Sentinel-2 images from every two months were sorted in descending order according to the "CLOUDY_PIXEL_PERCENTAGE" attribute and then stitched together.
[0039] Step S104: Segment the GEDI data, Sentinel-1 images, and Sentinel-2 images according to the study area boundary, and divide the data into training and validation sets according to an 8:2 structure: Masking was applied to the GEDI data, Sentinel-1 images, and Sentinel-2 images according to the study area boundaries. The GEDI data within the study area was divided into training and validation sets in an 8:2 ratio. The training set was used for model parameter updates and learning, while the validation set was used to evaluate the model's generalization ability and accuracy.
[0040] Step S105: Construct the DU-Net model to invert forest canopy height: First, Sentinel-1 and Sentinel-2 satellite imagery is segmented into 256x256 pixel image patches. The images are pre-trained using the DINOv2 self-supervised learning framework to obtain a highly generalizable image encoder specifically for vegetation features. This encoder is based on the ViT (Vision Transformer) architecture, decomposing the input image into a sequence of 16×16 image patches. During training, the teacher model receives a global view of the image, while the student model receives both local and global views. Simultaneously, a masked image modeling mechanism is used to randomly mask the input of the student model, forcing the network to learn contextual semantic information. The teacher and student models are jointly trained to align the feature representation output by the student model with that of the teacher model.
[0041] The pre-trained DINOv2 teacher model weights are used directly as the encoder, and all parameters of this encoder module are frozen to remain unchanged during subsequent training. This encoder is responsible for extracting multi-scale deep semantic features from the input satellite images.
[0042] A trainable U-Net decoder is constructed, which receives multi-level features extracted by the encoder via skip connections. Along the extension path, the feature maps are upsampled bilinearly to restore spatial resolution, then concatenated with the corresponding feature maps from the encoder. This concatenation is followed by feature fusion and reconstruction using 3x3 convolutional layers and Rectified Linear Units (ReLU). This process is repeated to progressively restore the image size.
[0043] The final layer generates a single-channel output with the same size as the input image through a 1x1 convolution. Preprocessed GEDI L2A forest canopy height data is used as ground truth labels to supervise the training of the decoder, ultimately outputting a high-precision forest canopy height map.
[0044] Step S106: Construct a Stacking ensemble learning model that includes multiple base learners (Random Forest Regressor, XGBoost, LightGBM, CatBoost, SVR, AdaBoost Regressor, KNN Regressor). The multi-source prediction results are fused through a meta-learner to achieve high-precision inversion of forest aboveground biomass. Features were extracted from the preprocessed Sentinel-1 and Sentinel-2 images, including band features (red, green, blue, near-infrared, VH, VV), vegetation index features (GNDVI, NDVI, DVI, RVI, SAVI), feature variables (VH+VV, VH-VV, VH / VV), as well as the minimum, maximum, median, mean, and standard deviation of the band features, vegetation index features, and feature variables in the time series, and texture features for each band extracted using 3x3, 5x5, 7x7, and 9x9 windows. GNDVI is the Green Normalized Difference Vegetation Index, which is sensitive to chlorophyll; NDVI is the Normalized Difference Vegetation Index, which utilizes the characteristic of plants absorbing red light and reflecting near-infrared light, and can effectively distinguish between wooded and non-wooded areas; DVI is the Difference Vegetation Index, which is sensitive to shading; RVI is the Ratio Vegetation Index, which is sensitive to high biomass areas; and SAVI is the Soil Adjusted Vegetation Index, which is sensitive to sparse forests. The formulas for calculating GNDVI, NDVI, DVI, RVI, and SAVI are as follows, where NIR represents the near-infrared band, Green represents the green light band, and Red represents the red light band:
[0045]
[0046]
[0047]
[0048]
[0049] The following texture features were extracted as feature variables using window sizes of 3x3, 5x5, 7x7, and 9x9. This represents the mean. The number of elements in the normalized gray-level co-occurrence matrix represents the number of elements in the normalized gray-level co- line, number The element values of the column, This represents the number of quantization levels for the image's grayscale. and This represents the grayscale index of a pixel in the image, and its value range is... arrive :
[0050]
[0051]
[0052]
[0053]
[0054]
[0055] To select the most useful feature variables, the Boruta algorithm is used to filter the above feature variables. The core idea of Boruta is to evaluate the importance of each original feature by comparing it with a random "image feature." Image features are generated by randomizing the values of the original features, so they have no real impact on the target variable. Boruta evaluates which original features are more informative than image features by repeatedly training the random forest model, thus selecting the features most useful for prediction.
