Snow depth sample generation method and device, electronic equipment and computer storage medium

Abstract

The invention relates to a remote sensing data processing technology, and provides a snow depth sample generation method and device, electronic equipment and a computer storage medium, and the method comprises the steps: obtaining a snow depth observation sample of a target region; determining a target sub-region in which the snow depth observation samples in the sub-regions meet a preset sparse condition from a plurality of sub-regions of the target region; and generating a snow depth virtual sample of the target sub-region according to a snow depth sample generation model obtained by training by taking the sample distribution characteristics of the snow depth observation samples in the other sub-regions except the target sub-region in the target region or the statistical prior characteristics of the snow depth observation samples of the target region as constraint conditions. According to the method, the snow depth virtual sample conforming to the real environment can be generated, and the model training precision and generalization ability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing data processing, in particular to a snow depth sample generation method and device, electronic equipment and computer storage medium. BACKGROUND

[0002] Snow depth is a key parameter representing the physical characteristics of snow, and accurate estimation of the spatial and temporal distribution of snow depth is of great significance for climate prediction, water resource management and disaster warning. Satellite remote sensing is the main means of large-scale snow depth monitoring, and deep learning technology, as a powerful data-driven method, has shown great potential in snow depth remote sensing inversion.

[0003] However, the performance of deep learning models is highly dependent on large-scale, high-quality and well-spatial and temporal representative training samples. In the snow depth estimation task, ground observation stations are the main source of "ground truth" data. Due to geographical, climatic and economic conditions, these observation stations are sparsely distributed in key areas such as complex mountainous areas and forests, resulting in a "small sample learning" dilemma of deep learning models facing sample scarcity, uneven distribution and insufficient representation. This seriously restricts the generalization ability of the model and the estimation accuracy in complex areas.

[0004] Existing data enhancement methods, such as simple interpolation, generate samples that are too different from the real environment, resulting in low precision, poor generalization ability and unreliable results of deep learning models trained using such samples. SUMMARY

[0005] The present application aims to provide a snow depth sample generation method, device, electronic equipment and computer storage medium.

[0006] Embodiments of the present application can be implemented as follows: In a first aspect, the present application provides a snow depth sample generation method, comprising: obtaining snow depth observation samples of a target area; determining a target sub-area in which the snow depth observation samples in a sub-area meet a preset sparsity condition from a plurality of sub-areas of the target area; generating snow depth virtual samples of the target sub-area according to a trained snow depth sample generation model, with the sample distribution characteristics of the snow depth observation samples in the remaining sub-areas of the target area other than the target sub-area or the statistical prior characteristics of the snow depth observation samples of the target area as constraint conditions.

[0007] In an optional implementation, the step of determining a target sub-area in which the snow depth observation samples in a sub-area meet a preset sparsity condition from a plurality of sub-areas of the target area comprises: determining an actual value range of the snow depth observation sample in each of the sub-regions; obtaining a preset value range of the preset sample of each of the sub-regions; calculating a sample coverage rate of each of the sub-regions according to the actual value range and the preset value range of each of the sub-regions; regarding the sub-region with the sample coverage rate less than a preset ratio as the target sub-region.

[0008] In an optional implementation, the step of determining the target sub-region in which the snow depth observation sample in each of the sub-regions satisfies the preset sparse condition from the plurality of sub-regions of the target region comprises: calculating a target region kernel density estimation distribution of the snow depth observation sample in the target region; calculating a sub-region kernel density estimation distribution of the snow depth observation sample in each of the sub-regions; calculating a dispersion of each of the sub-region kernel density estimation distributions and the target region kernel density estimation distribution; regarding the sub-region with the dispersion greater than a preset threshold as the target sub-region.

[0009] In an optional implementation, the method further comprises: constructing a conditional generative adversarial network model, the conditional generative adversarial network model comprising a generator and a discriminator; obtaining a snow depth training sample sequence of the target region; determining a training sub-region in which a snow depth training sample in each of the sub-regions satisfies the preset sparse condition from the plurality of sub-regions of the target region; using the snow depth training sample sequence in the training sub-region as a constraint condition, alternately training the generator and the discriminator until a preset termination condition is satisfied, and obtaining the trained snow depth sample generation model; a loss function of the generator is determined according to an adversarial loss, a statistical feature matching loss, and a time continuity physical loss, the adversarial loss representing a distribution difference between a snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region, the statistical feature matching loss representing a statistical feature difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region, and the time continuity physical loss representing a difference between a time continuity feature of the snow depth estimation sample sequence generated by the generator and a preset time continuity feature.

[0010] In an optional implementation, the step of generating the snow depth virtual sample of the target sub-region according to the trained snow depth sample generation model, with the sample distribution feature of the snow depth observation sample in the remaining sub-region in the target region except the target sub-region or the statistical prior feature of the snow depth observation sample of the target region as the constraint condition, comprises: generating, by using the snow depth sample generation model, a plurality of groups of snow depth virtual samples with a number of times different from the number of the snow depth observation samples in the target sub-region; merging the snow depth observation sample in the target sub-region and each group of snow depth virtual samples respectively to obtain each group of snow depth reference samples; training, by using each group of snow depth reference samples, the pre-constructed snow depth estimation model to obtain an initial estimation model corresponding to each group of snow depth reference samples; taking an initial estimation model that meets a preset performance index from the plurality of initial estimation models as the snow depth estimation model of the target sub-region.

[0011] In an optional implementation, the step of taking an initial estimation model that meets a preset performance index from the plurality of initial estimation models as the snow depth estimation model of the target sub-region comprises: obtaining a snow depth estimation model of the remaining sub-region, the snow depth estimation model of the remaining sub-region being trained according to snow depth training samples in the remaining sub-region; performing snow depth estimation on the remaining sub-region according to the snow depth estimation model of the remaining sub-region to obtain a snow depth estimation result of the remaining sub-region; performing snow depth estimation on the target sub-region according to the snow depth estimation model of the target sub-region to obtain a snow depth estimation result of the target sub-region; splicing the snow depth estimation result of the target sub-region and the snow depth estimation result of the remaining sub-region to obtain a snow depth estimation result of the target region.

