Biomass inversion sample screening optimization method and device, equipment and storage medium

Abstract

The application provides a biomass inversion sample screening optimization method, device and equipment and a storage medium, and relates to the technical field of forest ecological remote sensing monitoring. The method comprises the following steps: inputting a plurality of laser radar sample data into a trained laser radar biomass inversion model to obtain a biomass prediction value and an uncertainty measurement value of each laser radar sample data; performing scale conversion and aggregation on the biomass prediction value and the uncertainty measurement value of each laser radar sample data corresponding to the same satellite pixel; calculating an information quantity score of each satellite pixel based on a satellite image feature vector corresponding to each satellite pixel; and constructing a joint optimization selection function based on the uncertainty measurement value and the information quantity score to select part of all satellite pixels to form a satellite training sample set for training a satellite biomass inversion model. The application can balance reliability and information quantity, and improve the inversion accuracy and generalization ability of the satellite inversion model.

Description

Technical Field

[0001] This invention relates to the field of forest ecological remote sensing monitoring technology, and in particular to a method, apparatus, equipment and storage medium for screening and optimizing biomass inversion samples. Background Technology

[0002] Forest biomass is a key parameter for assessing carbon sinks and studying ecosystem functions. Currently, forest biomass retrieval techniques based on remote sensing technology include direct modeling and two-stage modeling. Direct modeling relies on a limited and unevenly distributed set of ground-based measured plots to directly train the satellite remote sensing model. Due to the scarcity and high cost of ground plots, and the significant mismatch between satellite pixel scale and ground plot scale, the training samples lack representativeness, resulting in poor model extrapolation ability and low retrieval accuracy over large areas with varying forest vegetation cover. Two-stage modeling utilizes UAV lidar to acquire high-precision, medium-range forest structure parameters and establishes a relationship with satellite imagery. Existing methods typically treat the inverted biomass from all lidar plots as equally weighted "true values" for training the satellite model.

[0003] However, the two-level modeling method has two inherent defects: First, the inversion process of lidar itself has uncertainties. Factors such as point cloud density, flight conditions, and inversion model errors can lead to different reliability levels of prediction results from different samples. However, existing methods treat all lidar samples equally as true values, treating high-uncertainty samples and reliable samples equally. As a result, the satellite model trained is susceptible to noise interference and lacks robustness. Second, the selection of lidar samples usually depends on experience or random sampling. A large number of samples with redundant features are included in the training set, while scarce feature regions are under-sampled, resulting in low data utilization efficiency. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for screening and optimizing biomass inversion samples to address the problems of low robustness of training samples and low data utilization efficiency in satellite biomass inversion models.

[0005] In a first aspect, embodiments of the present invention provide a method for screening and optimizing biomass inversion samples, comprising: Multiple lidar sample data are input into a trained lidar biomass inversion model to obtain the predicted biomass value and uncertainty metric value for each lidar sample data; the lidar biomass inversion model is trained based on a training sample composed of lidar sample data and true biomass values. The biomass prediction value and uncertainty metric value of each lidar sample data corresponding to the same satellite pixel are scaled and aggregated to obtain the biomass prediction value and uncertainty metric value of each satellite pixel. Based on the satellite image feature vector corresponding to each satellite pixel, the information content score of each satellite pixel is calculated. A joint optimization selection function is constructed based on the uncertainty metric and information score. Based on the value of the joint optimization selection function, a portion of satellite pixels are selected from all satellite pixels to form a satellite training sample set, which is then used to train the satellite biomass inversion model.

[0006] In one possible implementation, the lidar biomass inversion model is a random forest model; multiple lidar sample data are input into the trained lidar biomass inversion model to obtain the biomass prediction value and uncertainty metric value for each lidar sample data, including: For each lidar sample data, the lidar sample data is input into the trained lidar biomass inversion model to obtain the prediction results of each decision tree for the lidar sample data. The standard deviation of each prediction result is calculated to obtain the uncertainty measure of the lidar sample data.

[0007] In one possible implementation, the biomass prediction value and uncertainty metric value of each lidar sample data corresponding to the same satellite pixel are scaled and aggregated to obtain the biomass prediction value and uncertainty metric value for each satellite pixel, including: Calculate the intersection area of ​​each single-tree polygon that intersects with the first satellite pixel in spatial position and the first satellite pixel, and calculate the ratio of the intersection area to the area of ​​the single-tree polygon as the contribution ratio of the single-tree polygon to the first satellite pixel; wherein, the first satellite pixel is any satellite pixel, and each single-tree polygon corresponds to a lidar sample. The biomass prediction value of each single tree polygon is multiplied by the corresponding contribution ratio in the first satellite pixel to obtain the biomass contribution value of the single tree polygon in the first satellite pixel. The biomass contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the predicted biomass value of the first satellite pixel. The uncertainty metric of each single-tree polygon is multiplied by its corresponding contribution ratio in the first satellite pixel to obtain the uncertainty contribution value of the single-tree polygon in the first satellite pixel. The uncertainty contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the uncertainty metric value of the first satellite pixel.

[0008] In one possible implementation, based on the satellite image feature vector corresponding to each satellite pixel, the information content score of each satellite pixel is calculated, including: Clustering is performed on the satellite image feature vector corresponding to each satellite pixel to obtain multiple clusters; For each satellite pixel, the information content score of the satellite pixel is determined based on the reciprocal of the number of satellite image feature vectors in the cluster to which the satellite image feature vector of the corresponding satellite pixel belongs.

[0009] In one possible implementation, the joint optimization selection function is:

[0010] in, To jointly optimize the selection of function values, Information content score, This is a measure of uncertainty. This is a preset constant; A subset of satellite pixels is selected based on the joint optimization selection function values ​​to form a satellite training sample set, including: From all the satellites, select the N satellite pixels with the largest joint optimization selection function value to form a satellite training sample set; where N is a preset value.

