A Surface Parametric Mapping Method Based on Adaptive Fusion of Physical Constraints and Spatial Residues
By adopting a surface parametric mapping method based on adaptive fusion of physical constraints and spatial residuals, the problems of scale artifact magnification, lack of physicochemical mechanisms, and solidification of fusion weights in existing technologies are solved, and high-precision and robust surface parametric mapping is achieved.
Patent Information
- Application Number
- CN202611620034.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for large-scale, high-resolution surface parameter spatial inversion and digital mapping suffer from problems such as scale artifact amplification, missing physicochemical mechanisms, and solidified fusion weights, resulting in inaccurate mapping results and poor robustness.
A surface parameter mapping method based on physical constraints and spatial residual adaptive fusion is adopted. By acquiring multi-scale, multi-source heterogeneous spatial data, standardization and collaborative preprocessing are performed to explicitly extract high-frequency spatial residual labels. High-frequency residual learning and correction are carried out in combination with hydrophysical constraints, and the fusion ratio is dynamically adjusted to generate cross-scale fused prediction maps.
It achieves high-precision, robust cross-regional surface parameter mapping, effectively suppresses error amplification in complex terrain areas, improves mapping accuracy and physical boundary fidelity, and enhances cross-regional generalization capability.
Smart Images

Figure CN123118949A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of computer technology, and more particularly to the field of remote sensing information processing and geospatial information technology, specifically to a surface parameter mapping method based on adaptive fusion of physical constraints and spatial residuals. Background Technology
[0002] Currently, for spatial inversion and digital mapping of large-scale, high-resolution surface parameters (such as soil pH and organic carbon), the industry mainly adopts three types of technical solutions, and these solutions differ in terms of feature utilization and fusion strategies.
[0003] Option 1: Spatial interpolation and simple downscaling method.
[0004] These methods mainly rely on image resampling techniques (such as bilinear interpolation and cubic convolution interpolation) to resample existing coarse-resolution surface parameter products (such as spatial resolution from 250m to 1km) to the target high resolution, or use geostatistics (such as ordinary kriging and inverse distance weighting) to directly perform full-area spatial interpolation on measured sample points. The core logic is to change the pixel size or fill the spatial gaps through purely mathematical means.
[0005] Option 2: Machine learning method of directly stacking multiple source covariates.
[0006] This is currently the most widely used technical approach for large-scale mapping. Its basic principle is to directly treat the coarse-resolution prior map of surface parameters as a regular environmental covariate, and directly stack it with high-resolution remote sensing multispectral features and terrain attributes. The data are then fed into machine learning models such as Random Forest (RF) and Gradient Boosting Tree (GBDT) for training. The aim is to automatically discover the nonlinear mapping relationship between multidimensional covariates and surface parameters through a data-driven approach.
[0007] Option 3: Conventional residual correction and multi-scale fusion method.
[0008] To mitigate the systematic biases that can arise from the direct stacking method, some studies have proposed a conventional residual fusion framework. This method treats the coarse-resolution product as a macroscopic low-frequency background trend. First, it calculates the spatial residual between the measured sample points and this background value. Then, it fits the residual using only high-resolution environmental covariates (extracting high-frequency information). Finally, it directly adds the predicted high-resolution residual results to the coarse-resolution prior map, thereby achieving a simple linear combination of high-frequency residuals and low-frequency trends.
[0009] However, considering the practical engineering needs of large-scale, high-resolution mapping, the above-mentioned technologies have three major drawbacks in terms of local heterogeneity representation, physical constraint mechanisms, and fusion weight allocation.
[0010] (1) The lack of physical boundary constraints in the downscaling of coarse resolution products leads to scale mismatch and artifact amplification problems associated with multi-scale feature fusion.
[0011] Current fusion methods (including direct stacking and conventional residual fusion) rely heavily on conventional resampling techniques such as bilinear interpolation for spatial downscaling when introducing coarse-resolution prior products (e.g., 250m resolution mass-source soil products) to provide macroscopic background information. This purely mathematical approach fails to incorporate a realistic three-dimensional topographic framework (e.g., a spatial physical boundary framework constructed from a digital elevation model (DEM)) for rigid constraints. The resulting high-resolution mapping generally suffers from a "smoothing effect," causing the "blocky features" of coarse-resolution pixels to be incorrectly smoothed, failing to accurately represent local spatial heterogeneity. Furthermore, machine learning models are prone to learning these spurious boundaries as real environmental covariate features during fitting, ultimately producing severe "mosaic artifacts" in the mapping results. This disrupts the spatial topological continuity of natural surface physicochemical features and severely restricts the spatial detail representation capabilities of high-resolution remote sensing data.
[0012] (2) Pure data-driven models lack geoscientific mechanism constraints and are prone to anti-physics prediction problems.
[0013] Current mainstream high-resolution spatial prediction models for land surface parameters heavily rely on purely data-driven mechanisms. These models essentially focus on optimizing global statistical errors, with their optimization objective solely concerned with the mathematical convergence of these errors. They neglect the fact that the spatial evolution of land surface parameters is highly controlled by micro-topography, water transport (such as leaching and deposition), and the intrinsic geophysicochemical laws governing material deposition. This lack of physical understanding, a "black box" learning model, makes the models prone to producing anomalous predictions that contradict geological common sense in areas with sparse sample coverage or extreme topography (such as steep slopes and deep depressions with strong leaching). This results in poor robustness when extrapolating across regions, a lack of interpretability of the physicochemical mechanisms, and severely weakened cross-regional robustness and spatial generalization performance.
[0014] (3) Ignoring the problem of solidified weights in multi-scale fusion caused by spatial non-stationarity.
[0015] In the multi-scale fusion stage of separating low-frequency macro trends from high-frequency prediction residuals, such as when fusing coarse-resolution prior maps with high-resolution prediction results, current conventional residual correction methods generally adopt a globally unified linear superposition formula (i.e., the fusion weights of low-frequency trends and high-frequency residuals are mechanically fixed at 1:1). However, in different geographic regions, the accuracy of coarse-resolution low-frequency prior information is higher in areas with gentle terrain, while the error is extremely large in mountainous areas with rugged terrain and fragmented surfaces. The reliability of its high-frequency residual information exhibits significant spatial non-stationarity differences. This processing logic completely ignores the "spatial non-stationarity" characteristic that is prevalent in geography. It lacks a dynamic perception and evaluation mechanism based on spatial complexity, and the globally fixed weight allocation mechanism cannot perceive and isolate this spatial heterogeneity. As a result, the inherent errors of coarse-resolution products are directly transmitted and amplified in complex terrain areas. For example, the prediction errors inherent in low-resolution base maps are further transmitted and amplified in areas with fragmented and complex terrain, seriously weakening the effectiveness of high-resolution residual term correction.
[0016] In summary, current technologies have not yet provided an effective overall solution to the systemic defects in the fusion of multi-scale features of surface parameters, such as scale artifact amplification, lack of physicochemical mechanisms, and solidification of fusion weights.
[0017] Furthermore, in the field of Digital Soil Mapping (DSM), existing research has employed tree-based machine learning models coupled with residual kriging interpolation for spatial mapping. However, residual processing often relies on purely mathematical statistical interpolation, and the synthesis stage assumes that low-frequency trends and high-frequency residuals are globally fixed linear superpositions. Regarding physically constrained machine learning, cutting-edge research on physically heuristic machine learning focuses on constructing partial differential equations or embedding complex soil-hydrological mechanism models, requiring stringent underlying geological parameters, which is almost infeasible for large-scale operational calculations. In terms of spatial uncertainty quantification, existing research has output uncertainty maps based on deep learning, but their uncertainty relies solely on the model's own variance propagation and is not linked to the confidence level of the fusion process. Summary of the Invention
[0018] This disclosure provides a method, apparatus, device, and storage medium for surface parameter mapping based on adaptive fusion of physical constraints and spatial residuals.
[0019] According to a first aspect of this disclosure, a surface parametric mapping method based on adaptive fusion of physical constraints and spatial residuals is provided. The method includes: Multi-scale, multi-source heterogeneous spatial data of the target inversion area are acquired and standardized collaborative preprocessing is performed. The multi-scale, multi-source heterogeneous spatial data includes a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. The preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map base are reconstructed to generate a prior low-frequency map. Based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset, high-frequency spatial residual labels are explicitly extracted; By learning and correcting high-frequency residuals coupled with hydrophysical constraints, a high-frequency residual correction map is generated based on the high-frequency spatial residual labels and the preprocessed high-resolution temporal multispectral image set. Based on the dynamic credibility assessment driven by spatial nonstationarity, the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map is dynamically adjusted to generate a cross-scale fusion prediction map.