[0012] In an optional implementation, the sub-region to be estimated is the remaining sub-region or the target sub-region, and the step of performing snow depth estimation on the sub-region to be estimated to obtain a snow depth estimation result of the sub-region to be estimated comprises: dividing the sub-region to be estimated into a plurality of grids; inputting a snow depth observation sample in each grid into a snow depth estimation model of the sub-region to be estimated to obtain a snow depth estimation result of each grid; splicing the snow depth estimation results of all the grids to obtain a snow depth estimation result of the sub-region to be estimated.

[0013] In a second aspect, the present application provides a snow depth sample generation device, the device comprising: an acquisition module configured to acquire snow depth observation samples of a target region; a determination module configured to determine a target sub-region in which the snow depth observation samples in a sub-region of the target region satisfy a preset sparsity condition; a generation module configured to generate, as a constraint condition, a sample distribution feature of the snow depth observation samples in the remaining sub-regions of the target region except the target sub-region or a statistical prior feature of the snow depth observation samples of the target region, and generate snow depth virtual samples of the target sub-region according to a trained snow depth sample generation model.

[0014] In a third aspect, the present application provides an electronic device comprising a processor and a memory, wherein the memory is configured to store a program, and the processor is configured to implement the snow depth sample generation method according to any one of the preceding embodiments when executing the program.

[0015] In a fourth aspect, the present application provides a computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the snow depth sample generation method according to any one of the preceding embodiments.

[0016] Compared with the prior art, the present application has the following beneficial effects: The present application generates snow depth virtual samples of a target sub-region in which samples are sparse (i.e., the samples satisfy a preset sparsity condition) in a target region, using a sample distribution feature in the remaining sub-regions of the target region in which samples are not sparse (i.e., the samples do not satisfy the preset sparsity condition) or a statistical prior feature of the samples in the target region as a constraint condition, to ensure that the sample distribution feature of the generated snow depth virtual samples is consistent with the sample distribution feature of the samples in the remaining sub-regions in which samples are not sparse or consistent with the statistical prior feature of the samples in the target region, so that the generated snow depth virtual samples are more consistent with the real environment, and a model with high precision, strong generalization ability and reliable results can be trained using the snow depth virtual samples that are more consistent with the real environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0018] Figure 1 The flowchart of the snow depth sample generation method provided in the present embodiment is shown in the figure.

[0019] Figure 2 A framework example diagram of the snow depth sample generation model provided in the embodiment.

[0020] Figure 3 A framework example diagram of the snow depth estimation model provided in the embodiment.

[0021] Figure 4 An overall flow example diagram of the snow depth estimation provided in the embodiment.

[0022] Figure 5 An example diagram of the snow depth sample generation device provided in the embodiment. Figure 4 An example diagram of stage one in the embodiment.

[0023] Figure 6 An example diagram of the snow depth sample generation device provided in the embodiment. Figure 4 An example diagram of stage two in the embodiment.

[0024] Figure 7 A block example diagram of the snow depth sample generation device provided in the embodiment.

[0025] Figure 8 A block example diagram of the electronic device provided in the embodiment.

[0026] Icon: 10-electronic device; 11-processor; 12-memory; 13-bus; 100-snow depth sample generation device; 110-acquisition module; 120-determination module; 130-generation module; 140-estimation module. DETAILED DESCRIPTION

[0027] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0028] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0029] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0030] In the description of the present application, it should be noted that if the terms "upper", "lower", "inner", "outer" and the like indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the present application is usually placed, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0031] In addition, if the terms "first", "second" and the like appear, they are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0032] It should be noted that the features in the embodiments of the present application can be combined with each other without conflict.

[0033] Please refer to Figure 1 , Figure 1 The flowchart of the snow depth sample generation method provided in the present embodiment includes the following steps: Step S101, obtaining snow depth observation samples of a target area.

[0034] In the present embodiment, the target area refers to a large geographical area for which snow depth estimation is required, for example, it can be the entire national area, or a part of the area, for example, the northwest area, the southwest area, etc. The snow depth observation samples can include snow depth values collected by ground stations and matching multi-source remote sensing data, meteorological data, terrain data and other multi-dimensional prediction factors.

[0035] Step S102, determining a target sub-area in which the snow depth observation samples in the sub-area of the target area satisfy a preset sparsity condition.

[0036] In the present embodiment, the target area can be divided into a plurality of sub-areas according to preset geographical attributes or environmental attributes, the geographical attributes include but are not limited to terrain attributes, vegetation attributes, etc., and the environmental attributes include but are not limited to climate and weather attributes, remote sensing data attributes, etc. The target sub-area refers to a sub-area in which the snow depth observation samples satisfy the preset sparsity condition, i.e., a sub-area with insufficient sample representation, and conversely, a sub-area that does not satisfy the preset sparsity condition is a sub-area with sufficient sample representation.

[0037] For example, the target area is the entire national area, and the target area is divided into three sub-areas of forest sub-area, high mountain sub-area and other sub-area according to the terrain attribute (elevation > 2500m or surface relief > 200m) and the vegetation attribute (forest coverage > 20%). The high mountain sub-area is the area outside the forest coverage area.

[0038] Step S103. Snow depth virtual samples of the target sub-region are generated according to the trained snow depth sample generation model, with the sample distribution characteristics of the snow depth observation samples in the remaining sub-regions in the target region except the target sub-region or the statistical prior characteristics of the snow depth observation samples in the target region as constraint conditions.

[0039] In the embodiment, since the remaining sub-regions are sub-regions with sufficient representative samples, the sample distribution characteristics of the snow depth observation samples in the sub-regions can include joint distribution characteristics between the multi-dimensional prediction factors and the snow depth values of the sub-regions. With the sample distribution characteristics as constraint conditions, the generated snow depth virtual samples can be ensured to be reasonable in physical laws and more consistent with the actual scene. In addition, the statistical prior characteristics of the snow depth observation samples in the target region can also be used as constraint conditions. The statistical prior characteristics can include statistical characteristics of the snow depth of the target region as a whole, such as mean value, extreme value, seasonal variation law, etc. With the statistical prior characteristics as constraint conditions, the generated snow depth virtual samples can be ensured to be consistent with the real samples in statistical distribution and time evolution trend.

[0040] In the embodiment, the snow depth virtual samples can be data augmentation and supplement of the snow depth observation samples in the target sub-region, so as to effectively make up for the problem of insufficient number and poor representativeness of the original samples in the target sub-region.