[0011] In one possible implementation, before inputting multiple lidar sample data into the trained lidar biomass inversion model, the following steps are also included: Acquire multiple sets of one-to-one corresponding ground sample data and lidar sample data; For each lidar sample data, the true biomass value of that lidar sample data is calculated based on the corresponding ground sample data. Using LiDAR sample data as training samples and corresponding ground truth biomass values ​​as sample labels, the initial random forest model is trained to obtain a trained LiDAR biomass inversion model.

[0012] In one possible implementation, after selecting a subset of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, the following is also included: The satellite biomass inversion model was trained using a satellite training sample set to obtain a trained satellite biomass inversion model. Biomass inversion mapping of the target area is performed based on a trained satellite biomass inversion model to obtain a satellite inversion biomass distribution map; the target area contains multiple sub-regions. Based on the trained lidar biomass inversion model, biomass inversion mapping is performed on the target area to obtain a radar-inverted biomass distribution map. The satellite inversion value of each sub-region in the satellite inversion biomass distribution map is compared with the predicted biomass value of the same sub-region in the radar inversion biomass distribution map, and the regions with differences greater than a preset threshold are identified as high conflict areas. Local recalibration of the trained lidar biomass inversion model is performed based on high-conflict zones, and / or high-conflict zones are marked as data acquisition areas to guide the supplementary acquisition of lidar sample data.

[0013] In one possible implementation, the trained lidar biomass inversion model is locally recalibrated based on high-conflict zones, including: Obtain ground sample data from high-conflict areas as supplementary training samples; Based on supplementary training samples, the trained lidar biomass inversion model is incrementally trained to obtain a locally recalibrated lidar biomass inversion model.

[0014] Secondly, embodiments of the present invention provide a screening and optimization device for biomass inversion samples, comprising: The uncertainty measurement module is used to input multiple lidar sample data into the trained lidar biomass inversion model to obtain the biomass prediction value and uncertainty measurement value for each lidar sample data; wherein, the lidar biomass inversion model is trained based on a training sample composed of lidar sample data and true biomass values. The conversion and combination module is used to scale and aggregate the biomass prediction values ​​and uncertainty measure values ​​of each lidar sample data corresponding to the same satellite pixel to obtain the biomass prediction value and uncertainty measure value of each satellite pixel. The information content calculation module is used to calculate the information content score of each satellite pixel based on the satellite image feature vector corresponding to each satellite pixel. The joint optimization module is used to construct a joint optimization selection function based on the uncertainty metric and information score, and select a portion of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, which is then used to train the satellite biomass inversion model.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0017] The biomass inversion sample screening and optimization method, apparatus, equipment, and storage medium provided in this invention simultaneously output biomass prediction values ​​and uncertainty metrics through a lidar biomass inversion model. This allows for a quantitative assessment of the reliability of each lidar sample, providing a basis for eliminating low-quality samples and preventing high-uncertainty samples from contaminating the satellite model. Secondly, by calculating the information score of each sample based on the satellite image feature space, high-value samples with scarce features and high information gain can be actively identified, overcoming the blindness of traditional random or empirical selection. Finally, a selection function is jointly constructed based on the uncertainty metric and the information score, automatically selecting satellite pixels with the best comprehensive score from the satellite pixels to form a satellite training sample set, achieving a balance between reliability and information content. The satellite inversion model trained in this way, due to the higher quality and stronger representativeness of the input samples, significantly improves its inversion accuracy and generalization ability, while maximizing the efficiency of limited lidar acquisition resources. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the biomass inversion sample screening and optimization method provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the biomass inversion sample screening and optimization device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] See Figure 1 The flowchart illustrating the implementation of the biomass inversion sample screening and optimization method provided in this embodiment of the invention is described in detail below: Step 101: Input multiple lidar sample data into the trained lidar biomass inversion model to obtain the predicted biomass value and uncertainty metric value for each lidar sample data; wherein, the lidar biomass inversion model is trained based on a training sample composed of lidar sample data and true biomass values.

[0021] In this embodiment, the lidar sample data refers to a set of data collected by a lidar sensor mounted on a drone and preprocessed. Each sample corresponds to a spatial unit and contains multidimensional features that reflect the forest structure, such as tree height statistics, crown width characteristics, and point cloud intensity. The multidimensional feature vectors of the lidar sample data specifically include the following four categories: The first category is canopy variable features, such as average crown width, crown width standard deviation, maximum crown width, and minimum crown width. The second category is height variable features, including tree height percentiles (10th, 25th, 50th, 75th, and 90th percentiles), maximum value, average value, standard deviation, and coefficient of variation. The third category is canopy cover and density features, such as vegetation cover and leaf area index. The fourth category is point cloud intensity features, including average intensity and intensity variance. These features describe the three-dimensional structure and canopy characteristics of the forest within the sample plot from different dimensions.

[0022] It should be noted that this method involves three different scales of data sources: ground quadrats are used to provide true biomass values, with the smallest spatial range (usually 20m × 20m) and a limited number; lidar quadrats are divided into training and prediction types. Training lidar quadrats are collected within a small area of ​​the plot, corresponding one-to-one with ground quadrats and having the same scale; prediction lidar quadrats are collected within a medium area of ​​the plot, usually 1000m × 1000m, with a larger number and the same scale as the training lidar quadrats, and do not collect corresponding ground quadrats. The predicted biomass value is determined by the lidar biomass inversion model; satellite pixels are the basic unit for the final inversion (10m to 30m). A high-precision model is established by collecting training lidar samples from a small area. This model is then applied to lidar samples collected from a medium area to obtain a large number of biomass predictions with uncertainties. Finally, these predictions are distributed to satellite pixels through scale transformation to form a sample set for training the satellite model. This is used to train the satellite biomass inversion model, enabling it to predict biomass over a large area.