[0020] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein acquiring multi-scale, multi-source heterogeneous spatial data of the target inversion region and performing standardized collaborative preprocessing includes: Acquire multi-scale, multi-source heterogeneous spatial data of the target inversion area, including a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. Based on the time-series bare soil quality score, the high-resolution time-series multispectral image set is reconstructed to generate a high-resolution bare soil spectral base map; the scoring mechanism constraints of the time-series bare soil quality score include: the score is negatively correlated with the normalized vegetation index of the image, and positively correlated with the normalized burn ratio and the normalized differential bare soil index. Using the high-resolution bare soil spectral base map as the target spatial reference, spatial reference unification and environmental variable extraction are performed on the high-resolution digital elevation model, the coarse-resolution prior low-frequency map base, and the ground measured sample point surface parameter dataset, and a high-resolution environmental covariate atlas is generated.
[0021] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein reconstructing the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map basis to generate the prior low-frequency map includes: Using spatial micro-topography analysis algorithms, based on a preprocessed high-resolution digital elevation model, the slope and topographic relief of the target inversion area are extracted pixel by pixel. The slope and terrain undulation are subjected to extreme value normalization and then linearly fused with equal weights into a physical boundary skeleton matrix. Preliminary resampling is performed on the preprocessed coarse-resolution prior low-frequency map substrate; Based on the coarse-resolution prior low-frequency map base after preliminary resampling and the physical boundary skeleton matrix, terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints is performed to generate a prior low-frequency map.
[0022] As described above and in any possible implementation, a further implementation is provided, wherein the explicit extraction of high-frequency spatial residual labels based on the prior low-frequency map and the preprocessed ground-measured sample point surface parameter dataset includes: Multi-source data point topology matching and value extraction are performed on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset to generate prior low-frequency predicted values and actual observed values corresponding to the spatial locations of each ground measured sample point. Based on the prior low-frequency predicted values and true observed values, explicit spatial residual calculations are performed on all spatially matched sample points one by one to generate the corresponding true spatial residuals of the land surface parameters. Based on the true spatial residuals of the surface parameters and the high-resolution environmental covariate atlas, high-frequency spatial residual labels are constructed.
[0023] As described above and in any possible implementation, a further implementation is provided, wherein the generation of a high-frequency residual correction map based on the high-frequency spatial residual label and the preprocessed high-resolution temporal multispectral image set through high-frequency residual learning and correction coupled with hydrophysical constraints includes: Using the high-frequency spatial residual label as the target dependent variable and the high-resolution environmental covariate map as the input independent variable matrix, a basic nonlinear fitting framework is constructed based on a tree-based ensemble algorithm with a preset loss function. Construct a piecewise, continuously differentiable physical penalty loss function for extreme terrain conditions; Construct a total loss function based on the preset Huber loss function and the physical penalty loss function; The model is trained based on the aforementioned basic nonlinear fitting framework until the total loss function converges, and the physical constraint residual fitting model is output. The high-resolution environmental covariates of all effective pixels within the target inversion area are input into the physical constraint residual fitting model, and a high-frequency residual correction map is output.
[0024] In addition to the aspects and any possible implementations described above, an implementation is further provided in which the tree-based ensemble algorithm for the preset loss function includes at least one of extreme gradient boosting trees, tree-based models that support custom joint gradient optimization, or physically inspired neural network frameworks.
[0025] As described above and in any possible implementation, a further implementation is provided, wherein the dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined linearly with equal weights to generate a comprehensive terrain complexity index. An adaptive dynamic credibility weight matrix is constructed based on the decay function of the comprehensive terrain complexity index. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0026] As described above and in any possible implementation, a further implementation is provided, wherein the dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined linearly with equal weights to generate a comprehensive terrain complexity index. Based on the S-shaped nonlinear mapping function, and according to the decay function of the comprehensive terrain complexity index, an adaptive dynamic credibility weight matrix is constructed. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0027] As described above and in any possible implementation, a further implementation is provided, wherein the dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, local verification error and local landform features are calculated pixel by pixel. Using geographic weighted regression or spatial variable coefficient models, with local validation error as the dependent variable and local topographic features as the independent variable, an adaptive dynamic confidence weight matrix is implicitly solved and output through local matrix operations. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0028] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the method further includes: The reliability of the cross-scale fusion prediction map is evaluated, and the uncertainty at the spatial pixel level is quantified using an ensemble learning mechanism to generate a dual-track spatial digital layer with both deterministic and uncertain aspects.
[0029] According to a second aspect of this disclosure, a surface parametric mapping device based on adaptive fusion of physical constraints and spatial residuals is provided. The device includes: The acquisition module is used to acquire multi-scale, multi-source heterogeneous spatial data of the target inversion area and perform standardized collaborative preprocessing; the multi-scale, multi-source heterogeneous spatial data includes a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. The reconstruction module is used to reconstruct the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map base to generate the prior low-frequency map. The explicit extraction module is used to explicitly extract high-frequency spatial residual labels based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset. The residual correction module is used to learn and correct high-frequency residuals by coupling hydrophysical constraints, and generate a high-frequency residual correction map based on the high-frequency spatial residual labels and the preprocessed high-resolution temporal multispectral image set. The fusion module is used for dynamic credibility assessment driven by spatial nonstationarity, dynamically adjusting the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map.
[0030] According to a third aspect of this disclosure, an electronic device is provided. The electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.
[0031] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described above.
[0032] This application provides a surface parameter mapping method based on adaptive fusion of physical constraints and spatial residuals. This method acquires multi-scale, multi-source heterogeneous spatial data of the target inversion area, including a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground-measured sample point surface parameter dataset, and a high-resolution temporal multispectral image set, and performs standardized collaborative preprocessing. The preprocessed high-resolution digital elevation model and coarse-resolution prior low-frequency map base are then reconstructed to generate a prior low-frequency map, thereby achieving terrain-guided low-frequency reconstruction. This high-fidelity reconstruction of the physical boundaries of the low-frequency prior map significantly alleviates scale reduction from the underlying data structure. Mosaic artifacts during the process; based on prior low-frequency maps and preprocessed ground-measured sample point surface parameter datasets, high-frequency spatial residual labels are explicitly extracted; by coupling high-frequency residual learning and correction with hydrophysical constraints, a high-frequency residual correction map is generated based on the high-frequency spatial residual labels and preprocessed high-resolution temporal multispectral image set. This deeply couples the physical constraint penalty term based on hydrogeological and topographic features within the machine learning framework of spatial residual fitting, allowing it to directly participate in the underlying optimization of model training, achieving high-frequency residual learning with physical constraints, and breaking the black-box limitation of purely data-driven models; dynamic credibility assessment based on spatial nonstationarity is also implemented. The estimation dynamically adjusts the cross-scale fusion reconstruction ratio of prior low-frequency maps and high-frequency residual correction maps to generate cross-scale fusion prediction maps, thereby perceiving the spatial non-stationarity and reliability differences of different geomorphic regions in real time. This achieves dynamic weighted fusion of prior low-frequency reconstruction information and high-frequency prediction residuals at the pixel scale, effectively suppressing the error amplification effect in complex terrain areas. Based on this, a large-scale, high-resolution spatial inversion and digital mapping method for land surface parameters based on physical constraints and spatial residual adaptive fusion can be constructed. This method abandons the simple, mechanical splicing of various algorithm modules in current technologies, and instead constructs a "structural downscaling" model based on the law of geospatial non-stationarity. The complete spatial information collaborative processing closed loop of "weak physical constraints and strong spatial adaptation", namely the "credibility reconstruction mechanism driven by spatial non-stationarity", effectively breaks through the theoretical limitations of existing multi-scale feature fusion and empirical inversion models. It achieves synergistic improvement in mapping accuracy, physical boundary fidelity and cross-regional generalization ability. Furthermore, by combining coarse-resolution prior low-frequency map base, high-precision topographic features and high-resolution temporal remote sensing images, it achieves high-fidelity fusion of cross-scale spatial data through the above-mentioned three-stage collaborative mechanism of "physical boundary-guided low-frequency reconstruction", "hydrological physical constraint high-frequency residual learning" and "spatial non-stationarity adaptive weighting".
