Vegetation coverage inversion method considering inversion uncertainty and cross-scale transfer learning interpretability

Abstract

The invention discloses a vegetation coverage inversion method considering inversion uncertainty and cross-scale transfer learning interpretability, and the method comprises the steps: executing a feature selection operation, and generating a core feature set through the fusion of a particle swarm algorithm and an SHAP interpretation framework; quantifying uncertainty caused by space matching differences, and comparing error distribution of inversion values and measured values under different overlapping degree gradients; carrying out cross-scale vegetation coverage mapping by adopting a transfer learning strategy, and adjusting source domain model parameters by utilizing target domain data; and generating a high-time-space-continuity vegetation coverage product of the target area based on the optimized model. According to the method, uncertainty control over the 85% overlapping degree threshold value is achieved, and the migration drawing precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of vegetation cover inversion technology, and in particular relates to a vegetation cover inversion method that takes into account inversion uncertainty and cross-scale transfer learning interpretability. Background Technology

[0002] Vegetation cover (FVC) is a core parameter for grassland ecosystem monitoring, and its accurate retrieval plays a crucial role in ecological assessment. Currently, FVC acquisition mainly relies on two types of technologies: ground-based measurement methods (such as manual surveys and UAV sampling), which, while ensuring high spatial resolution, are limited by labor costs and coverage area, making it difficult to support large-scale, long-term monitoring needs; and satellite remote sensing retrieval methods, including empirical regression, pixel-based bisection, and machine learning models. Among these, machine learning is widely used because it can capture complex spectral-FVC mapping relationships, and the resulting global FVC products (such as GEOV1 / GLASS) generally have a spatial resolution greater than 250 meters.

[0003] The existing technology has the following drawbacks:

[0004] (1) Spatial matching error leads to unquantified inversion uncertainty:

[0005] While high-resolution satellites (such as Landsat 30m and Sentinel-210m) can improve detail capture capabilities, the spatial discrepancy between the center of the UAV-measured image range and the center of the satellite pixels leads to systematic errors in the inversion results. This error is further amplified, especially in heterogeneous underlying surface regions. Current technology lacks a quantification mechanism for the inversion uncertainties caused by spatial matching differences, making it impossible to assess error boundaries (e.g., errors can reach 27.88% at the Landsat 30m scale), thus reducing the reliability of ecological decision-making.

[0006] (2) Lack of cross-scale transfer learning mechanisms:

[0007] To alleviate the data gaps caused by low-frequency observations from high-resolution satellites (such as Landsat's 8-day revisit cycle) and cloud contamination, multi-source remote sensing collaborative mapping has become a feasible solution. While transfer learning has been applied in fields such as agricultural classification, it still has shortcomings in FVC inversion: the effectiveness of cross-scale transfer (e.g., 10m→30m) has not been verified, and it cannot solve the adaptation problem of high-resolution data to low-resolution target domains; the feature transfer mechanism is unclear, lacking an interpretable framework to analyze the transfer patterns of spectral features during scale transformations; and it is highly dependent on data, with high costs for obtaining large-scale validation samples, hindering the production of products with high spatiotemporal continuity.

[0008] (3) Insufficient interpretability of feature selection:

[0009] Traditional feature selection methods (such as particle swarm optimization) rely on black-box optimization, making them susceptible to feature redundancy due to initialization parameters. Existing techniques cannot simultaneously achieve feature importance ranking and interpretability parsing, reducing model generalization ability (e.g., redundant features hinder the inversion of R from Landsat data). 2 (Reduced by 8%). Summary of the Invention

[0010] To address the aforementioned technical issues, this invention proposes a vegetation cover inversion method that takes into account both inversion uncertainty and the interpretability of cross-scale transfer learning, achieving uncertainty control at an 85% overlap threshold and improving the accuracy of transfer mapping.

[0011] To achieve the above objectives, this invention provides a vegetation cover inversion method that takes into account inversion uncertainty and cross-scale transfer learning interpretability, comprising:

[0012] Perform feature selection operations to generate a core feature set by fusing the particle swarm optimization algorithm with the SHAP interpretation framework;

[0013] The uncertainty caused by spatial matching differences is quantified by comparing the error distribution between the inverted and measured values ​​under different overlap gradients.

[0014] A transfer learning strategy was used to map vegetation cover across scales, and the source domain model parameters were adjusted using target domain data.

[0015] Based on the optimized model, a high spatiotemporal continuity vegetation cover product for the target area is generated.