[0023] The raw lidar point cloud is denoised, filtered, and classified into digital elevation models (DEMs) and digital surface models. Subtracting the two yields the canopy height model. Then, a single-tree segmentation algorithm is used to extract features such as tree height, canopy width, and point cloud intensity for each tree, which are then aggregated to form a feature vector of the lidar sample data. The lidar biomass inversion model is a pre-trained mathematical model. Its input is lidar sample data, and its output is the predicted biomass value and the corresponding uncertainty metric. The predicted biomass value refers to the forest aboveground biomass estimated by the model, commonly expressed in tons per hectare. The uncertainty metric characterizes the reliability of the predicted biomass value; a higher value indicates a less reliable prediction. Sources of uncertainty may include insufficient point cloud density, poor flight conditions, and fitting errors of the inversion model on specific forest types. The model is trained based on a large amount of one-to-one lidar sample data and true biomass values, typically obtained from ground quadrat surveys. When large amounts of ground quadrat data are difficult to obtain, the biomass of individual trees can be estimated from diameter at breast height (DBH) and tree height using the allometric growth equation, and then aggregated to the quadrat scale.

[0024] in, DBH Diameter at breast height, a , b Fitting parameters for each tree species.

[0025] In practice, multiple lidar sample data are input into the model one by one. After internal calculations, the model outputs a biomass prediction and an uncertainty metric for each sample. This process ensures that the estimation result for each sample is accompanied by a quantitative indicator of its reliability. In this way, subsequent steps can identify and prioritize the use of samples with reliable predictions, while eliminating noisy samples with high uncertainty, thereby improving the accuracy of the entire inversion chain.

[0026] Step 102: The biomass prediction value and uncertainty metric value of each lidar sample data corresponding to the same satellite pixel are scaled and aggregated to obtain the biomass prediction value and uncertainty metric value of each satellite pixel.

[0027] In this embodiment, since lidar sample data typically uses individual trees as the basic unit, with each tree represented as an irregular vector polygon, while satellite imagery uses regular raster pixels as the basic unit, there is a difference in spatial scale. Therefore, after obtaining the biomass prediction and uncertainty metric values ​​for each lidar sample data, it is necessary to convert the individual tree-scale biomass prediction and uncertainty metric values ​​to raster data aligned with the satellite image pixel scale. Specifically, an area-weighted method is used for the conversion. For any target pixel in the satellite imagery, all individual tree polygons that intersect with that pixel in spatial location are first identified. For each intersecting individual tree polygon, the intersection area between the polygon and the target pixel is calculated, and then the proportion of this intersection area to the total area of ​​the individual tree polygon is calculated. Then, the biomass prediction value of the individual tree polygon is multiplied by the above proportion to obtain the biomass contribution value of that individual tree to that pixel. The biomass value of the target pixel is obtained by summing the contribution values ​​of all intersecting individual trees. The scaling transformation of the uncertainty metric values ​​employs the same area-weighted method: the uncertainty metric values ​​of each individual polygon are weighted by the same intersection area ratio and then summed to obtain the uncertainty metric value for that satellite pixel. Through this transformation, the lidar sample data is reorganized into raster data corresponding one-to-one with satellite image pixels. Each record represents the predicted biomass value and its uncertainty metric value at a satellite pixel location, providing scale-matched input data for subsequent sample selection and satellite model training.

[0028] Step 103: Calculate the information content score of each satellite pixel based on the satellite image feature vector corresponding to each satellite pixel.

[0029] In this embodiment, the satellite image feature vector refers to the set of features extracted from satellite multispectral images that correspond to the location of the lidar sample, such as reflectance in each band, vegetation index, and tassel transform components. To obtain these features, satellite images covering the target area, such as Sentinel-2 or Landsat series data, are first collected and subjected to radiometric calibration, atmospheric correction, geometric fine correction, and cloud masking to generate surface reflectance products. Then, within the coverage area of ​​each satellite pixel, multispectral features are extracted, including reflectance values ​​in each band and normalized vegetation index, enhanced vegetation index, ratio vegetation index, normalized differential moisture index, red-edge chlorophyll index, etc., calculated from these bands. Brightness, greenness, and humidity components obtained from tassel transform can also be calculated. These features are combined into a high-dimensional vector, which is the satellite image feature vector of that satellite pixel. All satellite image feature vectors together constitute a feature space.

[0030] The information score is designed to measure the scarcity or representativeness of each sample in the feature space. If a satellite pixel's feature pattern is unique among all candidate samples, meaning it differs significantly from the features of most other satellite pixels, then that satellite pixel has a high information score. This means that adding it to the training set can provide more new information to the satellite model, helping to improve the model's generalization ability. Conversely, if a satellite pixel's feature pattern appears frequently in the candidate set, its information score is low, indicating that the satellite pixel carries less new information. There are various methods for calculating the information score, such as distance-based nearest neighbor methods or clustering-based methods, but their core idea is to evaluate the uniqueness of a sample in the feature space. By calculating the information score of each sample, this embodiment quantifies the value of samples from an information theory perspective, providing a second dimension of evaluation criteria for subsequent intelligent screening.

[0031] Step 104: Construct a joint optimization selection function based on the uncertainty metric and information score, and select a portion of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, so as to train the satellite biomass inversion model based on the satellite training sample set.

[0032] In this embodiment, the joint optimization selection function integrates two dimensions—reliability and information content—into a comprehensive score. Using uncertainty metrics alone can only eliminate unreliable samples, but may tend to select samples with extremely low uncertainty but redundant features, resulting in a lack of diversity in the training set; using information content scores alone may select samples with high information content but extremely unreliable samples, introducing significant noise.

[0033] Therefore, the design principle of the joint optimization selection function is: the comprehensive score is positively correlated with the information content score and negatively correlated with the uncertainty metric, that is, samples with high information content and low uncertainty are preferentially selected. For each candidate sample, its uncertainty metric and information content score are substituted into the joint optimization selection function to calculate the comprehensive score. Then, all candidate samples are sorted from high to low according to the comprehensive score. Based on the preset target sample number, a number of satellite pixels with high ranking are selected, and the set of these data is called the satellite training sample set. Finally, the satellite biomass inversion model is trained based on this satellite training sample set.