[0041] The above method provided by the embodiment is used for the target sub-region with sample sparsity (i.e., the sample meets the preset sparsity condition) in the target region, and snow depth virtual samples of the target sub-region are generated according to the trained snow depth sample generation model, with the sample distribution characteristics of the samples in the remaining sub-regions with non-sparsity (i.e., the sample does not meet the preset sparsity condition) in the target region or the statistical prior characteristics of the samples in the target region as constraint conditions. The sample distribution characteristics of the generated snow depth virtual samples are consistent with the sample distribution characteristics of the samples in the remaining sub-regions with non-sparsity, or consistent with the statistical prior characteristics of the samples in the target region, so that the generated snow depth virtual samples are more consistent with the real environment, and a model with high precision, strong generalization ability and reliable result can be trained by using the snow depth virtual samples more consistent with the real environment.

[0042] In optional embodiments, in order to improve the rationality of the determination of the target sub-region, the embodiment provides at least two ways to determine the target sub-region: Way one: First, the actual value range of the snow depth observation samples in each sub-region is determined. Second, the preset value range of the preset samples of each sub-region is obtained. In the embodiment, the preset value range is the theoretical maximum range or the empirically reasonable range of the possible values of the sub-region in each predictor dimension, for example, the minimum value and the maximum value of a certain remote sensing factor that theoretically can occur in the geographical region within a certain period of time. The preset value range can be set based on historical data statistics or expert experience as a reference benchmark for sample coverage calculation.

[0043] Third, according to the actual value range and the preset value range of each sub-region, the sample coverage of each region is calculated; In the embodiment, the sample coverage is the proportion of the value range of the snow depth observation sample in the sub-region to the preset value range of the sub-region, which is usually expressed in percentage. If the actual value range of a certain sub-region is (a1, a2) and the preset value range is (b1, b2), the sample coverage of the sub-region is (a2-a1) / (b2-b1). For example, if the preset value range of a certain sub-region of a certain predictor is 0-100, and the value range of the actual observation sample of the region is 30-80, the sample coverage of the factor is 50%.

[0044] It should be noted that for multi-dimensional predictors, the coverage of each dimension can be calculated respectively, and then the average or weighted average is taken as the comprehensive coverage. Fourth, the sub-region with a sample coverage less than a preset ratio is taken as the target sub-region.

[0045] In the embodiment, the preset ratio can be set according to the actual scene needs, for example, the preset ratio can be set to 50%, if the coverage of a certain sub-region is lower than the threshold, it is considered that the sample of the region is not representative enough, and it belongs to the target sub-region that needs data enhancement. For example, in the Forest, Alpine and other sub-regions divided from the target region, the coverage of Forest and Alpine regions is 42.47% and 48.32% respectively, which is lower than the threshold of 50%, so they are identified as target sub-regions, while the coverage of Other region is 79.14%, which belongs to the remaining sub-region and does not need to be enhanced.

[0046] The first method is especially suitable for snow depth estimation tasks driven by remote sensing data, in which regions with sparse ground observation sites and uneven sample distribution are widespread. By quantifying the coverage index, the sub-regions with good and poor representativeness can be effectively distinguished, and the pertinence and effectiveness of the data enhancement strategy are ensured.

[0047] The second method: First, the target region kernel density estimation distribution of the snow depth observation sample in the target region is calculated; In the embodiment, the target region kernel density estimation distribution reflects the joint probability density distribution of the snow depth samples in the target region in each prediction factor dimension, and is used as a global reference for the sub-region sample distribution.

[0048] Secondly, the sub-region kernel density estimation distribution of the snow depth observation samples in each sub-region is calculated. In the embodiment, the sub-region kernel density estimation distribution reflects the joint probability density distribution of the snow depth samples in each prediction factor dimension.

[0049] Thirdly, the divergence of each sub-region kernel density estimation distribution and the target region kernel density estimation distribution is calculated. In the embodiment, the divergence refers to the difference between the sub-region sample distribution and the overall distribution of the target region, and is usually measured by statistical quantities such as KL divergence (Kullback-Leibler Divergence), JS divergence (Jensen-Shannon Divergence), or Euclidean distance. The greater the divergence value, the higher the deviation of the sub-region sample distribution from the overall distribution of the target region, and the poorer the sample representativeness. As an implementation manner, the divergence can be represented by KDE (Kernel Density Estimation) divergence, which represents the deviation between the joint probability density distribution of the snow depth samples in each prediction factor dimension of the target region and the joint probability density distribution of the snow depth samples in each prediction factor dimension of the sub-region.

[0050] Fourthly, the sub-region with a divergence greater than a preset threshold is regarded as a target sub-region.

[0051] In the embodiment, the preset threshold is set according to actual task requirements, for example, 10%. If the divergence of a sub-region exceeds the threshold, it is considered that the sample representativeness of the sub-region is insufficient, and the sub-region belongs to the target sub-region that needs to be enhanced. For example, the average KDE divergence of the Forest and Alpine sub-regions is 17.27% and 13.16% respectively, both of which exceed the threshold of 10%, and therefore are identified as target sub-regions; the KDE divergence of the Other region is only 3.65%, which does not exceed the threshold, and therefore does not belong to the target sub-region.

[0052] The second method establishes an overall sample distribution model through the KDE distribution of the target region, and then analyzes the KDE distribution of each sub-region and compares it with the overall model, so as to identify the sub-region with a larger sample distribution bias. This method not only considers the sparsity of the sample quantity, but also further improves the accuracy and robustness of the representativeness evaluation from the consistency of the sample distribution.

[0053] It should be noted that in some scenarios, mode one and mode two can be combined for use, that is, a sub-region that meets both mode one and mode two is taken as a target sub-region.

[0054] In an optional embodiment, in order to improve the quality of the generated snow depth virtual sample, the embodiment further provides an implementation of obtaining a trained snow depth sample generation model: Firstly, a conditional generative adversarial network model is constructed, and the conditional generative adversarial network model includes a generator and a discriminator; In the embodiment, the conditional generative adversarial network (cGAN) is an extended form of the generative adversarial network (GAN), and the core idea is to guide the generator to generate samples with specific attributes or characteristics by introducing additional condition information. In the cGAN, the generator and the discriminator both receive an additional condition input (such as the constraint condition in the embodiment), so that the generator can generate samples that meet the constraint condition, and the discriminator also refers to the constraint condition when judging the authenticity of the sample. The conditional generative adversarial network model is a model constructed based on the conditional generative adversarial network.