[0016] Optional feature selection operations include:

[0017] The particle swarm optimization algorithm is used to select the initial screening feature with the average root mean square error of multi-model cross-validation as the optimization target.

[0018] The initial features are sorted by SHAP value, and the features with a cumulative contribution of not less than a set threshold are retained to form the core feature set.

[0019] Optional, quantification of spatial matching differences includes:

[0020] Set up an overlap gradient sequence and simulate spatial matching error by moving the position of the measured image;

[0021] Extract vegetation cover values ​​with different degrees of overlap along the satellite pixel boundary direction;

[0022] Compare the error range between the inverted vegetation cover under different degrees of overlap and the measured value under complete overlap.

[0023] Optionally, the overlap gradient sequence decreases from 100% to 50% at 5% intervals, and the satellite pixel boundary directions include four directions: up, down, left, and right.

[0024] Optionally, the transfer learning strategy may specifically be:

[0025] In cross-sensor migration scenarios, bidirectional migration of Landsat 30m data and Sentinel-230m data is achieved;

[0026] In cross-scale migration scenarios, upscaling migration from Sentinel-210m to 30m resolution is achieved;

[0027] In the feature-enhanced migration scenario, a vegetation index is constructed by introducing the red-edge band to participate in the migration.

[0028] Optionally, in the transfer learning strategy, the source domain model is fine-tuned and optimized using 80% of the data from the target domain.

[0029] Optional, vegetation cover products include:

[0030] The XGBoost model was used to perform the inversion for Landsat data.

[0031] Inversion was performed on Sentinel-2 data using the LightGBM model.

[0032] Inversion errors in water areas are eliminated through water body masking.

[0033] Technical Effects of this Invention: This invention discloses a vegetation cover inversion method that considers inversion uncertainty and the interpretability of cross-scale transfer learning, addressing the technical shortcomings of existing technologies such as unquantified spatial matching errors and unclear cross-scale transfer mechanisms. First, feature selection interpretability is achieved by fusing particle swarm optimization (PSO) with the SHAP interpretation framework, generating a core feature set. Second, a 5% overlap gradient quantization mechanism is established to extract vegetation cover values ​​with different overlaps in the upper / lower / left / right directions of satellite pixels, systematically revealing the inversion uncertainty caused by spatial matching differences. Finally, a fine-tuning transfer learning strategy is adopted to optimize model parameters in three transfer scenarios: cross-sensor, cross-scale, and red-edge feature enhancement, generating a high spatiotemporal continuity vegetation cover product. Specifically, transfer learning uses 80% of the target domain data to adjust the source domain model; Landsat data uses the XGBoost model, and Sentinel-2 data uses the LightGBM model for inversion, with errors eliminated through water masking. This method achieves uncertainty control at the 85% overlap threshold, and the Sentinel-2 30m transfer mapping accuracy reaches R0. 2 =0.918. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0035] Figure 1 This is a flowchart illustrating a vegetation cover inversion method that takes into account inversion uncertainty and cross-scale transfer learning interpretability according to an embodiment of the present invention. Detailed Implementation

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0038] like Figure 1 As shown, this embodiment provides a vegetation cover inversion method that takes into account inversion uncertainty and cross-scale transfer learning interpretability, including:

[0039] Perform feature selection operations to generate a core feature set by fusing the particle swarm optimization algorithm with the SHAP interpretation framework;

[0040] The uncertainty caused by spatial matching differences is quantified by comparing the error distribution between the inverted and measured values ​​under different overlap gradients.

[0041] A transfer learning strategy was used to map vegetation cover across scales, and the source domain model parameters were adjusted using target domain data.

[0042] Based on the optimized model, a high spatiotemporal continuity vegetation cover product for the target area is generated.

[0043] Furthermore, feature selection operations include:

[0044] The particle swarm optimization algorithm is used to select the initial screening feature with the average root mean square error of multi-model cross-validation as the optimization target.

[0045] The initial features are sorted by SHAP value, and the features with a cumulative contribution of not less than a set threshold are retained to form the core feature set.

[0046] Furthermore, the quantitative spatial matching differences include:

[0047] Set up an overlap gradient sequence and simulate spatial matching error by moving the position of the measured image;

[0048] Extract vegetation cover values ​​with different degrees of overlap along the satellite pixel boundary direction;

[0049] Compare the error range between the inverted vegetation cover under different degrees of overlap and the measured value under complete overlap.