[0033] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description
[0034] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart is shown for a surface parametric mapping method based on adaptive fusion of physical constraints and spatial residuals according to an embodiment of the present disclosure; Figure 2 A flowchart is shown for another surface parametric mapping method based on adaptive fusion of physical constraints and spatial residuals according to an embodiment of the present disclosure; Figure 3 A block diagram of a surface parametric mapping apparatus based on adaptive fusion of physical constraints and spatial residuals according to an embodiment of the present disclosure is shown. Figure 4 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation
[0035] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0036] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0037] This disclosure presents a spatial inversion and digital mapping method for large-scale, high-resolution surface parameters based on adaptive fusion of physical constraints and spatial residuals. It abandons the simple, mechanical splicing of algorithm modules in current technologies, instead constructing a complete spatial information collaborative processing closed loop based on the law of geospatial nonstationarity: "structural downscaling - weak physical constraints - strong spatial adaptation," namely, a "reliability reconstruction mechanism driven by spatial nonstationarity." This effectively overcomes the theoretical limitations of existing multi-scale feature fusion and empirical inversion models, achieving synergistic improvements in mapping accuracy, physical boundary fidelity, and cross-regional generalization ability. Furthermore, by combining a coarse-resolution prior low-frequency map base, high-precision topographic features, and high-resolution temporal remote sensing imagery, it achieves high-fidelity fusion of cross-scale spatial data through the aforementioned three-stage collaborative mechanism of "physical boundary-guided low-frequency reconstruction," "hydrological and physical constraint high-frequency residual learning," and "spatial nonstationarity adaptive weighting."
[0038] Figure 1 A flowchart of a surface parameter mapping method 100 based on adaptive fusion of physical constraints and spatial residuals according to an embodiment of the present disclosure is shown.
[0039] In box 110, multi-scale, multi-source heterogeneous spatial data of the target inversion region are acquired and standardized collaborative preprocessing is performed.
[0040] In some embodiments, the target inversion region can be specifically set according to the user's actual needs.
[0041] In some embodiments, multi-scale, multi-source heterogeneous spatial data can be multi-scale spatial environment datasets obtained by users according to their actual needs for inversion, so that high-resolution bare soil spectral atlases without vegetation disturbance can be reconstructed through time-series quality scoring strategies, and standardized collaborative preprocessing can be completed, providing a standardized data foundation for subsequent cross-scale fusion and high-precision inversion.
[0042] In some embodiments, the acquisition of multi-scale, multi-source heterogeneous spatial data of the target inversion region and the subsequent standardized collaborative preprocessing specifically include: Acquire multi-scale, multi-source heterogeneous spatial data of the target inversion area. Multi-scale, multi-source heterogeneous spatial data includes, but is not limited to, high-resolution digital elevation models, coarse-resolution prior low-frequency map bases, ground measured sample point surface parameter datasets, and high-resolution time-series multispectral image sets. Based on the time-series bare soil quality score, a high-resolution time-series multispectral image set is reconstructed to generate a high-resolution bare soil spectral base map. The scoring mechanism constraints of the time-series bare soil quality score include: the score is negatively correlated with the normalized vegetation index of the image, and positively correlated with the normalized burn ratio and the normalized differential bare soil index. Using a high-resolution bare soil spectral base map as the target spatial reference, spatial reference unification and environmental variable extraction are performed on the high-resolution digital elevation model, the coarse-resolution prior low-frequency map base, and the surface parameter dataset of ground measured sample points, and a high-resolution environmental covariate atlas is generated.
[0043] In some embodiments, for the target inversion region, the following four types of multi-scale, multi-source heterogeneous spatial data can be systematically acquired to complete the acquisition of multi-source heterogeneous basic datasets: 1) Ground measured sample point surface parameter dataset: Obtain the in-situ measured coordinates (latitude and longitude) of the target surface parameters (including but not limited to soil pH, soil organic carbon, total salinity, heavy metal concentration and other physicochemical properties strongly affected by topography and hydrology) within the target inversion area, taking into account both spatial distribution uniformity and topographic characteristics, and their corresponding actual observed values of physicochemical analysis. 2) Coarse-resolution prior low-frequency map substrate: Obtain open-source or crowdsourced low-resolution surface parameter digital raster products covering the target inversion area, and use them as prior low-frequency data input to provide macro trend information. 3) High-resolution temporal multispectral image set: acquire long-term high-resolution multispectral satellite images covering the target inversion area (including but not limited to Landsat series, Sentinel-2 series, Gaofen series GF satellites, etc.), with spatial resolution defined as the preset target uniform resolution level (such as 10m, 30m or customized grid). 4) High-resolution digital elevation model (DEM): Acquire high-precision topographic data that matches the spatial resolution of the multispectral images mentioned above, which will be used for subsequent extraction of spatial physical boundary skeleton and calculation of hydrophysical features.
[0044] In some embodiments, to eliminate the masking interference of the spatiotemporal heterogeneity of vegetation cover in farmland and natural areas on the surface parametric spectral response, precise bare soil period optimization and mosaic fusion can be performed on high-resolution temporal multispectral image sets.
[0045] First, radiometric calibration, atmospheric correction, and high-precision cloud and cloud shadow masking are performed on all input high-resolution temporal multispectral images to obtain pure surface reflectance sequence data.
[0046] Secondly, a pixel-by-pixel comprehensive scoring model for bare soil quality is constructed. This is achieved by jointly evaluating the image using the Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR2), and Normalized Difference Bare Soil Index (NDBSI). The scoring mechanism is constrained by a negative correlation between the score and NDVI, and a positive correlation between the score and both NBR2 and NDBSI, thus quantifying and identifying the "bare soil purity" of each pixel at different observation phases.
[0047] Finally, employing a "best bare soil pixel priority" synthesis strategy, within a preset optimal observation window, effective observations with the highest bare soil quality scores (i.e., lowest vegetation cover and highest soil spectral signal-to-noise ratio) are extracted pixel by pixel for spatial stitching, outputting a high-resolution, seamless bare soil spectral base map free from cloud cover and vegetation interference. In summary, the reconstruction of a high-resolution bare soil spectral atlas based on time-series quality scores is complete.
[0048] In some embodiments, the synthesized high-resolution bare soil spectral base map can be used as the target spatial reference, and the coarse-resolution prior low-frequency map base, high-resolution DEM, and ground measured sample point surface parameter dataset can be uniformly projected and transformed to the same coordinate system. Subsequently, based on the high-resolution bare soil spectral base map, multidimensional spectral derivative indices sensitive to target surface parameters (such as green / near-infrared ratio, blue-green normalized difference, etc.) are calculated and extracted; based on the high-precision DEM, micro-topographic and hydrological attribute features such as elevation, slope, terrain ruggedness index (TRI), and topographic wetness index (TWI) are extracted, thus completing the spatial reference unification and environmental variable extraction.
[0049] The aforementioned spectral and topographic features together constitute a high-resolution environmental covariate atlas, providing standardized full-volume high-frequency feature inputs for subsequent construction of topographic boundary skeletons, computation of physical constraint penalty terms, and quantification of local topographic complexity.
[0050] In box 120, the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map basis are reconstructed to generate the prior low-frequency map.
[0051] In some embodiments, to overcome the smoothing effect and mosaic artifacts caused by current downscaling resampling (such as bilinear interpolation), a high-resolution terrain skeleton is introduced as a spatial physical prior. The coarse-resolution product is reconstructed with high fidelity of physical boundaries, and prior low-frequency information with clear geoscientific logical boundaries is extracted to complete terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints.
[0052] In some embodiments, the above-described reconstruction of the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map basis to generate a prior low-frequency map specifically includes: Using spatial micro-topography analysis algorithms, based on a preprocessed high-resolution digital elevation model, the slope and topographic relief of the target inversion area are extracted pixel by pixel. The slope and topographic relief are normalized to their extreme values and then linearly fused with equal weights into a physical boundary skeleton matrix. Preliminary resampling is performed on the preprocessed coarse-resolution prior low-frequency map substrate; Based on the coarse-resolution prior low-frequency map base and the physical boundary skeleton matrix after preliminary resampling, terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints is performed to generate a prior low-frequency map.
[0053] In some embodiments, the spatial distribution variation of surface physicochemical properties, such as soil pH and organic matter, is closely related to the transport of water and matter, and this process is highly controlled by the undulating boundaries of natural terrain. Spatial abrupt changes in slope and undulation, such as the edges of eroded valleys and transition zones between rolling hills and terraces, usually correspond strictly to the actual natural geographical boundaries where soil physicochemical properties change abruptly.