[0055] Secondly, a snow depth training sample sequence of the target region is obtained; In the embodiment, the snow depth training sample sequence is derived from observation data of multiple sub-regions in the target region, and usually exists in the form of a time sequence, for example, observation records of snow depth values and multi-dimensional prediction factors matched therewith changing over time.

[0056] Thirdly, a training sub-region in which the snow depth training samples in a sub-region meet a preset sparse condition is determined from the multiple sub-regions in the target region; In the embodiment, the manner of determining the training sub-region is similar to the manner of determining the target sub-region described above, and thus will not be described herein.

[0057] Fourthly, the snow depth training sample sequence in the training sub-region is used to alternately train the generator and the discriminator with the sample distribution characteristics of the snow depth training sample sequence in the sub-region other than the training sub-region in the target region or the statistical prior characteristics of the snow depth training sample sequence of the target region as a constraint condition until a preset termination condition is met, so as to obtain a trained snow depth sample generation model; The loss function of the generator is determined according to an adversarial loss, a statistical feature matching loss and a time continuity physical loss. The adversarial loss represents the distribution difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The statistical feature matching loss represents the statistical feature difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The time continuity physical loss represents the difference between the time continuity feature of the snow depth estimation sample sequence generated by the generator and the preset time continuity feature.

[0058] In the embodiment, the preset termination condition can be that the training error reaches a convergence state or reaches a set training round.

[0059] In the embodiment, the snow depth estimation sample sequence generated by the generator is a generated sample, the snow depth training sample sequence in the training sub-region is a real sample, the adversarial loss aims to improve the consistency of the generated sample with the real sample in the distribution level, the statistical feature matching loss is used to measure the difference between the generated sample and the real sample in the statistical feature (such as mean, variance, spatial correlation, etc.), and ensure that the generated sample is consistent with the real sample in the statistical characteristics; the time continuity physical loss focuses on the continuity performance of the generated sample in the time dimension, and measures whether the time variation trend of the generated sample conforms to the preset time continuity feature, for example, the natural evolution law of the snow depth with the season change. The joint optimization of the three types of losses, that is, the adversarial loss, the statistical feature matching loss and the time continuity physical loss, helps to improve the authenticity and physical rationality of the generated sample in the distribution, statistics and time evolution three dimensions.

[0060] As a specific implementation manner, the conditional generative adversarial network model is a conditional Wasserstein generative adversarial network gradient penalty model (ConWGP), and details can be referred to Figure 2 , Figure 2 A framework example diagram of the snow depth sample generation model provided in the embodiment is shown in FIG. 1. Figure 2 In the ConWGP model, the generator and the discriminator both adopt a bidirectional gated recurrent unit (Bi-GRU) as a backbone to effectively capture the time sequence dependency of the snow depth training sample sequence. When training the model, a constraint condition is introduced for guidance, for example, the sample distribution feature of a representative sufficient region or the statistical prior (such as mean, extreme value, etc.) of the target region can be used as a condition. The loss of the generator of the model includes not only the standard adversarial loss, but also the statistical feature matching loss (to ensure that the generated sample is consistent with the real sample in the feature statistics of the intermediate layer of the discriminator) and the time continuity physical constraint loss (to ensure that the time sequence change of the generated sample is smooth and conforms to the physical law). After the training is completed, the virtual snow depth sample sequence can be generated by using the model.

[0061] In an optional implementation, after training the snow depth sample generation model, in order to train a snow depth estimation model for the target sub-region with stronger predictive ability using the samples generated by the snow depth sample generation model, this embodiment also provides an implementation method for finally determining the snow depth estimation model for the sub-region using a screening mechanism: First, using a snow depth sample generation model, multiple sets of virtual snow depth samples are generated, with the number of samples being different multiples of the number of snow depth observation samples in the target sub-region. In this embodiment, different multiples can be set according to actual needs. For example, the number of generated virtual snow depth samples is 1 to 10 times the number of snow depth observation samples in the target sub-region, totaling 10 sets.

[0062] Secondly, the snow depth observation samples in the target sub-region are merged with each group of virtual snow depth samples to obtain each group of snow depth reference samples. In this embodiment, each set of snow depth reference samples consists of snow depth observation samples and a set of virtual snow depth samples, thereby significantly expanding the effective sample set within the target sub-region. The merging method can be simple data overlay, or it can be overlaying first and then processing the overlayed samples to ensure that each set of snow depth reference samples after processing is consistent with the snow depth observation samples within the target sub-region in terms of statistical distribution, spatial characteristics, and temporal evolution, so as to avoid introducing biases that affect the model training effect.

[0063] Third, the pre-built snow depth estimation model is trained using each set of snow depth reference samples to obtain the initial estimation model corresponding to each set of snow depth reference samples. In this embodiment, the snow depth estimation model can be a Bayesian bidirectional gated recurrent unit (BBGRU) model, which can simultaneously output the snow depth estimate and the corresponding prediction uncertainty. Please refer to... Figure 3 , Figure 3 This is a framework example diagram of the snow depth estimation model provided in this embodiment. Figure 3 The BBGRU snow depth estimation model in this paper is based on a three-layer Bi-GRU, with a Bayesian fully connected layer following the last layer to output the predicted mean and variance. During the training process of the pre-built snow depth estimation model, each set of snow depth reference samples is used as input. Through model parameter optimization, the model learns the spatial distribution patterns and variation laws of snow depth data, thereby gaining the ability to estimate the snow depth in a target sub-region.

[0064] Fourth, the initial estimation model that meets the preset performance index among multiple initial estimation models is used as the snow depth estimation model for the target sub-region.

[0065] In this embodiment, the preset performance metrics can be the mean squared error and coefficient of determination of the model on the validation set. evaluation indexes such as MAE, and can also be the stability performance of the model on the snow depth reference samples in each group, for example, whether the model trained on multiple groups of snow depth reference samples has similar prediction performance.

[0066] Through the above screening mechanism, low-performance models caused by low-quality snow depth virtual samples or unstable training process can be effectively eliminated, the diversity and representativeness of the training data of the snow depth estimation model are effectively improved, the prediction ability of the model in the region is enhanced, and the robustness and generalization ability of the finally selected snow depth estimation model are improved.