[0050] Furthermore, the overlap gradient sequence decreases from 100% to 50% at 5% intervals, and the satellite pixel boundary directions include four directions: up, down, left, and right.

[0051] Furthermore, the transfer learning strategy specifically includes:

[0052] In cross-sensor migration scenarios, bidirectional migration of Landsat 30m data and Sentinel-230m data is achieved;

[0053] In cross-scale migration scenarios, upscaling migration from Sentinel-210m to 30m resolution is achieved;

[0054] In the feature-enhanced migration scenario, a vegetation index is constructed by introducing the red-edge band to participate in the migration.

[0055] Furthermore, in the transfer learning strategy, 80% of the data from the target domain is used to fine-tune the parameters of the source domain model for optimization.

[0056] Furthermore, the generated vegetation cover products include:

[0057] The XGBoost model was used to perform the inversion for Landsat data.

[0058] Inversion was performed on Sentinel-2 data using the LightGBM model.

[0059] Inversion errors in water areas are eliminated through water body masking.

[0060] Specifically, the implementation process of this embodiment includes the following steps:

[0061] Data acquisition and processing:

[0062] Unmanned Aerial Vehicle (UAV) Data Acquisition and Processing:

[0063] In this embodiment, during the lush vegetation season of July and August 2023, a DJI Phantom 4 multispectral drone (P4M) was used to collect nine 500m × 500m multispectral data points at different vegetation covers and altitudes in the Yellow River source area. The P4M integrates six 1 / 2.9-inch CMOS sensors with 2.08 million pixels, covering visible light, blue (450nm ± 16nm), green (560nm ± 16nm), red (650nm ± 16nm), red edge (730nm ± 16nm), and near-infrared (840nm ± 26nm). Furthermore, the drone not only possesses multi-frequency, multi-system high-precision RTK GNSS positioning capabilities with a vertical accuracy of 1.5cm + 1ppm (RMS) and a horizontal accuracy of 1cm + 1ppm (RMS), but also features a light intensity sensor for easier subsequent drone data processing. This embodiment selected clear, windless, and flat, open terrain conditions for data collection to ensure data quality. The drone automatically flew along a planned route based on the solar altitude angle calculated by DJI GS Pro, taking vertically downward images at an altitude of 40m. The forward overlap was 80%, and the lateral overlap was 70%. Under these conditions, the spatial resolution of the drone imagery was 2.1cm, allowing for relatively clear differentiation between vegetation and non-vegetation information. This embodiment used Pix4Dmapper software to automatically process the drone's data across different bands and output single-band orthophotos with reflectance information. Then, the NDVI thresholding method was used to extract vegetation and non-vegetation information from the multispectral data, and the proportion of vegetation pixels to total pixels was calculated as the measured FVC used in this embodiment.

[0064] Remote sensing data acquisition and processing:

[0065] This embodiment extracts reflectance data from Sentinel-2A, Sentinel-2B, Landsat 8OLI, and Landsat 9OLI 2 sensors within the study area during July-August 2023 using the GEE platform. First, this embodiment utilizes a cloud removal algorithm to eliminate the influence of clouds and converts the data using scaling factors and gain factors to obtain standard remote sensing data. Research shows that using the maximum NDVI synthesis method can further mitigate cloud pollution and atmospheric effects, thereby obtaining a high-quality remote sensing image dataset. Therefore, this embodiment employs pixel fusion to extract pixels with the maximum NDVI value from the Landsat 8OLI and Landsat 9OLI 2 data as the research basis, further improving the temporal resolution (8 days) and spatial integrity of the Landsat data (hereafter referred to as L8 / 9). This embodiment uses the GEE platform to call the JavaScript code: ee.ImageCollection(“COPERNICUS / S2_SR_HARMONIZED”) to obtain Sentinel 2A and Sentinel 2B data. Then, the pixel with the highest NDVI value from both datasets was extracted as the basis to improve the temporal resolution of the Sentinel 2 data to 5 days (hereafter referred to as S2A / B). Next, global 30m DEM data for the study area was extracted using the GEE platform. Furthermore, this embodiment uniformly uses bilinear interpolation to resample the non-10m resolution bands of S2A / B and the DEM data to 10m resolution to match the highest resolution of the S2A / B data. Finally, based on the above data sources, the feature dataset required for this embodiment was created, as shown in Table 1.