[0054] Therefore, based on the high-precision digital elevation model (DEM) obtained above, spatial micro-topography analysis algorithms can be used to extract topographic attribute features such as slope and terrain relief (TRI) for each pixel across the entire area. Subsequently, the extracted topographic attribute features are subjected to extreme value normalization and then linearly fused with equal weights to form a feature matrix reflecting the high-frequency spatial texture attributes of the land surface, which is defined as the guidance image (I). This guidance image constructs the "spatial physical boundary skeleton" of the target area, with its spatial resolution strictly consistent with the target's high resolution, thus completing the construction of the geoscientific mechanism-based spatial physical boundary skeleton and guidance image.
[0055] In some embodiments, the above-mentioned terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints, performed on the coarse-resolution prior low-frequency map basis after preliminary resampling and the physical boundary skeleton matrix to generate a prior low-frequency map, specifically includes: Using the coarse-resolution prior low-frequency map base after initial resampling as the input image and the physical boundary skeleton matrix as the guiding image, terrain-guided filtering processing that preserves spatial edges is performed to generate the prior low-frequency map.
[0056] Based on this, by introducing a spatial physical boundary skeleton built on a high-resolution digital elevation model (DEM) and combining it with a terrain edge-guided filtering mechanism, high-fidelity reconstruction of the physical boundaries of low-frequency prior maps can be achieved, significantly alleviating mosaic artifacts in the downscaling process from the underlying data structure.
[0057] Specifically, the coarse-resolution prior image basis obtained above is initially resampled to the target high resolution and used as the input image (denoted as...). Subsequently, the high-resolution physical boundary skeleton matrix constructed above... As a guide image, a guided filtering operation that preserves spatial edges is performed.
[0058] The core mathematical assumption of guided filtering is: in pixels local window centered Internally, output image (denoted as) ) and guide image There exists a local linear mapping relationship: In the formula, The reconstructed output cell values, To guide image metadata, and For local windows The linear coefficients within the input coarse-resolution prior image are determined. To solve for these linear coefficients and prevent overfitting, a least-squares cost function is constructed, which minimizes the difference between the output image and the input coarse-resolution prior image. While reconstructing the error, a regularization term is introduced to penalize the coefficient. Disorder mutations: In the formula, The smoothing regularization parameter controls the strength of boundary preservation, and its recommended value range is set to [value range missing]. Local window The recommended analysis radius is set to Each pixel. Through the above mathematical process, the algorithm forces the guiding image to... The high-frequency topographic abrupt change gradients included, namely at least one micro-topographic and hydrological attribute feature among elevation, slope, topographic relief (TRI), and topographic wetness index (TWI), are directionally transferred to the coarse-resolution input image. In this process, the physical boundaries are precisely reshaped, and the terrain-guided filtering process that preserves the spatial physical boundaries is completed.
[0059] In some embodiments, after the terrain-guided filtering process described above, the low-frequency macroscopic trend attributes of coarse-resolution pixels (such as large-scale climate and soil-forming background baselines) are deeply coupled with the high-resolution terrain boundary skeleton. The final output is a priori low-frequency map consistent with the target resolution. This map, while maintaining the distribution regularity of macroscopic parameters without distortion, perfectly matches the real topological boundaries at the local pixel scale, eliminating block artifacts caused by spatial scale mismatch from the underlying data structure. This provides high-fidelity physical boundary-based base data for subsequent high-frequency residual learning and spatial non-stationary fusion, thus completing the output of a priori low-frequency map with high-fidelity boundaries.
[0060] In some embodiments, the core logic of the specific mathematical calculation model or machine learning algorithm in "terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints" is to use a high-resolution terrain skeleton as a structural prior to perform edge-preserving smoothing and reconstruction on the coarse-resolution product. In addition to the guided image filtering, the following equivalent alternatives can also be used, such as joint bilateral filtering or optimized reconstruction based on local spatial Markov random field (MRF).
[0061] In some embodiments, the above-mentioned terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints, performed according to the coarse-resolution prior low-frequency map basis after preliminary resampling and the physical boundary skeleton matrix to generate a prior low-frequency map, further includes: Using the coarse-resolution prior low-frequency map base after initial resampling as the input image, and the physical boundary skeleton matrix or the high-resolution comprehensive terrain feature matrix as the joint guiding map, the similarity weight of the terrain feature values is introduced while calculating the spatial distance weight, and joint bilateral filtering is performed to generate the prior low-frequency map. This can also force the reconstructed prior low-frequency map to fit the real physical boundary of the landform.
[0062] In some embodiments, the above-described reconstruction of the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map basis to generate the prior low-frequency map may further include: Using spatial micro-topography analysis algorithms, based on a preprocessed high-resolution digital elevation model, the slope and topographic relief of the target inversion area are extracted pixel by pixel. Preliminary resampling is performed on the preprocessed coarse-resolution prior low-frequency map substrate; Based on the coarse-resolution prior low-frequency map base after initial resampling, the slope and terrain undulation are used as spatial smoothing prior terms in the energy function of the local spatial Markov random field. The prior low-frequency map is generated by iterative solution.
[0063] It can be seen that the resampling of coarse-resolution products can also be regarded as an energy minimization problem, and the extracted high-resolution slope and terrain undulation can be used as spatial smoothing priors in the MRF energy function. High-fidelity boundary reconstruction can be achieved through iterative solution.
[0064] In box 130, high-frequency spatial residual labels are explicitly extracted based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset.
[0065] In some embodiments, the macroscopic background trend and local microscopic fluctuations of the spatial distribution of surface parameters can be separated from the mathematical level, and the current single-objective direct fitting can be transformed into high-frequency residual learning oriented towards local detail variations, thereby providing accurate optimization targets (label data) for the subsequent introduction of physical mechanism constraints, and completing the explicit extraction of high-frequency spatial residual labels of measured samples.
[0066] In some embodiments, the explicit extraction of high-frequency spatial residual labels based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset specifically includes: Multi-source data point topology matching and value extraction are performed on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset to generate prior low-frequency predicted values and actual observed values corresponding to the spatial locations of each ground measured sample point. Based on prior low-frequency predicted values and true observation values, explicit spatial residual calculations are performed on all spatially matched sample points one by one to generate the corresponding true spatial residuals of surface parameters. High-frequency spatial residual labels are constructed based on the true spatial residuals of surface parameters and high-resolution environmental covariate atlases.
[0067] In some embodiments, based on the acquired ground-measured sample point surface parameter dataset, spatial topological intersection and projection alignment are performed on the generated prior low-frequency map with high-fidelity boundaries using the measured latitude and longitude coordinates of each sample point. The nearest neighbor extraction method is employed to accurately obtain the prior low-frequency predicted value corresponding to the spatial location of each measured sample point, defined as... Meanwhile, the actual observed value obtained from in-situ field measurements and laboratory physicochemical analysis of this sample point is defined as... This allows for topological matching and value extraction of multi-source data points.
[0068] In some embodiments, explicit spatial residual calculation is performed on each of the spatially matched sample points. The calculation formula is as follows: In the formula, This represents the true spatial residual of the surface parameters at the location of the measured sample point. This residual value has a clear physical meaning, namely its positive and negative deviation direction and magnitude. It quantitatively characterizes the degree of high-frequency deviation of the true distribution of surface parameters from the macroscopic prior low-frequency trend under the influence of micro-topographic cutting and local hydrological transport at the local spatial scale, thus completing the explicit calculation of the high-frequency spatial residual label.
[0069] In some embodiments, all calculated sample points The residual dataset is compiled into a target variable set. Before entering the core constraint fitting model, data quality control and outlier detection are performed on the residual dataset, such as using the Laida criterion. The Isolation Forest algorithm, also known as the Isolation Forest algorithm, aims to remove outlier residuals caused by sampling location errors, laboratory analysis errors, or interference from extreme ground features. It retains high-quality residuals after processing. The data, along with the extracted high-resolution environmental covariates (including bare soil spectral derivation indices and micro-topography and hydrological properties) corresponding to the sample point locations, together construct a standard input / output (I / O) feature label dataset to drive the subsequent physical constraint fitting module. Through the above process, the prediction target of the inverse model can be transformed from high-dimensional and complex absolute values of surface parameters into high-frequency residuals highly correlated with local topography and hydrology. This removes obstacles to the smooth injection of underlying physical laws (such as leaching and convergence constraints) from a data structure perspective, thereby completing the construction and quality control of the physical constraint target dataset.