[0067] In an optional implementation, in order to accurately estimate the snow depth of the entire target region, the target sub-region and the remaining sub-regions are estimated separately, and the estimation results are finally combined. One implementation is as follows: First, the snow depth estimation model of the remaining sub-regions is obtained, which is trained according to the snow depth training samples in the remaining sub-regions; In this embodiment, since the snow depth training samples in the remaining sub-regions are sufficient, have high sample density and representativeness, and can accurately reflect the spatial distribution rule and variation trend of the snow depth in the remaining sub-regions, when there are multiple remaining sub-regions, a corresponding snow depth estimation model can be trained according to the snow depth training samples in each remaining sub-region.

[0068] Second, the snow depth of the remaining sub-regions is estimated according to the snow depth estimation model of the remaining sub-regions, and the snow depth estimation result of the remaining sub-regions is obtained; Third, the snow depth of the target sub-region is estimated according to the snow depth estimation model of the target sub-region, and the snow depth estimation result of the target sub-region is obtained; In this embodiment, as can be known from the generation process of the snow depth estimation model of the target sub-region above, the snow depth estimation model is trained based on the snow depth reference samples formed by combining the multiple groups of snow depth virtual samples generated by the snow depth sample generation model and the original observation samples. The initial estimation model meeting the preset performance index is selected by training multiple groups of reference data of different sample sizes, so as to ensure that the estimation model still has good prediction ability and stability under the condition of sample sparseness.

[0069] Fourth, the snow depth estimation result of the target sub-region and the snow depth estimation result of the remaining sub-regions are spliced to obtain the snow depth estimation result of the target region.

[0070] In the embodiment, the splicing is the integration in the spatial dimension, for example, the estimation results of each sub-region are integrated according to the geographical boundary, to ensure the continuity and consistency of the estimation values between the sub-regions. The splicing process can be the superposition of the snow depth estimation results, or the processing of the superposed snow depth estimation results in the manner of considering the transition zone between adjacent sub-regions, spatial interpolation correction, data format unification, and the like, so that the final snow depth estimation results of the target region reach the unified standard in the spatial coverage range and the data quality.

[0071] In the optional implementation, in order to improve the spatial resolution and local accuracy of the estimation results, the embodiment provides an implementation manner of dividing the sub-region into a grid, and performing snow depth estimation on each grid to obtain the snow depth estimation results. This manner can be applied to the remaining sub-regions or the target sub-region. For the convenience of description, the two kinds of sub-regions are collectively referred to as a to-be-estimated sub-region. The implementation manner of performing snow depth estimation on the to-be-estimated sub-region to obtain the snow depth estimation results of the to-be-estimated sub-region can be as follows: First, the to-be-estimated sub-region is divided into a plurality of grids. In the embodiment, the division of the to-be-estimated sub-region into a plurality of grids means that the to-be-estimated region is regularly divided in the spatial dimension to form a plurality of non-overlapping grid units with a fixed spatial resolution. The grid is a common discretization processing manner in geographical spatial analysis, each grid corresponds to a certain geographical range, and serves as a basic spatial unit for snow depth estimation. The purpose of dividing the grid is to decompose the estimation task of the entire to-be-estimated sub-region into a plurality of local spatial unit estimation tasks, thereby improving the spatial resolution and local accuracy of the estimation results.

[0072] In the embodiment, the division of the grid can be based on the spatial rasterization technology in the geographical information system, and the continuous geographical space is divided into a grid structure with uniform size and shape according to a preset spatial resolution (such as 1 km x 1 km, 500 m x 500 m, etc.). The fineness of the grid division can be adjusted according to the actual application requirements, the limitation of the computing resources, and the availability of the data, to ensure that the division result can meet the estimation accuracy requirements and will not cause excessive computing burden.

[0073] Secondly, the snow depth observation samples in each grid are input into the snow depth estimation model of the to-be-estimated sub-region to obtain the snow depth estimation result of each grid. In the embodiment, the snow depth estimation processes of a plurality of different grids can be executed in parallel to improve the snow depth estimation performance of the to-be-estimated sub-region.

[0074] Finally, the snow depth estimation results of all the grids are spliced to obtain the snow depth estimation result of the to-be-estimated sub-region.

[0075] In this embodiment, stitching together the snow depth estimation results of all grids refers to arranging and integrating the estimation results of each grid in an orderly manner according to the geographic coordinate index at the time of grid division, ensuring that the estimation results meet the specifications of geospatial data in terms of spatial continuity and boundary consistency. During the stitching process, operations such as data format conversion, missing value imputation, and edge smoothing can be used to improve the visualization effect and spatial consistency of the overall estimation results.

[0076] To make the above snow depth estimation process more intuitive, please refer to... Figure 4 , Figure 4 This is an example diagram illustrating the overall snow depth estimation process provided in this embodiment. Figure 4 In China, the snow depth estimation process includes the following four stages: Phase 1: Where are the synthetic samples generated? Phase 1 includes data collection, sub-region division, snow depth sample allocation, and sample representativeness diagnosis, ultimately resulting in sub-regions that do not require sample data augmentation (i.e., other sub-regions) and sub-regions that require sample augmentation (i.e., target sub-regions).

[0077] Please refer to Figure 5 , Figure 5 Provided for this embodiment Figure 4 Example diagram of Phase 1, Figure 5 The remote sensing data includes, but is not limited to, TB (Top of Atmosphere Brightness Temperature), FTC (Freeze / Thaw Cycle), and Topo (Topography).

[0078] Phase Two: How to generate samples? In Phase Two, a snow depth sample generation model is trained using constraints, and the trained model is then used to generate virtual snow depth samples. Please refer to [link / reference needed]. Figure 6 , Figure 6 Provided for this embodiment Figure 4 Example diagram of Phase 2, Figure 6 In this model, both the generator and discriminator employ a three-layer Bi-GRU network. The generator's input consists of random noise and constraints, and it comprises a latent variable z, a three-layer Bi-GRU, a fully connected FC layer, and an output layer. The discriminator comprises an input layer, a three-layer Bi-GRU, a temporal attention layer, a fully connected FC layer, and an output layer.

[0079] Phase 3: How many synthetic samples need to be generated? In the third stage, a plurality of combined samples (i.e., original snow depth samples and virtual snow depth samples with different multiples of the original snow depth samples) are generated by using the trained snow depth sample generation model, and a plurality of snow depth estimation models are trained by using the plurality of combined samples, and the optimal number of combined samples is evaluated.