[0034] Before selecting some satellite pixels from all satellite pixels based on the joint optimization selection function value, satellite pixels with uncertainty values ​​exceeding a preset threshold can be filtered out. Then, based on the joint optimization selection function value, satellite pixels with moderate to low uncertainty and extremely high information content are selected from the remaining satellite pixels to form a satellite training sample set.

[0035] The satellite biomass retrieval model is a regression model that takes the satellite image feature vector of each sample as input and outputs the biomass value. During training, the satellite image feature vector of each sample in the training sample set is used as the input feature, while the corresponding lidar biomass prediction value is used as the training label—since these samples have undergone uncertainty screening, their labels are considered reliable. Through training, the satellite model learns the mapping relationship from satellite image features to forest biomass. The training samples selected through a joint optimization selection function are both reliable and information-rich, enabling the satellite model to achieve better generalization performance under limited training sample conditions.

[0036] This invention employs a lidar biomass inversion model that simultaneously outputs predicted biomass and uncertainty metrics. This allows for a quantitative assessment of the reliability of each lidar sample, providing a basis for eliminating low-quality samples and preventing high-uncertainty samples from contaminating the satellite model. Secondly, by calculating the information score of each sample based on the satellite image feature space, it proactively identifies high-value samples with scarce features and high information gain, overcoming the blindness of traditional random or empirical selection. Finally, a selection function is jointly constructed based on the uncertainty metric and information score, automatically selecting the sample with the best comprehensive score from the candidate samples to form the satellite training sample set, achieving a balance between reliability and information content. The satellite inversion model trained in this way, due to its higher quality and more representative input samples, significantly improves its inversion accuracy and generalization ability, while maximizing the efficiency of limited lidar acquisition resources.

[0037] In one possible implementation, the lidar biomass inversion model is a random forest model; multiple lidar sample data are input into the trained lidar biomass inversion model to obtain the biomass prediction value and uncertainty metric value for each lidar sample data, including: For each lidar sample data, the lidar sample data is input into the trained lidar biomass inversion model to obtain the prediction results of each decision tree for the lidar sample data. The standard deviation of each prediction result is calculated to obtain the uncertainty measure of the lidar sample data.

[0038] In this embodiment, the lidar biomass inversion model can specifically adopt a random forest model. Random forest is an ensemble learning algorithm composed of multiple decision trees. Each decision tree uses a different bootstrap sampling training set and random feature subset during training, thus possessing natural diversity.

[0039] When LiDAR sample data is input into a trained random forest model, each decision tree in the model independently outputs a predicted biomass value. If there are T decision trees in the random forest, then T predicted values ​​can be obtained. The standard deviation of these T predicted values ​​is calculated and used as a measure of the uncertainty of the LiDAR sample data. A larger standard deviation indicates greater discrepancy in the predictions of the decision trees for that sample, meaning that the sample's features match the patterns in the training data less closely, or that the sample is in a region where the model's generalization ability is weak; therefore, the reliability of the prediction results is lower. Conversely, a smaller standard deviation indicates that the predictions of the decision trees tend to be consistent, and the prediction results are more reliable.

[0040] In practical applications, the number of trees in a random forest can be adjusted based on the forest type and point cloud quality. For example, setting T=500 can ensure the stability of the standard deviation estimate. Utilizing the ensemble properties of random forests, the uncertainty measure of each prediction result can be obtained without changing the model structure, offering high computational efficiency and natural compatibility with the model training process.

[0041] In one possible implementation, the biomass prediction value and uncertainty metric value of each lidar sample data corresponding to the same satellite pixel are scaled and aggregated to obtain the biomass prediction value and uncertainty metric value for each satellite pixel, including: Calculate the intersection area of ​​each single-tree polygon that intersects with the first satellite pixel in spatial position and the first satellite pixel, and calculate the ratio of the intersection area to the area of ​​the single-tree polygon as the contribution ratio of the single-tree polygon to the first satellite pixel; wherein, the first satellite pixel is any satellite pixel, and each single-tree polygon corresponds to a lidar sample. The biomass prediction value of each single tree polygon is multiplied by the corresponding contribution ratio in the first satellite pixel to obtain the biomass contribution value of the single tree polygon in the first satellite pixel. The biomass contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the predicted biomass value of the first satellite pixel. The uncertainty metric of each single-tree polygon is multiplied by its corresponding contribution ratio in the first satellite pixel to obtain the uncertainty contribution value of the single-tree polygon in the first satellite pixel. The uncertainty contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the uncertainty metric value of the first satellite pixel.

[0042] In this embodiment, the above-mentioned scale transformation and aggregation process can be implemented through the following specific steps. Taking any satellite pixel as the first satellite pixel for explanation, firstly, the spatial geometric information, biomass prediction value, and uncertainty metric value of each individual tree polygon are obtained from all lidar sample data. It should be noted that after individual tree segmentation and feature extraction, each lidar sample data is represented as a single-tree polygon corresponding to that sample. In other words, the lidar sample and the individual tree polygon have the same spatial geometric meaning; that is, each sample itself is a closed polygonal region, and its boundary contains the comprehensive features of all trees within that sample. Therefore, during scale transformation, the basic unit processed is these individual tree polygons, and their area is the canopy projection area of ​​that individual tree.

[0043] Then, calculate the intersection area of ​​each monolithic polygon that intersects the first satellite pixel in spatial location with that first satellite pixel. This calculation can be performed using a vector polygon intersection algorithm, such as a polygon clipping algorithm in computer graphics.

[0044] Next, the intersection area of ​​each individual polygon is divided by the total area of ​​that polygon itself to obtain the contribution ratio of that polygon to the first satellite pixel. This contribution ratio is between 0 and 1. When the individual polygon is completely inside the first satellite pixel, the contribution ratio is 1, and when the two are only tangent at their edges, the contribution ratio approaches 0.

[0045] Then, the predicted biomass value of the single-tree polygon is multiplied by the above contribution ratio to obtain the biomass contribution value of the single-tree polygon in the first satellite pixel; the uncertainty metric value of the single-tree polygon is multiplied by the above contribution ratio to obtain the uncertainty contribution value.