[0066] Table 1

[0067]

[0068]

[0069] This embodiment preprocesses UAV multispectral data and uses the NDVI thresholding method to extract vegetation and non-vegetation pixel information of the study area. Subsequently, Sentinel-2 and Landsat data are acquired through the GEE platform, and FVC and feature information are extracted under conditions of complete overlap between satellite remote sensing imagery and UAV measured imagery using vector grids of corresponding resolutions. Simultaneously, FVC is further extracted at different overlap levels using a 5% overlap gradient, and outliers are removed using interquartile range combined with an isolated forest algorithm to obtain the base dataset for this embodiment. Next, this dataset is randomly divided into training and test sets in an 8:2 ratio. The training set is used for feature selection, hyperparameter optimization, and machine learning model training, while the test set is used for accuracy verification. Finally, using measured FVC data with overlapping differences, the inversion uncertainty caused by the difference in pixel spatial matching between UAV measured imagery and satellite remote sensing imagery at two scales is quantified. Simultaneously, the transfer learning performance of machine learning and the interpretability of its scale transfer are explored, and its reliability in FVC mapping is verified, producing a high-precision FVC dataset for the Yellow River source region. The specific technical route is as follows: Figure 1 .

[0070] Machine learning models:

[0071] Random Forest (RFR) is an ensemble model that combines multiple independent decision trees to form a decision forest and averages the predictions of each tree to obtain the overall prediction result of the forest, exhibiting strong robustness. XGBoost (XGBR) is a boosting ensemble model that optimizes the model by minimizing the gradient information of the loss function. Its hierarchical decision tree growth technique reduces model complexity, offering high applicability and scalability. LightGBM (LGBMR) is a boosting ensemble model that uses a histogram method to determine the optimal leaf split, offering significant advantages in improving computational efficiency and optimizing memory usage. 61] CatBoost (CATBR) is a nonlinear regression boosting ensemble model that uses ordered boosting and target encoding schemes, and can effectively address the bias problem caused by data reuse.

[0072] Feature and hyperparameter selection:

[0073] Feature and hyperparameter selection is a crucial task for maximizing machine learning performance. This embodiment uses a feature selection algorithm that couples PSO and SHAP: PSO-SHAP. Particle Swarm Optimization (PSO) is a heuristic intelligent optimization algorithm that mimics the social behavior of flocks of birds or schools of fish.

[64] Each particle represents a possible solution, and the particle updates its position in the search space based on its velocity, position, and the current optimal solution. 65]Cooperation among particles can accelerate the search process, allowing the global optimum to gradually approach the true optimum. Compared to other heuristic algorithms, PSO has the advantages of ease of coding, fewer parameters, and strong scalability. 66] In this embodiment, to reduce the PSO algorithm's dependence on a single model and further enhance its robustness, the average RMSE after 3-fold cross-validation of the four tree models is used as the objective function. Furthermore, interpretability schemes such as SHAP can reveal the contribution of features to the machine learning model's predictions both globally and locally. Therefore, to enhance feature interpretability, this embodiment calculates the SHAP value of each feature in the four tree models based on the features selected by PSO, sorts them according to the magnitude of the average result, and then selects the feature with 95% interpretability as the final feature used in this embodiment.

[0074] Optuna is a tool for hyperparameter optimization based on existing hyperparameter combinations and corresponding objective functions, which constructs and updates probabilistic models. It is widely used due to its high flexibility, efficiency, and scalability. This embodiment uses the RMSE obtained after 5-fold cross-validation as the objective parameter for Optuna and conducts 1000 trials. Furthermore, this embodiment uses the Bayesian-optimized TPE (Tree-structured Parzen Estimator) sampling method and enables multivariate optimization (multivariate = True) to ensure that TPE fully considers the interactions between multiple hyperparameters. To further conserve computational resources, this embodiment also uses the HyperbandPruner pruning strategy to reduce computational overhead and accelerate the optimization process. The model and its hyperparameters used in this embodiment are shown in Table 2.

[0075] Table 2

[0076]

[0077] Quantification of uncertainty in the inversion of spatial matching differences:

[0078] In-situ measurement data is crucial for the accuracy and reliability of surface parameter inversion. However, during FVC inversion, there may be spatial matching differences between the UAV-measured image range and satellite remote sensing image pixels (UAV-satellite), leading to some uncertainty in the inversion results. To fully quantify this uncertainty, this embodiment first extracts vector grids of corresponding scales from Sentinel 2A / B and Landsat 8 / 9 remote sensing images, and obtains the FVC in each grid as the measured FVC that perfectly matches the satellite pixel. Secondly, by gradually moving the UAV image to create a certain degree of overlap with the satellite remote sensing image (this overlap will decrease by 5% in increments of 5% within the range of 100%-50%), the measured FVC under different overlap levels is extracted. To further ensure the reliability of the research results, this embodiment repeats the overlap setting and FVC extraction work in the four directions (up, down, left, and right) of the same satellite remote sensing image pixel. Finally, the errors between the inversion results under different overlap degrees and the measured FVC under self-overlap and full overlap degrees are calculated, and the mean error values ​​in four directions are used to reveal the inversion error and its uncertainty variation caused by the machine-satellite matching difference. Furthermore, this embodiment, based on the Monte Carlo randomized trial concept, randomly selects 100 numbers between 0 and 1000 as random seeds for data segmentation and machine learning training, and conducts 100 randomized trials under different overlap degrees to further study the inherent uncertainty of the machine learning model inversion and its overlap sensitivity.

[0079] Transfer learning:

[0080] High-precision measured FVC datasets are the cornerstone of remote sensing inversion model construction. Although UAV sampling has played a significant role, it still suffers from the inherent drawback of exponentially increasing costs in acquiring multi-scale validation data. Furthermore, the numerous parameter selections required for machine learning models further increase computational costs. This embodiment systematically designs a multi-dimensional transfer learning framework to explore the cross-sensor and cross-scale transfer performance of machine learning. This embodiment uses five types of datasets: (1) L8 / 9 30m resolution data (L8 / 9_30m); (2) S2A / B 10m resolution data (S2A / B_10m) and its upscaled 30m resolution data (S2A / B_30m); (3) S2A / B 10m red-edge enhanced dataset (S2A / B_10m_RE) and its upscaled 30m resolution data (S2A / B_30m_RE). Among them, (1) and (2) share the spectral feature space, while (3) introduces differentiated vegetation response information through the red-edge band. Based on the aforementioned data architecture, five progressive transfer scenarios are constructed: Scenarios 1 and 3 conduct bidirectional cross-sensor transfer between L8 / 9_30m and S2A / B_30m. Scenarios 2 and 4 realize cross-scale transfer from S2A / B_10m to L8 / 9_30m data and from S2A / B_30m. Scenario 5 explores cross-scale transfer from S2A / B_10m_RE to S2A / B_30m_RE. This embodiment employs a fine-tuning transfer learning strategy, utilizing 80% of the target domain data to fine-tune parameters and 20% to validate the performance of fine-tuning transfer learning, rapidly adapting to the target domain feature distribution while preserving prior knowledge from the source domain. Furthermore, to analyze the mechanistic characteristics of multi-scale transfer, this embodiment innovatively introduces the SHAP interpretability framework to track the dynamic evolution of feature importance before and after transfer. By comparing the distribution differences of SHAP values ​​between the source and target domains, this embodiment reveals the transfer characteristics of spectral features at different spatial scales, providing a theoretical basis for optimizing multi-source remote sensing transfer schemes.

[0081] High-precision FVC mapping and extrapolation verification:

[0082] This embodiment utilizes L8 / 9 and S2A / B data, along with their perfectly matched measured FVC data, to invert high-precision FVC datasets at 10m and 30m resolutions for the study area based on four machine learning models and the optimal transfer learning scheme. This includes the S2A / B 10m resolution dataset, the S2A / B 30m dataset obtained through the optimal transfer learning scheme, and the L8 / 9 30m resolution dataset. Furthermore, this embodiment directly uses the S2A / B_10m source domain model to invert the S2A / B 30m resolution dataset from the S2A / B_30m data to compare the differences in results from various 30m FVC mapping schemes. Simultaneously, this embodiment uses the MNDWI index to extract water system data for the study area (threshold set to -0.2), and performs masking processing on the FVC mapping data based on this data to further ensure the quality of the FVC mapping data. Then, 40% of the independent data within the validation sites are randomly selected for extrapolation validation of the mapping data. The validation dataset consisted of 3294 data points at a 10m scale and 381 data points at a 30m scale. Finally, based on the optimal mapping scheme and validation results, a set of high-precision FVC data for the Yellow River source region at 10m and 30m resolutions was produced respectively.

[0083] Accuracy assessment:

[0084] To accurately evaluate the inversion performance of all models, this embodiment uses three metrics to evaluate model performance, including the coefficient of determination (R²). 2 ), root mean square error (RMSE), and mean absolute error (MAE). Where R... 2 The larger the value, the smaller the RMSE and MAE, indicating better model performance and higher inversion accuracy.