[0080] Stage four: large-scale regional snow depth estimation In the fourth stage, for the sub-regions that do not require sample data enhancement, a snow depth estimation model is constructed based on Bayesian Bi-GRU using the original real samples, for the sub-regions that require sample data enhancement, a snow depth estimation model is constructed based on Bayesian Bi-GRU using the optimal combined samples, and the respective snow depth estimation models are used to estimate the snow depth of the respective sub-regions, and finally the snow depth estimation results in the nationwide range are obtained.

[0081] Based on Figure 4 The embodiment also provides a specific example of estimating the snow depth distribution of the snow season of 2019-2020 in the nationwide region as a target region, including the following steps: Step one: sample representativeness diagnosis (1) Data preparation: obtain the daily snow depth observation data of 457 ground stations of China Meteorological Administration as real data. Obtain the corresponding multi-source prediction factor data, including AMSR2 passive microwave brightness temperature, MODIS NDSI, terrain data (elevation, slope, etc.) and meteorological forcing data (air temperature, precipitation, etc.), a total of 37-dimensional time series prediction factors. All data are unified to 0.1° spatial resolution to construct a time series sample set.

[0082] (2) Regional division: according to the forest coverage (FTC> 20%) and the terrain complexity (elevation> 2500m or surface relief> 200m), the nationwide region is divided into three sub-regions of forest (Forest), alpine (Alpine) and other (Other). Alpine is the region outside the forest coverage area.

[0083] (3) Representative evaluation: The coverage rate of the samples of each sub-region is calculated. The results show that the average coverage rates of the Forest and Alpine regions are 42.47% and 48.32% respectively, both of which are lower than the threshold of 50%. The Other region is 79.14%.

[0084] The KDE distribution of the samples of each sub-region is calculated and compared with the KDE distribution of the overall characteristics of the region. The results show that the average KDE dispersion of the Forest and Alpine regions is 17.27% and 13.16% respectively, both of which are higher than the threshold of 10%. The Other region is only 3.65%.

[0085] Conclusion: In general, Forest and Alpine regions are under-represented and need data augmentation. Other region is well-represented and does not need augmentation.

[0086] Step two: Conditioned sample generation For Forest and Alpine regions, we construct ConWGP models to augment data.

[0087] (1) Model architecture: Both generator and discriminator use three-layer Bi-GRU networks. The input of the generator is random noise and conditional information (here we use the sample distribution of Other region as strong prior condition), and the output is the synthesized 38-dimensional time series samples (37-dimensional predictors + 1-dimensional snow depth).

[0088] (2) Loss function: We use a composite loss function for training, including: Adversarial loss: based on Wasserstein distance and gradient penalty.

[0089] Statistical feature matching loss: force the generated data to match the mean and standard deviation of the real data at the intermediate layer of the discriminator.

[0090] Temporal continuity loss: L1 regularization on the change of adjacent time steps of the generated snow depth sequence to ensure its physical reasonableness.

[0091] (3) Sample generation: After training, we generate virtual sample sets 1 to 10 times the original number of real samples for Forest and Alpine regions, respectively.

[0092] Step three: Optimal augmentation scale determination (1) Model construction: We construct a BBGRU snow depth estimation model. This model is based on three-layer Bi-GRU and connects a Bayesian fully connected layer after the last layer to output the predicted mean and variance.

[0093] (2) Performance evaluation: In the Alpine region, we combine 20 real training samples with 1-10 times (i.e. 20-200) virtual samples to train 10 different BBGRU models. Evaluation on the independent test set shows that when introducing 6 times virtual samples, the model's reaches the highest (0.9), the RMSE and MAE are the lowest, and the prediction uncertainty is also at a relatively low and stable level (Table 1). Therefore, we determine the optimal augmentation scale for the Alpine region as 6 times.

[0094] In the Forest region, we perform the same operation. Evaluation shows that when introducing 7 times virtual samples, the model performs best on the test set ( =0.8). Therefore, the optimal augmentation scale of the Forest region is determined to be 7 times.

[0095] The performance of the Bayesian BI-GRU snow depth estimation model on the three sub-regions is quantitatively evaluated in Table 1.

[0096] Table 1

[0097] T1 in Table 1 is the training set, T2 is the test set, and the units of RMSE and MAE are cm. X1-X10 are combinations of real training samples and virtual samples with 1-10 times.

[0098] Step four: large-scale snow depth estimation and synthesis (1) Model training: Alpine region: The final BBGRU model is trained using an augmented training set consisting of its original 21 real samples and 122 (6 times) optimal virtual samples.

[0099] Forest region: The final BBGRU model is trained using an augmented training set consisting of its original 33 real samples and 226 (7 times) optimal virtual samples.

[0100] Other region: The BBGRU model is directly trained using its 313 original real samples.

[0101] (2) Snow depth mapping: The trained three models are applied to all grids in the corresponding sub-region, and the predicted factor data is input to obtain the snow depth and uncertainty distribution of each sub-region.

[0102] (3) Result synthesis: The estimated results of the three sub-regions are spliced to form the snow depth distribution map and uncertainty map of the 2012-2021 snow season covering the entire national region.

[0103] In order to perform the corresponding steps in the above embodiments and various possible implementation manners, an implementation manner of a snow depth sample generation device 100 is given below. Please refer to Figure 7 , Figure 7 The block diagram of the snow depth sample generation device provided in this embodiment needs to be explained. The basic principle and technical effects of the snow depth sample generation device 100 provided by the present application are the same as those of the corresponding above-mentioned embodiments, and for brief description, part of this embodiment is not mentioned.

[0104] The snow depth sample generation device 100 includes an acquisition module 110, a determination module 120, and a generation module 130.

[0105] The acquisition module 110 is configured to acquire snow depth observation samples of a target region. determining, by a determining module 120, a target sub-region in which the snow depth observation samples in the sub-region satisfy a preset sparsity condition from a plurality of sub-regions of the target region; generating, by a generating module 130, a snow depth virtual sample of the target sub-region according to the trained snow depth sample generation model, with a sample distribution feature of the snow depth observation samples in the remaining sub-regions of the target region except the target sub-region or a statistical prior feature of the snow depth observation samples of the target region as a constraint condition.

[0106] In an optional implementation, the determining module 120 is specifically configured to: determine an actual value range of the snow depth observation samples in each sub-region; obtain a preset value range of the preset samples of each sub-region; calculate a sample coverage rate of each sub-region according to the actual value range and the preset value range of each sub-region; regard the sub-region with the sample coverage rate less than a preset ratio as the target sub-region.