[0046] Repeat the above calculation for all individual tree polygons intersecting with the first satellite pixel. Finally, sum the biomass contribution values ​​of all individual tree polygons to the first satellite pixel to obtain the biomass prediction value of the first satellite pixel; sum the uncertainty contribution values ​​of all individual tree polygons to the first satellite pixel to obtain the uncertainty metric value of the first satellite pixel.

[0047] In this way, the predicted biomass value of each satellite pixel is a comprehensive result of the contribution of all surrounding individual trees proportionally to their area. This avoids the information loss caused by simple averaging and ensures the consistency of uncertainty measurement transmission. When multiple lidar samples overlap at the boundary, the same individual tree polygon will not be counted repeatedly because each individual tree polygon has a unique spatial identifier. If different lidar samples cover the same area but the individual tree segmentation results are inconsistent, the sample data with higher point cloud density and more recent acquisition time is given priority, or fusion and deduplication are performed during the preprocessing stage. This embodiment provides a rigorous and repeatable scale conversion scheme to ensure lossless information transfer from the individual tree scale to the satellite pixel scale.

[0048] In one possible implementation, based on the satellite image feature vector corresponding to each satellite pixel, the information content score of each satellite pixel is calculated, including: Clustering is performed on the satellite image feature vector corresponding to each satellite pixel to obtain multiple clusters; For each satellite pixel, the information content score of the satellite pixel is determined based on the reciprocal of the number of satellite image feature vectors in the cluster to which the satellite image feature vector of the corresponding satellite pixel belongs.

[0049] In this embodiment, the satellite image feature vector corresponding to each satellite pixel is first clustered to obtain multiple clusters. The K-means algorithm can be used for clustering, and the number of clusters can be dynamically determined based on the total number of candidate samples. For example, the number of clusters can be set to the integer part of the number of candidate samples divided by 20, and not less than 5. The number of clustering iterations is set to 100, and the convergence threshold is set to 0.001. After clustering, each satellite image feature vector is assigned to a unique cluster, and each cluster contains several samples. Thus, the feature space is divided into several regions, and the samples within each region have similar spectral characteristics.

[0050] For any satellite pixel, its cluster is determined, and the number of samples in that cluster is counted, denoted as C. The information score I of that satellite pixel is then defined as 1 divided by the sum of C and a smoothing coefficient λ, i.e., I = 1 / (C + λ). The smoothing coefficient λ is a small positive number (e.g., 1) to prevent the denominator from being zero when C is zero. Clearly, the smaller C is, the larger I is; conversely, the larger C is, the smaller I is. This means that if a satellite pixel's cluster contains only a few satellite image feature vectors, the satellite pixel is considered to have a high information score because its feature pattern is relatively unique in the current sample set, and adding it to the training set can significantly expand the satellite model's coverage of the feature space. Conversely, if the cluster of a satellite pixel contains a large number of satellite image feature vectors, it indicates that its feature pattern has been sufficiently sampled, resulting in a lower information score. This information score evaluation method based on cluster frequency is intuitive and requires no additional prior knowledge, effectively guiding sample selection towards feature-sparse regions.

[0051] In one possible implementation, the joint optimization selection function is:

[0052] in, To jointly optimize the selection of function values, Information content score, This is a measure of uncertainty. This is a preset constant; Based on the joint optimization selection function value, a subset of satellite pixels are selected from all satellite pixels to form a satellite training sample set, including: Among all satellite pixels, the N satellite pixels with the largest joint optimization selection function value are selected to form a satellite training sample set; where N is a preset value.

[0053] In this embodiment, the natural logarithmic transformation is used to compress the numerical scale of the uncertainty metric U when its dynamic range is large, preventing a few samples with extremely high U values ​​from having such large denominators that they are completely excluded from selection. Simultaneously, the logarithmic transformation preserves the monotonic relationship that a larger U results in a smaller score. (Constant) Take a very small positive number, for example Its purpose is to prevent the denominator from being zero when log(1+U)=0 when U equals zero. The design of this function reflects the idea of ​​trade-offs: the larger the information score I is, the better, and the smaller the uncertainty measure U is, the better. Therefore, samples with high comprehensive scores are usually those with high information content and low uncertainty, that is, golden samples that are both reliable and scarce.

[0054] The value of N can be determined based on the complexity of the satellite biomass inversion model, the desired training set size, and computational resources. For example, with a total of 500 candidate samples, N can be set to 100; or N can be set to 20% of the total number of candidate samples. After scoring and ranking, the top N samples are selected, and the rest are excluded. This embodiment provides a quantifiable and repeatable sample selection criterion, making the selection results deterministic and interpretable.

[0055] In one possible implementation, before inputting multiple lidar sample data into the trained lidar biomass inversion model, the following steps are also included: Acquire multiple sets of one-to-one corresponding ground sample data and lidar sample data; For each lidar sample data, the true biomass value of that lidar sample data is calculated based on the corresponding ground sample data. Using LiDAR sample data as training samples and corresponding ground truth biomass values ​​as sample labels, the initial random forest model is trained to obtain a trained LiDAR biomass inversion model.

[0056] In this embodiment, before executing step 101, the lidar biomass inversion model needs to be pre-trained. First, multiple sets of one-to-one corresponding ground sample data and lidar sample data are acquired. The ground sample data consists of fixed quadrats deployed in a typical forest area, typically 20 meters by 20 meters in size. All trees within the quadrats with a diameter at breast height (DBH) greater than 5 cm are measured individually, recording the tree species, DBH, height, and latitude and longitude coordinates for each tree. The biomass of each tree is calculated based on the allometric growth equations for each tree species, and the true biomass values ​​at the quadrats are summarized. The lidar sample data is obtained by collecting point clouds in the same area using a drone equipped with lidar and then processing them. Then, for each quadrat, the lidar sample data is used as the feature input for the training samples, and the corresponding ground sample biomass values ​​are used as the sample labels. The initial random forest model is trained using a random forest algorithm.