[0070] In box 140, high-frequency residual correction maps are generated based on high-frequency spatial residual labels and preprocessed high-resolution temporal multispectral image sets through high-frequency residual learning and correction coupled with hydrophysical constraints.
[0071] In some embodiments, to break the "black box" limitation of current pure data-driven machine learning models, topographic and hydrophysical penalty constraints can be deeply coupled into the underlying optimization logic of model training. This forces the model to follow the natural geochemical evolution law when fitting high-frequency spatial residuals, fundamentally curbing anti-physical prediction under extreme geomorphic units, and completing high-frequency residual learning and correction map generation with coupled hydrophysical constraints.
[0072] In some embodiments, the high-frequency residual learning and correction coupled with hydrophysical constraints described above, generating a high-frequency residual correction map based on the high-frequency spatial residual labels and the preprocessed high-resolution temporal multispectral image set, specifically includes: Using high-frequency spatial residual labels as the target dependent variable and high-resolution environmental covariate atlases as the input independent variable matrix, a basic nonlinear fitting framework is constructed based on a tree-based ensemble algorithm with a preset loss function. Construct a piecewise, continuously differentiable physical penalty loss function for extreme terrain conditions; Construct the total loss function based on the preset Huber loss function and the physical penalty loss function; The model is trained based on the basic nonlinear fitting framework until the total loss function converges, and the physical constraint residual fitting model is output. The high-resolution environmental covariates of all effective pixels within the target inversion area are input into the physical constraint residual fitting model, and the high-frequency residual correction map is output.
[0073] In some embodiments, the explicitly extracted high-quality spatial residual labels can be... As the target dependent variable, the extracted high-resolution environmental covariate atlas (containing multidimensional spectral derivation indices and micro-topographic attributes) is used as the input independent variable matrix. A tree-based ensemble algorithm that supports custom loss functions, such as Extreme Gradient Boosting Tree (XGBoost), is selected as the basic nonlinear fitting framework to explore the mapping relationship between high-dimensional remote sensing features and high-frequency residuals, thereby completing the construction of the high-frequency residual feature set and the setting of the basic fitting framework.
[0074] In some embodiments, within the natural soil physicochemical laws, the spatial transport height of surface parameters, such as soil pH or salinity, is controlled by topography and moisture conditions. For example, in strongly leached areas with extremely steep slopes and very low TWIs, soluble basic ions are easily lost, and the residual should not show a significant positive deviation; in low-lying catchment areas with extremely high TWIs, substances easily accumulate with runoff, and their residual should not show a significant negative deviation due to evaporation and concentration. Based on this geoscientific mechanism, a piecewise continuous differentiable physical penalty loss function can be constructed for extreme topographic conditions (…). ): In the formula, This represents the predicted residual output by the model at the current iteration step. The aforementioned closed interval setting ensures that the function remains within critical points (e.g., ...). The continuity of the algorithm. To ensure the algorithm's ability to execute automatically across regions, a physical extreme value trigger threshold is required. Instead of being manually hard-coded constants, these are adaptively and dynamically extracted based on the Empirical Cumulative Density Function (eCDF) of the target region's environmental covariates. For any input feature vector... Its eCDF definition is: In the formula, The total number of valid pixels in the entire area. This is an indicator function. The inverse function is obtained by solving for it. Get dynamic threshold: Set , , This adaptive calibration mechanism ensures the scale invariance of the penalty term across different geomorphic watersheds, thereby completing the mathematical quantification and threshold calibration of the topographic-hydrological-physical penalty function.
[0075] In some embodiments, the total loss function of the underlying machine learning algorithm is reconstructed. This is defined as a joint form of data approximation loss and physical prior penalty loss: In the formula, The Huber loss function (whose hyperparameters are used to reduce outlier observation noise) is used to reduce noise interference from outliers. Adaptive optimization calibration is performed based on the extreme value distribution characteristics of the measured residuals. This represents the physical constraint weight penalty coefficient, used to balance the data fitting accuracy with the strength of geoscientific mechanism constraints. To adapt to the solution requirements of the XGBoost algorithm's custom objective function, [this is implemented / targeted]. Calculate the first gradient ( ) and second gradient ( The formula is as follows: In each gradient descent iteration of model training, the algorithm first calculates the gradient of the base Huber loss. ) and the gradient of the aforementioned physical penalty term ( ), then weighted The total first gradient is obtained by linear addition. ) and the total second gradient ( Substituting the total gradient into the second-order Taylor expansion of the total loss function, the algorithm directly intervenes in the calculation of node split gain in the decision tree through backpropagation, curbing the generation of anti-physics branches from the bottom layer. It also combines this with a spatial block cross-validation strategy to adjust hyperparameters (such as learning rate, tree depth, etc.). Optimize the value until... The model converges and outputs a physically constrained residual fitting model, thereby completing the underlying joint optimization driven by both physical mechanisms and data.
[0076] In some embodiments, the high-resolution environmental covariate matrix of all effective pixels within the target inversion area is batch input into the trained physical constraint residual fitting model to perform pixel-by-pixel spatial extrapolation. The final output is a high-resolution high-frequency residual correction map covering the entire area, defined as... This layer accurately depicts the high-frequency spatial fluctuations of surface parameters at the pixel scale relative to the macroscopic low-frequency background. Furthermore, its predicted extreme values are subject to the mandatory backstop of the underlying hydrophysical laws of the algorithm, which eliminates false prediction noise that violates natural laws from a mechanism perspective. This provides a high-quality local correction basis for the final dynamic credibility fusion, thereby completing the spatial extrapolation output of the high-frequency residual correction map.
[0077] In some embodiments, the tree-based ensemble algorithm with the aforementioned preset loss function further includes at least one of extreme gradient boosting trees, tree-based models that support custom joint gradient optimization, or physically inspired neural network frameworks.
[0078] For example, in the case of "high-frequency residual learning coupled with hydrophysical constraints", the core logic is to quantify the natural laws of surface material transport as a penalty loss and intervene in the optimization direction of the pure data-driven model. In addition to the extreme gradient boosting tree (XGBoost) used above, other equivalent alternatives can also be adopted.
[0079] Specifically, other tree-based models that support custom joint gradient optimization, such as LightGBM or CatBoost, can be used, as long as the piecewise continuous physical penalty function containing hydrological extreme triggering conditions is constructed. By sinking down to its underlying loss function to calculate the first and second gradients, high-frequency residual correction maps that conform to geoscientific logic can be output.
[0080] Alternatively, physics-informed neural network (PINNs) frameworks can be used. For example, when fitting the residuals using a multilayer perceptron (MLP) or convolutional neural network (CNN), the designed topographic and hydrological penalty equation can be directly added as a regularization term to the backpropagation loss function of the neural network. In this study, gradient descent is used to limit the output of neurons to anti-physical extremes in extreme terrain.
[0081] In summary, a physical constraint penalty term based on hydro-topographic features can be deeply coupled into the machine learning framework of spatial residual fitting, and directly participate in the underlying optimization of model training. This mechanism breaks the black-box limitation of purely data-driven models, avoids anti-physics prediction anomalies from the algorithmic level, and improves the geoscientific rationality of high-frequency residual learning results.
[0082] In box 150, based on the dynamic credibility assessment driven by spatial nonstationarity, the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map is dynamically adjusted to generate a cross-scale fusion prediction map.
[0083] In some embodiments, to achieve high-fidelity collaboration of multi-source and multi-scale data, and to break the mechanical paradigm of "adding the prior map and the predicted residual with globally fixed equal weights (e.g., 1:1)" in the current residual fusion method, it is possible to use the theory of "spatial non-stationarity" in geography to quantify the local terrain complexity and perceive the error changes of prior low-frequency information in different spatial regions in real time. This enables pixel-level dynamic reconstruction of credibility and optimal configuration of the two information streams, completing dynamic credibility assessment and cross-scale fusion reconstruction driven by spatial non-stationarity.