[0107] In an optional implementation, the determining module 120 is specifically further configured to: calculate a target region kernel density estimation distribution of the snow depth observation samples in the target region; calculate a sub-region kernel density estimation distribution of the snow depth observation samples in each sub-region; calculate a dispersion of the sub-region kernel density estimation distribution and the target region kernel density estimation distribution; regard the sub-region with the dispersion greater than a preset threshold as the target sub-region.

[0108] In an optional implementation, the generating module 130 is further configured to: construct a conditional generative adversarial network model, the conditional generative adversarial network model including a generator and a discriminator; obtain a snow depth training sample sequence of the target region; determine a training sub-region in which the snow depth training samples in the sub-region satisfy a preset sparsity condition from a plurality of sub-regions of the target region; use the snow depth training sample sequence in the training sub-region, alternately train the generator and the discriminator with a sample distribution feature of the snow depth training sample sequence in the sub-region of the target region except the training sub-region or a statistical prior feature of the snow depth training sample sequence of the target region as a constraint condition, until a preset termination condition is satisfied, and obtain a trained snow depth sample generation model; The loss function of the generator is determined according to an adversarial loss, a statistical feature matching loss and a time continuity physical loss. The adversarial loss represents a distribution difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The statistical feature matching loss represents a statistical feature difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The time continuity physical loss represents a difference between a time continuity feature of the snow depth estimation sample sequence generated by the generator and a preset time continuity feature.

[0109] The snow depth sample generation apparatus 100 further comprises an estimation module 140.

[0110] The estimation module 140 is configured to: generate a plurality of groups of snow depth virtual samples with a number of times of a number of snow depth observation samples in the target sub-region by using the snow depth sample generation model; merge the snow depth observation samples in the target sub-region and each group of snow depth virtual samples respectively to obtain each group of snow depth reference samples; train the pre-constructed snow depth estimation model by using each group of snow depth reference samples to obtain an initial estimation model corresponding to each group of snow depth reference samples; select an initial estimation model satisfying a preset performance index from the plurality of initial estimation models as the snow depth estimation model of the target sub-region.

[0111] In an optional implementation, the estimation module 140 is further configured to: obtain snow depth estimation models of the remaining sub-regions, wherein the snow depth estimation models of the remaining sub-regions are trained according to snow depth training samples in the remaining sub-regions; perform snow depth estimation on the remaining sub-regions according to the snow depth estimation models of the remaining sub-regions to obtain snow depth estimation results of the remaining sub-regions; perform snow depth estimation on the target sub-region according to the snow depth estimation model of the target sub-region to obtain a snow depth estimation result of the target sub-region; splicing the snow depth estimation result of the target sub-region and the snow depth estimation results of the remaining sub-regions to obtain a snow depth estimation result of the target region.

[0112] In an optional implementation, the sub-region to be estimated is the remaining sub-region or the target sub-region, and the estimation module 140 is specifically further configured to: divide the sub-region to be estimated into a plurality of grids; input snow depth observation samples in each grid into the snow depth estimation model of the sub-region to be estimated to obtain a snow depth estimation result of each grid; splicing the snow depth estimation results of all the grids to obtain a snow depth estimation result of the sub-region to be estimated.

[0113] The embodiment of the present application further provides a block schematic diagram of the electronic device 10, and the electronic device 10 implements the snow depth sample generation method of the foregoing embodiment. Figure 8 , Figure 8 The block schematic diagram of the electronic device 10 is provided for the embodiment, and the electronic device 10 comprises a processor 11, a memory 12 and a bus 13, and the processor 11 and the memory 12 are connected through the bus 13.

[0114] The processor 11 can be an integrated circuit chip with a signal processing capability. In the implementation process, each step of the snow depth sample generation method of the foregoing embodiment can be completed by the integrated logic circuit of hardware or the instruction in the form of software in the processor 11. The processor 11 can be a general processor, including a CPU (Central Processing Unit, central processor), an NP (Network Processor, network processor) and the like; and can also be a DSP (Digital Signal Processor, digital signal processor), an ASIC (Application Specific Integrated Circuit, application specific integrated circuit), an FPGA (Field Programmable Logic Gate Array, field programmable logic gate array) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0115] The memory 12 is used for storing a program for implementing the snow depth sample generation method, and the program can be a software function module stored in the memory 12 in the form of software or firmware or solidified in the OS (Operating System, operating system) of the electronic device 10.

[0116] The processor 11 executes the program to implement the snow depth sample generation method of the foregoing embodiment after receiving an execution instruction.

[0117] The embodiment provides a computer storage medium, and a computer program is stored on the computer storage medium, and the computer program is executed by the processor to implement the snow depth sample generation method as described in the foregoing embodiment.

[0118] In summary, the embodiment of the present application provides a snow depth sample generation method and device, electronic equipment and computer storage medium, the method comprises: obtaining the snow depth observation sample of the target area; determining the target sub-area in which the snow depth observation sample in the sub-area of the target area meets the preset sparse condition from the plurality of sub-areas of the target area; taking the sample distribution characteristics of the snow depth observation sample in the remaining sub-area of the target area except the target sub-area or the statistical prior characteristics of the snow depth observation sample of the target area as the constraint condition, and generating the snow depth virtual sample of the target sub-area according to the trained snow depth sample generation model. Compared with the prior art, the present embodiment has at least the following advantages: (1) for the target sub-area in which the sample in the target area is sparse (that is, the sample meets the preset sparse condition), the sample distribution characteristics of the remaining sub-area in which the sample is not sparse (that is, the sample does not meet the preset sparse condition) or the statistical prior characteristics of the sample in the target area are taken as the constraint condition, and the snow depth virtual sample of the target sub-area is generated according to the trained snow depth sample generation model, which ensures that the sample distribution characteristics of the generated snow depth virtual sample are consistent with the sample distribution characteristics of the remaining sub-area in which the sample is not sparse, or consistent with the statistical prior characteristics of the sample in the target area, so that the generated snow depth virtual sample is more consistent with the real environment, and then a model with high precision, strong generalization ability and reliable result can be trained using the snow depth virtual sample which is more consistent with the real environment; (2) a closed-loop framework from "sample diagnosis-sample enhancement-result estimation-result evaluation" is constructed, and the core problem of small sample remote sensing inversion is systematically solved; (3) by using ConWGP and combining physical constraints, the generated virtual sample performs excellently in statistical consistency, diversity and physical rationality, effectively avoiding model collapse and providing high-quality training data for downstream models; (4) by performing targeted data enhancement on the sample sparse area, the generalization ability and estimation accuracy of the deep learning model in these complex areas are significantly improved, and the overfitting problem is effectively alleviated; (5) by using the Bayesian deep learning model, not only high-precision snow depth estimation values can be provided, but also the uncertainty of each prediction result can be quantified, providing key reliability information for scientific research and decision-making applications; (6) it has good scalability and is not only suitable for snow depth estimation, but also can be extended to small sample remote sensing inversion tasks of soil moisture, vegetation parameters and other surface parameters.