[0057] During training, hyperparameters need to be set, such as setting the number of decision trees to 500, the maximum tree depth to 20, and the minimum number of sample splits to 5, and using 5-fold cross-validation for parameter optimization. After sufficient training, the random forest model learns the nonlinear mapping relationship between lidar features and biomass, at which point a trained lidar biomass inversion model is obtained. This model can be called in step 101 to output biomass predictions and uncertainty metrics for new candidate lidar sample data.

[0058] In one possible implementation, after selecting a subset of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, the following is also included: The satellite biomass inversion model was trained using a satellite training sample set to obtain a trained satellite biomass inversion model. Biomass inversion mapping of the target area is performed based on a trained satellite biomass inversion model to obtain a satellite inversion biomass distribution map; the target area contains multiple sub-regions. Based on the trained lidar biomass inversion model, biomass inversion mapping is performed on the target area to obtain a radar-inverted biomass distribution map. The satellite inversion value of each sub-region in the satellite inversion biomass distribution map is compared with the predicted biomass value of the same sub-region in the radar inversion biomass distribution map, and the regions with differences greater than a preset threshold are identified as high conflict areas. Local recalibration of the trained lidar biomass inversion model is performed based on high-conflict zones, and / or high-conflict zones are marked as data acquisition areas to guide the supplementary acquisition of lidar sample data.

[0059] In this embodiment, the satellite biomass inversion model is first trained using the satellite training sample set assembled in step 103 to obtain the trained satellite biomass inversion model. The satellite biomass inversion model can employ a gradient boosting tree algorithm (such as XGBoost or LightGBM), with the satellite image feature vector for each sample as input and biomass value as output.

[0060] After training, biomass inversion mapping of the target area is performed based on the model. The target area comprises multiple sub-regions, such as those defined by satellite imagery pixel grids or administrative divisions. The trained satellite biomass inversion model is applied to satellite imagery of the entire target area, calculating biomass pixel by pixel to generate a satellite-inverted biomass distribution map. Simultaneously, using a pre-trained lidar biomass inversion model, biomass prediction is performed on all sub-regions within the target area covered by lidar samples, resulting in a lidar-inverted biomass distribution map. Since lidar inversion has high accuracy in local areas, this result can be used as a reference benchmark. The satellite inversion value of each sub-region in the satellite-inverted biomass distribution map is compared sub-region by sub-region with the lidar biomass prediction value of the same sub-region in the lidar-inverted biomass distribution map, and the relative error is calculated. The relative error is equal to the absolute value of the difference between the satellite inversion value and the lidar prediction value divided by the lidar prediction value. If the relative error of a certain sub-region exceeds a preset threshold (e.g., 30%), the region is identified as a high-conflict area.

[0061] For identified high-conflict areas, this invention provides two processing methods: one is to locally recalibrate the trained lidar biomass inversion model; the other is to mark the high-conflict areas as data acquisition areas to guide the supplementary acquisition of lidar sample data. Both methods can be used simultaneously, or one method can be selected according to the actual situation.

[0062] When local recalibration is selected, the system acquires ground sample data from high-conflict areas as supplementary training samples to incrementally train the lidar biomass inversion model, resulting in a locally recalibrated lidar biomass inversion model. When supplementary data collection is selected, the system marks high-conflict areas as priority data collection areas and dispatches UAVs or ground survey teams to conduct supplementary flights or quadrat deployments in these areas. Subsequently, using the supplementally collected data or the locally recalibrated model, steps 101 to 103 and the feedback steps of this iteration are re-executed. After multiple iterations, the area proportion of high-conflict areas gradually decreases, and the inversion accuracy of the system continuously improves. This closed-loop optimization mechanism enables the system to proactively identify its weaknesses and make targeted improvements, achieving a leap from static one-time modeling to dynamic adaptive learning.

[0063] In one possible implementation, the trained lidar biomass inversion model is locally recalibrated based on high-conflict zones, including: Obtain ground sample data from high-conflict areas as supplementary training samples; Based on supplementary training samples, the trained lidar biomass inversion model is incrementally trained to obtain a locally recalibrated lidar biomass inversion model.

[0064] In this embodiment, after identifying a high-conflict zone, ground sample data within that zone is acquired as supplementary training samples. Ground sample data can be acquired by newly establishing ground quadrats within the high-conflict zone, measuring each tree, and calculating the true biomass based on the allometric growth equation; alternatively, if historical ground survey data already exists for the area but has not yet been used for initial training, this historical data can be reused. These newly acquired ground sample data are paired with corresponding lidar sample data to form a small-scale supplementary training sample set. Then, based on these supplementary training samples, the already trained lidar biomass inversion model is incrementally trained.

[0065] One approach to incremental training is to add supplementary training samples to the original training set and fully retrain the model using the same random forest hyperparameters. Due to the incremental learning characteristics of random forests, the model can also be adjusted using only new samples. For example, for each decision tree, the new samples can be propagated down the tree to update the predicted values ​​of the leaf nodes, or new decision trees can be added to fit the residuals of the new samples. After incremental training, the model's prediction bias in high-conflict areas is corrected, resulting in more reliable biomass predictions and uncertainty metrics. When these updated LiDAR sample data are used again for satellite model sample selection, samples from high-conflict areas are more likely to be selected for the satellite training sample set due to reduced uncertainty, thereby improving the satellite model's extrapolation capability in that region and creating positive feedback. This local recalibration method avoids the high computational cost of global retraining, can quickly respond to weaknesses identified by the system, and achieves efficient utilization of limited ground data resources.

[0066] When performing iterative feedback optimization, stopping conditions need to be set to avoid infinite loops. When the total area of ​​high-conflict zones falls below a preset threshold (e.g., 5%), it indicates that the satellite model and the lidar model have largely reached agreement in that area, and the system accuracy has stabilized; at this point, the iteration can be terminated. Additionally, when the number of iterations reaches a preset upper limit (e.g., 3 times), iteration should also stop even if the high-conflict zones have not been completely eliminated to prevent excessive consumption of computational resources. In practical applications, the above thresholds and upper limits can be flexibly adjusted according to the accuracy requirements of the monitoring task and the data acquisition cost.