[0084] In some embodiments, the above-mentioned dynamic credibility assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, specifically includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined with equal weights to generate a comprehensive terrain complexity index. An adaptive dynamic credibility weight matrix is constructed based on the decay function of the comprehensive terrain complexity index. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0085] In some embodiments, the degree of spatial fragmentation of the terrain directly determines the uncertainty of the coarse-resolution product at the local scale. This can be addressed by utilizing extracted high-precision DEM data and employing a moving analysis window, such as... or Pixel neighborhood, calculate the spatial statistical index characterizing the micro-topographic complexity on a pixel-by-pixel basis, namely the local elevation standard deviation ( (reflecting surface roughness) and local slope variation coefficient ( (This reflects the degree of dramatic topographic relief).
[0086] To eliminate the difference in dimensions, the two indicators mentioned above ( and The extreme value normalization process is performed, and the terrain complexity index is generated by equal-weight linear combination. ): The index The range of values is Its spatial statistical significance lies in the fact that in the heart of flat and contiguous plains, the index approaches 0; while in the edges of gullies or mountains with severe incision and serious soil erosion, the index increases rapidly, thus quantitatively characterizing the intensity of spatial heterogeneity of the entire region's surface and completing the spatial quantification and normalization of local topographic complexity.
[0087] In some embodiments, based on the principle of spatial nonstationarity, the prediction reliability of prior low-frequency maps is higher in areas with gentle terrain, while in areas with rugged terrain, the error is amplified sharply and the reliability decreases due to the spatial scale effect. Based on this mechanism, a system for predicting... Spatial adaptive credibility dynamic weighting function: In the formula, target coordinates The dynamic confidence weight coefficient assigned to the "prior low-frequency map" at each pixel has a value range of [value range missing]. ; This is a sensitivity parameter for regulating the weight decay rate.
[0088] To ensure the objectivity of this attenuation mechanism, the parameters... Instead of being arbitrarily preset, the calibration is obtained through a validation set of spatial blocks reserved during the model training phase, with the objective function of minimizing the global root mean square error (RMSE), and using gradient optimization algorithms (or grid search) for adaptive calibration. This function mechanism ensures a dynamic response to changes in the weight space, meaning that the more complex the terrain (… Increase), weight The exponential decay forces the fusion process to automatically reduce its dependence on low-frequency prior maps with large errors, thereby completing the construction and calibration of the adaptive dynamic confidence weighting function.
[0089] In some embodiments, the dynamic credibility weight matrix covering the entire region is constructed and calibrated as described above. The generated prior low-frequency map with high-fidelity physical boundaries (corresponding pixel values are denoted as...) ), and the high-frequency residual correction map of the coupled hydrophysical constraints derived (corresponding pixel values are denoted as ), This involves performing pixel-level dynamic fusion. The mathematical essence of cross-scale fusion is not a weighted average of absolute values, but rather dynamic compensation of residuals based on the basis. The mathematical equation for fusion reconstruction is modified and defined as follows: In the formula, This is the final predicted absolute value of the surface parameters at that pixel. After this dynamic reliability fusion reconstruction, a final high-resolution cross-scale fusion prediction map covering the entire area is output. In flat and open areas ( Approaching 1), the weight of the residual compensation term approaches 0, and the model retains the gradual stability of the prior low-frequency background to the greatest extent; while in the landform fractured area ( (Sharply reduced), residual compensation item weight By increasing the model's full activation, the high-frequency remote sensing residual correction capability with physical constraints is fully activated, thus clearly and accurately reproducing the details of local variations caused by natural processes such as terrain erosion, and completing the pixel-level reconstruction output of the cross-scale fusion prediction map.
[0090] In some embodiments, the core logic of "spatial nonstationarity-driven dynamic credibility assessment" is based on geomorphic heterogeneity, and spatially adaptively and dynamically adjusts the fusion ratio of the two information streams. In addition to the exponential decay function based on comprehensive terrain complexity used above, other equivalent alternatives can be adopted, such as the sigmoid function or spatial variable coefficient allocation based on geographically weighted regression (GWR).
[0091] In some embodiments, the above-mentioned dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, further includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined with equal weights to generate a comprehensive terrain complexity index. Based on the S-shaped nonlinear mapping function, an adaptive dynamic credibility weight matrix is constructed according to the decay function of the comprehensive terrain complexity index. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0092] The above methods can be used In this form, by setting a center translation threshold With slope control parameters Similarly, it can achieve smooth and rapid decay of macroscopic low-frequency weights under complex terrain.
[0093] In some embodiments, the above-mentioned dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, further includes: Based on the preprocessed high-resolution digital elevation model, local verification error and local landform features are calculated pixel by pixel. Using geographic weighted regression or spatial variable coefficient models, with local validation error as the dependent variable and local topographic features as the independent variable, an adaptive dynamic confidence weight matrix is implicitly solved and output through local matrix operations. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
[0094] In this way, instead of relying on explicit mathematical attenuation equations, we can use GWR or spatially varying coefficient (SVC) models, with the local validation error of the sample points as the dependent variable and the local geomorphological features as the independent variable, to implicitly solve and output the weight matrix covering the entire area through local matrix operations, and finally complete the dynamic transformation to a cross-scale fusion prediction map.
[0095] In summary, an adaptive confidence evaluation function based on the Local Terrain Complexity Index (TCI) can be established to perceive the spatial nonstationarity and confidence differences of different geomorphic regions in real time, realize the dynamic weighted fusion of prior low-frequency reconstruction information and high-frequency prediction residuals at the pixel scale, and effectively suppress the error amplification effect in complex terrain regions.
[0096] It should be noted that, in order to address the systemic defects in the fusion of multi-scale features of surface parameters, such as scale artifact amplification, lack of physicochemical mechanisms, and solidification of fusion weights, this disclosure provides a high-resolution surface parameter mapping method based on adaptive fusion of physical constraints and spatial residuals through the above-mentioned specific process. It abandons the simple and mechanical splicing of various algorithm modules in the current technology, and instead constructs a complete spatial information collaborative processing closed loop of "structural downscaling - weak physical constraints - strong spatial adaptation" based on the law of geospatial nonstationarity.
[0097] Specifically, addressing the scale mismatch and prediction distortion problems encountered in existing large-scale, high-resolution surface parameter inversion methods, this disclosure breaks through the theoretical limitations of current "direct stacking method" and "fixed-weight residual fusion method," proposing a credibility reconstruction mechanism based on "spatial nonstationarity-driven" principles. Its core technical mechanism encompasses the following three dimensions: (1) Credibility spatial decoupling based on geoscience “spatial nonstationarity”.
[0098] Current residual fusion techniques generally employ globally fixed weights (such as simple 1:1 linear addition) to synthesize low-frequency macroscopic trends and high-frequency prediction residuals. This processing logic ignores the "spatial non-stationarity" law prevalent in geography. In fact, due to spatial heterogeneity, coarse-resolution products have extremely high prediction reliability in areas with flat terrain and continuous landscapes; however, in complex terrain areas with rugged terrain and fragmented surfaces, their inherent uncertainty is drastically amplified. This disclosure uses "local terrain complexity" as the driving factor for reliability allocation, abandoning the "one-size-fits-all" fusion paradigm, and realizing pixel-level dynamic reconstruction of the weights of prior low-frequency information and high-frequency residual information in different spatial regions.
[0099] (2) The natural mapping relationship between topographic physical boundaries and surface physicochemical variations.
[0100] The spatial differentiation of surface parameters (such as soil pH and organic carbon) is not random, but highly controlled by water and material transport processes driven by natural topographic undulations, such as gravitational leaching and depression convergence. This disclosure utilizes this geoscientific causal relationship, using the geomorphic abrupt change gradient extracted from a high-resolution digital elevation model (DEM) as a "spatial physical boundary skeleton" to constrain and reconstruct a smooth, coarse-resolution prior map. This mechanism can recover the true natural topological boundaries from the underlying data structure without introducing additional spectral noise, effectively overcoming the mosaic artifacts caused by traditional downscaling algorithms.
[0101] (3) The three-stage collaborative mechanism of “downscaling-weak constraint-strong adaptation”.
[0102] The technical solution disclosed herein is not a simple accumulation of existing AI algorithms, but a three-stage collaborative system with interconnected steps.
[0103] Phase 1 (Terrain-Guided Low-Frequency Reconstruction): Focuses on preserving the fidelity of macroscopic trends. Edge-guided filtering is performed using a 3D terrain skeleton to overcome spatial scale conversion barriers and generate a priori low-frequency maps with clear physical boundaries.