[0119] The above is only various embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generating snow depth samples, characterized in that, The method includes: Obtain snow depth observation samples for the target area; From multiple sub-regions of the target region, determine the target sub-regions within which the snow depth observation samples satisfy the preset sparsity condition; Using the sample distribution characteristics of snow depth observation samples in the remaining sub-regions of the target region other than the target sub-region, or the statistical prior characteristics of snow depth observation samples in the target region as constraints, a virtual snow depth sample of the target sub-region is generated according to the trained snow depth sample generation model.

2. The method according to claim 1, characterized in that, The step of determining the target sub-region from multiple sub-regions of the target region where the snow depth observation samples satisfy the preset sparsity condition includes: Determine the actual range of snow depth observation samples within each sub-region; Obtain the preset value range of the preset sample for each sub-region; Calculate the sample coverage rate of each sub-region based on the actual value range and the preset value range of each sub-region; The sub-regions with sample coverage less than a preset ratio are designated as the target sub-regions.

3. The method according to claim 1, characterized in that, The step of determining the target sub-region from multiple sub-regions of the target region where the snow depth observation samples satisfy the preset sparsity condition includes: Calculate the target area kernel density estimation distribution of snow depth observation samples within the target area; Calculate the subregion kernel density estimate distribution of snow depth observation samples within each of the aforementioned subregions; Calculate the deviation between the kernel density estimation distribution of each sub-region and the kernel density estimation distribution of the target region; The sub-regions with a deviation greater than a preset threshold are designated as the target sub-regions.

4. The method according to claim 1, characterized in that, The method further includes: A conditional generative adversarial network (GAN) model is constructed, which includes a generator and a discriminator. Obtain the snow depth training sample sequence for the target area; From multiple sub-regions of the target region, determine the training sub-regions in which the snow depth training samples within the sub-regions satisfy the preset sparsity condition; Using the sample distribution characteristics of the snow depth training sample sequence in the sub-regions other than the training sub-region in the target region or the statistical prior characteristics of the snow depth training sample sequence in the target region as constraints, the generator and the discriminator are trained alternately using the snow depth training sample sequence in the training sub-region until a preset termination condition is met, thereby obtaining the trained snow depth sample generation model. The loss function of the generator is determined based on adversarial loss, statistical feature matching loss, and temporal continuity physical loss. The adversarial loss characterizes the distribution difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The statistical feature matching loss characterizes the statistical feature difference between the snow depth estimation sample sequence generated by the generator and the snow depth training sample sequence in the training sub-region. The temporal continuity physical loss characterizes the difference between the temporal continuity features of the snow depth estimation sample sequence generated by the generator and the preset temporal continuity features.

5. The method according to claim 1, characterized in that, The step of generating virtual snow depth samples for the target sub-region based on a trained snow depth sample generation model, using the sample distribution characteristics of snow depth observation samples in the remaining sub-regions of the target region excluding the target sub-region or the statistical prior characteristics of snow depth observation samples in the target region as constraints, includes: Using the snow depth sample generation model, multiple sets of virtual snow depth samples are generated, each a different multiple of the number of snow depth observation samples within the target sub-region. The snow depth observation samples within the target sub-region are merged with each group of virtual snow depth samples to obtain each group of snow depth reference samples. Using each set of snow depth reference samples, the pre-constructed snow depth estimation model is trained to obtain the initial estimation model corresponding to each set of snow depth reference samples. The initial estimation model that meets the preset performance index among the multiple initial estimation models is used as the snow depth estimation model for the target sub-region.

6. The method according to claim 5, characterized in that, After the step of using the initial estimation model that meets the preset performance index from among the multiple initial estimation models as the snow depth estimation model for the target sub-region, the following is included: Obtain the snow depth estimation model for the remaining sub-regions, wherein the snow depth estimation model for the remaining sub-regions is trained based on the snow depth training samples in the remaining sub-regions; Based on the snow depth estimation model of the remaining sub-regions, the snow depth of the remaining sub-regions is estimated to obtain the snow depth estimation results of the remaining sub-regions; The snow depth of the target sub-region is estimated based on the snow depth estimation model of the target sub-region, and the snow depth estimation result of the target sub-region is obtained. The snow depth estimation results of the target sub-region and the snow depth estimation results of the other sub-regions are spliced ​​together to obtain the snow depth estimation result of the target region.

7. The method according to claim 6, characterized in that, The steps for estimating the snow depth in the sub-region to be estimated, which is either the remaining sub-regions or the target sub-region, include: The sub-region to be estimated is divided into multiple grids; The snow depth observation samples in each grid are input into the snow depth estimation model of the sub-region to be estimated, and the snow depth estimation results of each grid are obtained. The snow depth estimation results of all the grids are stitched together to obtain the snow depth estimation result of the sub-region to be estimated.

8. A snow depth sample generation device, characterized in that, The device includes: The acquisition module is used to acquire snow depth observation samples of the target area; The determination module is used to determine, from multiple sub-regions of the target region, the target sub-region where the snow depth observation sample within the sub-region satisfies a preset sparsity condition; The generation module is used to generate virtual snow depth samples of the target sub-region based on the trained snow depth sample generation model, using the sample distribution characteristics of snow depth observation samples in the remaining sub-regions of the target region other than the target sub-region or the statistical prior characteristics of snow depth observation samples in the target region as constraints.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory being used to store a program, and the processor being used to implement the snow depth sample generation method as described in any one of claims 1-7 when executing the program.

10. A computer storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the snow depth sample generation method as described in any one of claims 1-7.

Images