[0056] Using the selected feature variables and the forest canopy height obtained by inversion from the DU-Net model as input features, multiple differentiated base learners (Random Forest Regressor, XGBoost, LightGBM, CatBoost, SVR, AdaBoost Regressor, KNN Regressor) are trained. A linear regression model is used as a meta-learner to fuse the prediction results of the base learners, thereby constructing a Stacking ensemble learning model to predict forest aboveground biomass.
[0057] Step S107: Use the trained model to estimate forest aboveground biomass and obtain the forest aboveground biomass result map: The optimal model parameters saved in steps S105 and S106 are loaded respectively, and then the samples are input into the trained network model. The output is the estimated value of forest aboveground biomass in the study area.
[0058] This invention innovatively combines lidar and deep learning technologies. LiDAR can directly acquire three-dimensional forest structure data, while deep learning technology can better process the spatiotemporal information in remote sensing data, achieving efficient feature extraction and pattern recognition. This improves the accuracy and efficiency of forest aboveground biomass estimation. It brings significant results to forest resource management, ecological environment protection, and climate change monitoring, taking an important step towards sustainable development.
[0059] The parts of this invention not described in detail are common knowledge to those skilled in the art.
Claims
1. A method for estimating forest aboveground biomass based on spaceborne lidar, characterized in that, include: Preprocessing of GEDI L2A and L4A data was performed to extract forest canopy height and biomass information; Preprocess the Sentinel-1 GRD data; Cloud removal from Sentinel-2 Level-2A images; The preprocessed GEDI data, Sentinel-1 images, and Sentinel-2 images were segmented according to the boundary of the study area, and the data were divided into training set and validation set according to an 8:2 structure. Constructing a DU-Net model to invert forest canopy height; A Stacking ensemble learning model is constructed, which integrates multi-source prediction results through a meta-learner to achieve high-precision inversion of forest aboveground biomass; The constructed DU-Net model and Stacking ensemble learning model were trained and validated using the training and validation sets; The preprocessed real-time satellite data is input into the trained DU-Net model and Stacking ensemble learning model to obtain the forest aboveground biomass result map.
2. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, The following method is used to preprocess GEDI L2A data: GEDI L2A data with the following characteristics were removed: quality_flag ≠ 1, degrade_flag ≠ 0, Sensitivity < 0.95, num_detectedmodes ≥ 15, solar_elevation ≥ 0, rx_assess_flag ≠ 0, |elev_lowestmode –SRTM DEM| > 40, |elev_lowestmode – TanDEM-X| > 50, ground height difference extracted by each algorithm setting > 2m, and slope > 20 degrees. Here, quality_flag is a quality flag (1 indicates acceptable data quality), degrade_flag is a degradation flag (0 indicates normal, 1 indicates degraded data), sensitivity is the probability of the laser beam penetrating the canopy and detecting the ground, num_detectedmodes is the number of detected peaks, solar_elevation is the solar altitude angle, rx_assess_flag is the receiver evaluation flag, and |elev_lowestmode – SRTMDEM| is the height difference between the ground height measured by GEDI and the SRTM elevation data, and |elev_lowestmode - TanDEM-X| is the height difference between the ground height measured by GEDI and the TanDEM-X elevation data; Add back GEDI L2A data that differ from the canopy height in "Global Forest Canopy Height, 2019 | GLAD" by less than 2 meters; For the GEDI L2A data processed above, RH95 is calculated, and after removing the maximum and minimum values of RH95, the average value of the remaining RH95 is calculated. This average value is used as the forest canopy height of the light spot range. Light spots with a canopy height of less than 3m are considered non-forest and removed.
3. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, The following method is used to preprocess GEDIL4A data: For GEDI L4A data, only data with the same spot size as the preprocessed GEDI L2A data are retained, and data with l4_quality_flag ≠ 1 are deleted. Then, the agbd attribute is extracted as the point-scale biomass, where l4_quality_flag is a quality flag bit, and 1 indicates that the data quality is acceptable.
4. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, The preprocessing steps for Sentinel-1 GRD data include: orbit correction, radiometric calibration, multi-view processing, Refined Lee filtering, terrain correction, decibel conversion, and cropping.
5. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, The method for removing clouds from Sentinel-2 Level-2A images is as follows: Pixels classified as "Clouds Low Probability", "Clouds Medium Probability", "Clouds High Probability", and "CloudShadows" are removed based on the "Scene Classification" band built into Sentinel-2. The Sentinel-2 images from every two months are sorted and stitched in descending order according to the "CLOUDY_PIXEL_PERCENTAGE" attribute.
6. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, The DU-Net model is constructed to invert forest canopy height, as detailed below: The Sentinel-1 and Sentinel-2 images were segmented into 256x256 pixel images, and DINOv2 self-supervised learning was used to obtain an image encoder specifically for vegetation images. The image encoder is divided into two versions: a teacher model and a student model. The teacher model receives a global view, while the student model receives both local and global views. The teacher model and the student model are jointly trained to output similar feature representations, thereby obtaining a pre-trained image encoder; The pre-trained teacher model weights are transferred to the encoder part of the U-Net model and frozen to construct a trainable decoder part. Using preprocessed GEDI data as labels and Sentinel images as input, the decoder of the DU-Net model is trained under supervision to generate a forest canopy height map.
7. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 1, characterized in that, Forest aboveground biomass is retrieved using a Stacking ensemble learning model, as detailed below: Features were extracted from the preprocessed Sentinel-1 and Sentinel-2 images, including band features (red, green, blue, near-infrared, VH, VV), vegetation index features (GNDVI, NDVI, DVI, RVI, SAVI), feature variables (VH+VV, VH-VV, VH / VV), and texture features for each band extracted using windows of 3x3, 5x5, 7x7, and 9x9. Select the features most useful for prediction from all the features and feature variables mentioned above; Using the selected feature variables and the forest canopy height obtained by inversion from the DU-Net model as input features, multiple differentiated base learners are trained, and a linear regression model is used as a meta-learner to fuse the prediction results of the base learners, thereby constructing a Stacking ensemble learning model to predict forest aboveground biomass.
8. The method for estimating forest aboveground biomass based on spaceborne lidar according to claim 7, characterized in that, The formulas for calculating vegetation indices GNDVI, NDVI, DVI, RVI, and SAVI are as follows: NIR represents the near-infrared band, Green represents the green light band, and Red represents the red light band.
9. A method for estimating forest aboveground biomass based on spaceborne lidar according to claim 7, characterized in that, Texture features include: in, This represents the mean. The number of elements in the normalized gray-level co-occurrence matrix represents the number of elements in the normalized gray-level co- line, number The element values of the column, This represents the number of quantization levels for the image's grayscale. and This represents the grayscale index of a pixel in the image, and its value range is... arrive .
10. A method for estimating forest aboveground biomass based on spaceborne lidar according to claim 7, characterized in that, The most useful features for prediction are selected using the following method: The Boruta algorithm is used to evaluate which original feature variables are more informative than image features by repeatedly training the random forest model, thereby selecting the feature variables most useful for prediction.