[0067] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0068] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0069] Figure 2 A schematic diagram of the biomass inversion sample screening and optimization device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 2 As shown, the biomass inversion sample screening and optimization device 2 includes: Uncertainty measurement module 21 is used to input multiple lidar sample data into the trained lidar biomass inversion model to obtain the biomass prediction value and uncertainty measurement value of each lidar sample data; wherein, the lidar biomass inversion model is trained based on the training sample composed of lidar sample data and true biomass values. The conversion and aggregation module 22 is used to scale and aggregate the biomass prediction value and uncertainty measure value of each lidar sample data corresponding to the same satellite pixel to obtain the biomass prediction value and uncertainty measure value of each satellite pixel. The information content calculation module 23 is used to calculate the information content score of each satellite pixel based on the satellite image feature vector corresponding to each satellite pixel. The joint optimization module 24 is used to construct a joint optimization selection function based on the uncertainty metric and information score, and select a portion of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, so as to train the satellite biomass inversion model based on the satellite training sample set.

[0070] In one possible implementation, the lidar biomass inversion model is a random forest model; the uncertainty measurement module 21 is specifically used for: For each lidar sample data, the lidar sample data is input into the trained lidar biomass inversion model to obtain the prediction results of each decision tree for the lidar sample data. The standard deviation of each prediction result is calculated to obtain the uncertainty measure of the lidar sample data.

[0071] In one possible implementation, the conversion aggregation module 22 is specifically used for: Calculate the intersection area of ​​each single-tree polygon that intersects with the first satellite pixel in spatial position and the first satellite pixel, and calculate the ratio of the intersection area to the area of ​​the single-tree polygon as the contribution ratio of the single-tree polygon to the first satellite pixel; wherein, the first satellite pixel is any satellite pixel, and each single-tree polygon corresponds to a lidar sample. The biomass prediction value of each single tree polygon is multiplied by the corresponding contribution ratio in the first satellite pixel to obtain the biomass contribution value of the single tree polygon in the first satellite pixel. The biomass contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the predicted biomass value of the first satellite pixel. The uncertainty metric of each single-tree polygon is multiplied by its corresponding contribution ratio in the first satellite pixel to obtain the uncertainty contribution value of the single-tree polygon in the first satellite pixel. The uncertainty contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the uncertainty metric value of the first satellite pixel.

[0072] In one possible implementation, the information calculation module 23 is specifically used for: Clustering is performed on the satellite image feature vector corresponding to each satellite pixel to obtain multiple clusters; For each satellite pixel, the information content score of the satellite pixel is determined based on the reciprocal of the number of satellite image feature vectors in the cluster to which the satellite image feature vector of the corresponding satellite pixel belongs.

[0073] In one possible implementation, the joint optimization selection function is:

[0074] in, To jointly optimize the selection of function values, Information content score, This is a measure of uncertainty. This is a preset constant; Joint optimization module 24 is specifically used for: Among all satellite pixels, the N satellite pixels with the largest joint optimization selection function value are selected to form a satellite training sample set; where N is a preset value.

[0075] In one possible implementation, the uncertainty measurement module 21 is also used for: Before inputting multiple lidar sample data into the trained lidar biomass inversion model, multiple sets of one-to-one corresponding ground sample data and lidar sample data are obtained. For each lidar sample data, the true biomass value of that lidar sample data is calculated based on the corresponding ground sample data. Using LiDAR sample data as training samples and corresponding ground truth biomass values ​​as sample labels, the initial random forest model is trained to obtain a trained LiDAR biomass inversion model.

[0076] In one possible implementation, the joint optimization module 24 is also used for: After selecting a portion of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, the satellite biomass inversion model is trained using the satellite training sample set to obtain the trained satellite biomass inversion model. Biomass inversion mapping of the target area is performed based on a trained satellite biomass inversion model to obtain a satellite inversion biomass distribution map; the target area contains multiple sub-regions. Based on the trained lidar biomass inversion model, biomass inversion mapping is performed on the target area to obtain a radar-inverted biomass distribution map. The satellite inversion value of each sub-region in the satellite inversion biomass distribution map is compared with the predicted biomass value of the same sub-region in the radar inversion biomass distribution map, and the regions with differences greater than a preset threshold are identified as high conflict areas. Local recalibration of the trained lidar biomass inversion model is performed based on high-conflict zones, and / or high-conflict zones are marked as data acquisition areas to guide the supplementary acquisition of lidar sample data.

[0077] In one possible implementation, the joint optimization module 24 is specifically used for: Obtain ground sample data from high-conflict areas as supplementary training samples; Based on supplementary training samples, the trained lidar biomass inversion model is incrementally trained to obtain a locally recalibrated lidar biomass inversion model.

[0078] This invention employs a lidar biomass inversion model that simultaneously outputs predicted biomass and uncertainty metrics. This allows for a quantitative assessment of the reliability of each lidar sample, providing a basis for eliminating low-quality samples and preventing high-uncertainty samples from contaminating the satellite model. Secondly, by calculating the information score of each sample based on the satellite image feature space, it proactively identifies high-value samples with scarce features and high information gain, overcoming the blindness of traditional random or empirical selection. Finally, a selection function is jointly constructed based on the uncertainty metric and information score, automatically selecting the sample with the best comprehensive score from the candidate samples to form the satellite training sample set, achieving a balance between reliability and information content. The satellite inversion model trained in this way, due to its higher quality and more representative input samples, significantly improves its inversion accuracy and generalization ability, while maximizing the efficiency of limited lidar acquisition resources.