[0104] The second stage (physically constrained high-frequency residual learning) focuses on logical correction of local variations. It breaks the "black box" limitation of pure data-driven machine learning, quantifies topographic and hydrological mechanisms (such as extreme leaching / catchment rules) as penalty terms and directly incorporates them into the optimization loss function at the bottom layer of the model. This guides remote sensing features to accurately fit high-frequency local residuals and generates high-frequency residual correction maps that conform to geochemical logic.
[0105] The third stage (dynamic credibility fusion) focuses on the optimal configuration of global information. Using the Comprehensive Terrain Complexity Index (TCI) as a medium, the heterogeneity of each pixel is perceived in real time, and the fusion ratio of the "prior low-frequency map" and the "high-frequency residual correction map" is dynamically adjusted to finally output a high-fidelity cross-scale fused prediction map.
[0106] Through the synergy of the aforementioned underlying mechanisms, this disclosure fundamentally alleviates industry pain points such as the amplification of errors in coarse-resolution base maps under complex terrain and the tendency of pure data AI models to generate anti-physics predictions, achieving a comprehensive breakthrough in large-scale surface parametric spatial mapping in terms of boundary fidelity, physical logic rationality, and cross-regional generalization ability.
[0107] In some embodiments, the above method further includes: The reliability of the cross-scale fusion prediction map is assessed, and the uncertainty at the spatial pixel level is quantified by an ensemble learning mechanism to generate a dual-track spatial digital layer with both deterministic and uncertain aspects.
[0108] In some embodiments, the reliability of the cross-scale fusion prediction map generated by the above adaptive weighted fusion can be assessed, and the uncertainty at the spatial pixel level can be quantified using an ensemble learning mechanism to output a dual-track map with both deterministic and uncertain aspects to assist downstream application decision-making.
[0109] In some embodiments, to avoid the "falsely high accuracy" caused by spatial autocorrelation of geographic data in random cross-validation, a spatial block cross-validation mechanism is introduced.
[0110] Based on the measured latitude and longitude coordinates of the sample points, a spatial clustering algorithm was used to divide the entire sample set into... Each iteration consists of independent, non-overlapping blocks in space. The samples from each block are used as the training set, and the samples from the remaining 1 spatially independent block are used as the independent validation set, and the process is repeated cyclically. Next, on the validation set, the final cross-scale fusion prediction results will be output. With the corresponding actual observed values To compare, the root mean square error (RMSE) and the coefficient of determination ( ) were calculated. This allows for an objective assessment of the model's true spatial generalization ability across geographical units, thereby completing the accuracy assessment based on spatial block cross-validation.
[0111] In some embodiments, when performing global spatial extrapolation, since the core model employs a gradient boosting architecture (such as XGBoost), two independent uncertainty extrapolation model branches are constructed in parallel outside the main prediction model. This is achieved by setting the target loss function to an asymmetric quantile / pinball loss function, with... and For the target quantile, independently train and output the 5th quantile of the target parameter prediction results ( ) and the 95th percentile ( Considering that the final fusion result is affected by dynamic confidence weights. The uncertainty of the prior low-frequency map is greatly suppressed in the flat region due to the modulation. Therefore, the effective high-frequency uncertainty of the pixel-level 90% prediction interval width (PIW) needs to be weighted and scaled, and its rigorous calculation formula is defined as: The smaller the value of a pixel, the higher the confidence level of the inference after cross-scale fusion; conversely, the larger the value of a pixel (usually located in the sparse area of the sample or at the edge of extreme terrain, where the residual variance increases in tandem with the weight), the higher the uncertainty of the real space after fusion, thus completing the pixel-level uncertainty inference based on the quantile loss gradient.
[0112] In some embodiments, the above calculation results are integrated to finally output a dual-track spatial digital layer: the main layer is the cross-scale fusion prediction map generated in step five, which combines macroscopic physical boundaries and local high-frequency details; the auxiliary layer is the synchronously generated pixel-level uncertainty distribution map (i.e., (Layer). This dual-track output mechanism provides underlying input support for downstream surface process simulation or precision agricultural management, balancing high-precision data with quantitative confidence levels, thereby completing the output of dual-track spatial digital maps.
[0113] To facilitate understanding, a typical black soil farming area will be selected as the target inversion verification area to further elaborate on the above-mentioned surface parameter mapping method based on the adaptive fusion of physical constraints and spatial residuals.
[0114] It should be noted that in recent years, due to the long-term impact of high-intensity agricultural development, soil degradation in the black soil region has become increasingly prominent, with soil pH exhibiting extremely complex spatial heterogeneity and non-stationarity. Spatially, in some areas, influenced by the intermingling of saline-alkali land, the soil pH range is extremely wide (roughly between 4.26 and 9.69), exhibiting dramatic spatial variability; while in other areas, the pH range is relatively narrow (4.74 to 7.72), with less spatial variability. This significant difference in "spatial non-stationarity" between different landforms and soil-forming environments makes it extremely easy for digital mapping models using globally fixed weights to experience "accuracy collapse" when making cross-regional predictions. Furthermore, the existing 250m resolution prior low-frequency map base cannot characterize the differences in water transport and material deposition driven by micro-topography under complex terrain (such as the edges of erosion gullies). There is an urgent need to conduct high-resolution, high-precision spatial reconstruction of soil pH at the 30m level by introducing physical boundary constraints and spatial adaptive fusion mechanisms.
[0115] Based on this, a typical black soil farming area was selected as the target inversion verification area. This area covers a wide range, has a relatively typical topography, and includes a variety of landform types, with the main soil types including black soil, black calcareous soil, meadow soil, and dark brown soil.
[0116] To fully verify the three-stage collaborative mechanism of "downscaling-physical constraints-adaptive fusion" disclosed herein, this embodiment constructs a standardized data base covering "satellite-terrain-priority-ground" to prepare multi-source spatial datasets.
[0117] High-resolution temporal multispectral image set: Landsat 8 / 9 satellite data with a spatial resolution of 30m were selected. To eliminate the spectral masking effect of vegetation and farmland stubble in spring, the focus was on the bare soil exposure window in spring, such as April to May. By integrating the Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR2), and Normalized Differential Bare Soil Index (NDBSI), a temporal composite of the "optimal bare soil pixel" was performed to obtain a seamless 30m bare soil spectral base map with the best signal-to-noise ratio.
[0118] High-resolution terrain data (DEM): A high-precision 12.5m digital elevation model (ALOSPALSAR DEM) covering the entire area was acquired. This data was used not only to extract slope and topographic relief to construct a "physical boundary skeleton" guide map, but also to calculate the topographic wetness index (TWI) to provide hydrophysical penalty constraints, and to quantify local topographic complexity (TCI).
[0119] (Coarse resolution) Prior low-frequency map substrate: Existing national or global scale soil pH digital mapping products (spatial resolution of approximately 250m) are used as the substrate input for model downscaling reconstruction and providing macroscopic low-frequency background.
[0120] Ground-based measured sample point (surface parameter) dataset: During the same period as satellite transit, 641 independent surface (0-20cm) soil pH measurement sample points were obtained through geospatial stratified random sampling. This measured dataset covers an extremely wide pH range and diverse micro-topographic units, and will be used to extract high-frequency residual labels, construct spatial block cross-validation sets, and conduct an ablation study to validate the synergistic effects of the models.
[0121] like Figure 2 As shown, for the above-mentioned implementation environment, the specific steps of the surface parameter mapping method based on the adaptive fusion of physical constraints and spatial residuals include the following.
[0122] Step 1: Acquisition and standardized collaborative preprocessing of multi-source heterogeneous spatial data.
[0123] The preprocessing process in this embodiment first involves acquiring and collaboratively processing multi-source heterogeneous datasets within the spatial scope of the black soil region to establish a standardized feature data foundation that supports subsequent downscaling and residual fusion.
[0124] 1) Time-series image acquisition and high-quality bare soil spectral atlas reconstruction.
[0125] For the target inversion area, Landsat 8 / 9 Level-2 surface reflectance products were acquired in batches during the spring (April 1 to May 31). To minimize the spectral masking effects of early crop emergence, natural vegetation regreening, and farmland stubble during this period, Normalized Difference Vegetation Index (NDVI), Normalized Burning Ratio (NBR2), and Normalized Difference in Bare Soil Index (NDBSI) were calculated pixel-by-pixel for the preprocessed images. An NDVI > 0.3 threshold was set as the vegetation masking threshold, and a comprehensive bare soil quality score was applied to the remaining pixels (the score weight was negatively correlated with NDVI and positively correlated with NBR2 and NDBSI). Within the three-year spring time window, a "best pixel priority" strategy was adopted, extracting the effective observations with the highest bare soil score (i.e., the purest soil background signal and the highest signal-to-noise ratio) for spatial mosaicking, ultimately generating a high-quality, seamless bare soil spectral base map with a 30m resolution covering the entire black soil region.