[0085] This embodiment, based on multi-source remote sensing image data and combined with a large amount of UAV-measured FVC data matching the spatial scale of satellite remote sensing images, studies the interpretable feature selection for FVC inversion, the uncertainty quantification of the difference between the UAV-measured image range and the pixel spatial matching of satellite remote sensing images, and the analysis of cross-scale transfer learning mechanisms. High-precision FVC data at 10m and 30m scales for the Yellow River source region were also produced. The main conclusions are as follows:

[0086] (1) The PSO-SHAP feature selection algorithm can effectively balance feature importance and interpretability in FVC inversion at different scales. 2 Maximum increase of 8%.

[0087] (2) The spatial matching difference between the measured range of UAV images and the pixels of satellite remote sensing images leads to uncertainties in FVC inversion and between machine learning models. These uncertainties gradually decrease with increasing overlap and tend to stabilize after 85% overlap. Uncertainties exhibit a certain scale correlation; the larger the scale, the greater the uncertainty. Specifically, R at the L8 / 9 scale...2 The highest uncertainties for RMSE and MAE are 7.08%, 27.88%, and 22.18%, respectively, and 2.99%, 7.74%, and 5.95% at the S2A / B scale.

[0088] (3) Machine learning has certain transfer performance between different scales of the same source and between different sources of the same scale. In particular, models trained on high-resolution data containing red-edge exponential features have the best inversion accuracy in cross-scale fine-tuning transfer learning of low-resolution data of the same source. SHAP can analyze the cross-scale fine-tuning transfer learning process from the perspective of feature transfer and reveal the scale invariance advantage of dominant features.

[0089] (4) The S2A / B data exhibits superior cartographic accuracy and spatial detail compared to the L8 / 9 data at a 30m resolution. The LGBMR model demonstrates the best cartographic performance on 10m data: R 2 =0.87, RMSE=0.110, MAE=0.075, the XGBR model in cross-scale fine-tuning transfer learning has the best mapping performance on 30m data: R 2 =0.90, RMSE=0.097, MAE=0.069. This embodiment promotes the transparency and practicality of machine learning in vegetation remote sensing, providing methodological innovation and high-precision data support for regional ecological management.

[0090] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vegetation cover inversion method that takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, characterized in that, include: Perform feature selection operations to generate a core feature set by fusing the particle swarm optimization algorithm with the SHAP interpretation framework; The uncertainty caused by spatial matching differences is quantified by comparing the error distribution between the inverted and measured values ​​under different overlap gradients. A transfer learning strategy was used to map vegetation cover across scales, and the source domain model parameters were adjusted using target domain data. Based on the optimized model, a high spatiotemporal continuity vegetation cover product for the target area is generated.

2. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, is characterized in that... Feature selection operations include: The particle swarm optimization algorithm is used to select the initial screening feature with the average root mean square error of multi-model cross-validation as the optimization target. The initial features are sorted by SHAP value, and the features with a cumulative contribution of not less than a set threshold are retained to form the core feature set.

3. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, is characterized in that... Differences in quantified spatial matching include: Set up an overlap gradient sequence and simulate spatial matching error by moving the position of the measured image; Extract vegetation cover values ​​with different degrees of overlap along the satellite pixel boundary direction; Compare the error range between the inverted vegetation cover under different degrees of overlap and the measured value under complete overlap.

4. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and cross-scale transfer learning interpretability, is characterized in that... The overlap gradient sequence decreases from 100% to 50% at 5% intervals, and the satellite pixel boundary directions include four directions: top, bottom, left, and right.

5. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, is characterized in that... The specific transfer learning strategy is as follows: In cross-sensor migration scenarios, bidirectional migration of Landsat 30m data and Sentinel-230m data is achieved; In cross-scale migration scenarios, upscaling migration from Sentinel-210m to 30m resolution is achieved; In the feature-enhanced migration scenario, a vegetation index is constructed by introducing the red-edge band to participate in the migration.

6. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, is characterized in that... In the transfer learning strategy, 80% of the data from the target domain is used to fine-tune the parameters of the source domain model for optimization.

7. The vegetation cover inversion method as described in claim 1, which takes into account inversion uncertainty and the interpretability of cross-scale transfer learning, is characterized in that... The vegetation cover products generated include: The XGBoost model was used to perform the inversion for Landsat data. Inversion was performed on Sentinel-2 data using the LightGBM model. Inversion errors in water areas are eliminated through water body masking.

Images