[0079] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0080] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0081] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0082] The processor 30 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0083] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the electronic device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0084] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0085] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0086] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0087] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0088] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0089] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for screening and optimizing biomass inversion samples, characterized in that, include: Multiple lidar sample data are input into a trained lidar biomass inversion model to obtain the predicted biomass value and uncertainty metric value for each lidar sample data; wherein, the lidar biomass inversion model is trained based on a training sample composed of lidar sample data and true biomass values. The biomass prediction value and uncertainty metric value of each lidar sample data corresponding to the same satellite pixel are scaled and aggregated to obtain the biomass prediction value and uncertainty metric value of each satellite pixel. Based on the satellite image feature vector corresponding to each satellite pixel, the information content score of each satellite pixel is calculated. A joint optimization selection function is constructed based on the uncertainty metric and the information score. Based on the value of the joint optimization selection function, a portion of satellite pixels are selected from all satellite pixels to form a satellite training sample set, which is then used to train the satellite biomass inversion model.

2. The method for screening and optimizing biomass inversion samples according to claim 1, characterized in that, The lidar biomass inversion model is a random forest model; the process of inputting multiple lidar sample data into the trained lidar biomass inversion model to obtain the predicted biomass value and uncertainty metric value for each lidar sample data includes: For each lidar sample data, the lidar sample data is input into the trained lidar biomass inversion model to obtain the prediction results of each decision tree for the lidar sample data. The standard deviation of each prediction result is calculated to obtain the uncertainty measure of the lidar sample data.

3. The method for screening and optimizing biomass inversion samples according to claim 1, characterized in that, The step of scaling and aggregating the biomass prediction values ​​and uncertainty metrics of each lidar sample data corresponding to the same satellite pixel to obtain the biomass prediction value and uncertainty metric value for each satellite pixel includes: Calculate the intersection area of ​​each single-tree polygon that intersects with the first satellite pixel in spatial position with the first satellite pixel, and calculate the ratio of the intersection area to the area of ​​the single-tree polygon as the contribution ratio of the single-tree polygon to the first satellite pixel; wherein, the first satellite pixel is any satellite pixel, and each single-tree polygon corresponds to a lidar sample. The biomass prediction value of each single-tree polygon is multiplied by the corresponding contribution ratio in the first satellite pixel to obtain the biomass contribution value of the single-tree polygon in the first satellite pixel. The biomass contribution values ​​of all single-tree polygons in the first satellite pixel are summed to obtain the predicted biomass value of the first satellite pixel. The uncertainty metric of each single-tree polygon is multiplied by its corresponding contribution ratio in the first satellite pixel to obtain the uncertainty contribution value of that single-tree polygon in the first satellite pixel. The uncertainty contribution values ​​of all single-polygon units in the first satellite pixel are summed to obtain the uncertainty metric value of the first satellite pixel.

4. The method for screening and optimizing biomass inversion samples according to claim 1, characterized in that, The calculation of the information content score for each satellite pixel based on the satellite image feature vector corresponding to each satellite pixel includes: Clustering is performed on the satellite image feature vector corresponding to each satellite pixel to obtain multiple clusters; For each satellite pixel, the information content score of the satellite pixel is determined based on the reciprocal of the number of satellite image feature vectors in the cluster to which the satellite image feature vector of the corresponding satellite pixel belongs.

5. The method for screening and optimizing biomass inversion samples according to claim 1, characterized in that, The joint optimization selection function is: in, To jointly optimize the selection of function values, Information content score, This is a measure of uncertainty. This is a preset constant; The method of selecting a subset of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set includes: Among all satellite pixels, the N satellite pixels with the largest joint optimization selection function value are selected to form a satellite training sample set; where N is a preset value.

6. The method for screening and optimizing biomass inversion samples according to claim 1, characterized in that, Before inputting multiple lidar sample data into the trained lidar biomass inversion model, the method further includes: Acquire multiple sets of one-to-one corresponding ground sample data and lidar sample data; For each lidar sample data, the true biomass value of that lidar sample data is calculated based on the corresponding ground sample data. Using LiDAR sample data as training samples and corresponding ground truth biomass values ​​as sample labels, the initial random forest model is trained to obtain a trained LiDAR biomass inversion model.

7. The method for screening and optimizing biomass inversion samples according to any one of claims 1 to 6, characterized in that, After selecting a subset of satellite pixels from all available satellites based on the joint optimization selection function value to form a satellite training sample set, the following steps are also included: The satellite biomass inversion model was trained using the aforementioned satellite training sample set to obtain the trained satellite biomass inversion model. Biomass inversion mapping of the target area is performed based on a trained satellite biomass inversion model to obtain a satellite inversion biomass distribution map; wherein, the target area contains multiple sub-regions; Based on the trained lidar biomass inversion model, the target area is mapped by biomass inversion to obtain a radar-inverted biomass distribution map. The satellite inversion value of each sub-region in the satellite inversion biomass distribution map is compared with the predicted biomass value of the same sub-region in the radar inversion biomass distribution map, and the regions with differences greater than a preset threshold are identified as high conflict areas. The trained lidar biomass inversion model is locally recalibrated based on the high-conflict zones, and / or the high-conflict zones are marked as data acquisition areas to guide the supplementary acquisition of lidar sample data.

8. A device for screening and optimizing biomass inversion samples, characterized in that, include: An uncertainty measurement module is used to input multiple lidar sample data into a trained lidar biomass inversion model to obtain the predicted biomass value and uncertainty measurement value for each lidar sample data; wherein, the lidar biomass inversion model is trained based on a training sample composed of lidar sample data and true biomass values. The transformation and aggregation module is used to scale and aggregate the biomass prediction values ​​and uncertainty measure values ​​of each LiDAR sample data corresponding to the same satellite pixel to obtain the biomass prediction value and uncertainty measure value of each satellite pixel. The information content calculation module is used to calculate the information content score of each satellite pixel based on the satellite image feature vector corresponding to each satellite pixel. The joint optimization module is used to construct a joint optimization selection function based on the uncertainty metric and information score, and select a portion of satellite pixels from all satellite pixels based on the joint optimization selection function value to form a satellite training sample set, so as to train the satellite biomass inversion model based on the satellite training sample set.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images