[0126] 2) Acquisition of terrain and prior data and unification of spatial benchmarks.
[0127] Simultaneously acquire 12.5m high-precision ALOS PALSAR DEM data covering the study area, as well as existing-scale soil pH digital mapping products with a spatial resolution of approximately 250m (defined as the prior low-frequency map base in this embodiment, denoted as...). Using a 30m bare soil spectral base map as the spatial reference, all spatial layers were uniformly projected onto the WGS 1984 UTM coordinate system. Among these, the 250m resolution... The 12.5m DEM data was initially resampled to 30m using bilinear interpolation (as the initial input for subsequent terrain-guided filtering), and then resampled to a 30m cell grid to ensure strict alignment of the multi-source dataset in spatial topology and cell resolution, thus constructing a standardized multi-scale data cube.
[0128] 3) Targeted extraction of environmental variables and topological matching of measured points.
[0129] Based on the unified data cube, we conduct targeted extraction of environmental variables to provide full feature support for subsequent core algorithm modules.
[0130] Blue and green light bands and their derived indices (such as the green / near-infrared ratio and the normalized difference between blue and green light) were calculated and extracted from the 30m bare soil spectral base map for subsequent high-frequency spatial residual fitting.
[0131] Slope and terrain undulation (TRI) are extracted from the 30m DEM for subsequent construction of the spatial physical boundary skeleton.
[0132] The topographic wetness index (TWI) was further extracted from the 30m DEM and used to construct hydrophysical penalty constraints.
[0133] Subsequently, the GPS coordinates of 641 independent measured sample points collected in the early stage were imported into the system. The high-resolution environmental covariate values and macroscopic prior low-frequency background values (pH_ref) corresponding to each point were accurately extracted using the nearest neighbor method. This completed the spatial matching between the actual observed pH values of the surface soil of the sample points and the multi-source feature data, forming the benchmark feature matrix for model training and residual extraction.
[0134] Step 2: Terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints.
[0135] This embodiment addresses the complex geomorphological features of the black soil region by introducing a high-resolution topographic skeleton as a spatial physical prior constraint. It performs high-fidelity reconstruction of the physical boundaries of the coarse-resolution macroscopic soil pH product, overcoming the mosaic artifacts caused by conventional downscaling and resampling from the underlying data structure.
[0136] 1) Construction of high-resolution physical boundary skeleton and guide graph.
[0137] Based on the spatially aligned 30m resolution DEM data from Step 1, the slope and topographic relief (TRI) of the study area were calculated pixel-by-pixel using the spatial micro-topography analysis module. In the context of black soil topography, abrupt changes in slope and TRI typically correspond precisely to transition zones or the boundaries between different micro-topographic units. These physical boundaries are often the true natural boundaries where spatial variations in soil moisture transport and physicochemical properties (such as the degree of salt leaching or deposition) occur.
[0138] The extracted slope and topographic relief features were normalized to their extreme values and linearly fused into a comprehensive topographic feature matrix reflecting the high-frequency spatial texture attributes of the land surface. This matrix was defined as a guidance image with a spatial resolution of 30m (denoted as ). This means that the "spatial physical boundary skeleton" of the target inversion region was constructed.
[0139] 2) Terrain-guided filtering processing that preserves spatial physical boundaries.
[0140] The initial resampling in step one, performed using bilinear interpolation, is reduced to 30m. The input image (denoted as) is used as the filter. Subsequently, using the 30m physical boundary skeleton matrix constructed above... As a guide image, guided image filtering that preserves spatial edges is performed.
[0141] In the specific engineering parameter settings of this embodiment, it is assumed that a linear mapping relationship exists within the local window, and the local window radius of the guided filter is set accordingly. (Right now (Analysis neighborhood of a cell), smoothing penalty regularization parameter By solving the problem of minimizing the output image relative to the input coarse resolution... The algorithm establishes a linear mapping coefficient between the terrain skeleton and soil pH within a local window, considering the reconstruction error and smoothing cost function. This processing procedure forces the guiding image... The high-frequency topographic abrupt gradients (i.e., physical boundaries of landforms) contained therein are directionally migrated to the relatively blurred coarse-resolution pH input image.
[0142] 3) Prior low-frequency map output with high-fidelity boundary.
[0143] After guided filtering, a 30m resolution prior low-frequency map is output. Compared with the smoothed maps generated by current bilinear or cubic convolution resampling, this map preserves the original image quality on a macroscopic global scale. Large-scale soil pH gradients show a smooth transition from strongly alkaline to slightly acidic regions, while at the local pixel scale, clear natural topography and landform boundaries are precisely reconstructed.
[0144] This physical mechanism-based reconstruction process effectively corrects the local smoothing effect caused by spatial scale mismatch, eliminates mosaic artifacts in the image, and provides a base data layer with high physical fidelity for the accurate extraction of high-frequency spatial residuals of measured sample points in subsequent steps.
[0145] Step 3: Explicit extraction of high-frequency spatial residual labels from actual samples.
[0146] This step aims to mathematically separate the macroscopic background trend and local microscopic fluctuations of soil pH in the black soil region, and to establish an accurate mapping between high-resolution multispectral data and the high-frequency variation residuals of surface parameters.
[0147] 1) Spatial topology matching of multi-source datasets of sample points.
[0148] Based on the standard spatial benchmark established in the preceding steps, the GPS coordinates of 641 ground-measured sample points were imported into the processing system. Using coordinate projection transformation technology, all sample points were precisely located into their corresponding 30m pixel grids. Subsequently, the nearest neighbor extraction method was employed to extract the predicted prior low-frequency map value (denoted as ) at each sample point location after physical boundary reconstruction. At the same time, record the actual ground measurement value of the sample point (denoted as ). ), as well as high-resolution environmental covariates extracted simultaneously, including multispectral derivation indices and micro-topographic features.
[0149] 2) Explicit calculation of high-frequency spatial residuals of surface parameters.
[0150] For 641 successfully matched sample points, the true high-frequency spatial residual of soil pH was calculated point by point. The calculation formula is as follows: Calculated The study quantitatively characterized the high-frequency variation amplitude of soil pH value on the macroscopic a priori low-frequency trend under typical landforms, caused by local micro-topographic undulations, variations in water conditions, or human activities (such as local application of soil conditioners at the field scale).
[0151] 3) Construction and quality control of the physical constraint target dataset.
[0152] To ensure the training quality of subsequent machine learning models coupled with physical constraints, rigorous data cleaning was performed on the obtained residual label dataset. An improved Laida criterion (such as...) was adopted. Using the median absolute deviation method, outlier residuals that may be caused by sampling location errors or laboratory analysis mistakes are removed. After quality control, high-quality, high-frequency spatial residual labels are retained. As the dependent variable (prediction target) of the core constraint fitting model, the synchronously extracted high-resolution environmental covariate map is used as the independent variable matrix to construct a standardized input / output (I / O) feature sample set to drive the subsequent physical constraint fitting module. This transformation process effectively avoids the model directly fitting the global pH absolute value with an extremely wide span and strong spatial non-stationarity, significantly reducing the risk of overfitting in cross-regional predictions.
[0153] Step 4: High-frequency residual learning and correction graph generation with coupled hydrophysical constraints.
[0154] This embodiment addresses the complex topography and strong spatial variability of soil physicochemical properties in the black soil region. It breaks through the limitations of the pure data-driven black box approach of current machine learning models by incorporating local topographic and hydrological processes as prior physical constraints into the underlying framework of residual fitting, thus ensuring the geoscientific rationality and cross-regional generalization ability of the high-resolution prediction results.
[0155] 1) Input feature set construction and model initialization.
[0156] Based on the residual labels of 641 high-quality measured sample points obtained in step three ( The target dependent variable was selected. For the independent variable input, conventional feature set stacking was abandoned to reduce the curse of dimensionality. High-resolution environmental covariates most sensitive to soil pH variations in the black soil region were extracted as the input matrix, including: 30 mLandsat bare soil spectral-derived indices, such as the green / near-infrared ratio and the normalized difference between blue and green light, etc., and micro-topographic attributes derived from the 12.5 m DEM (slope, topographic wetting index TWI). Extreme gradient boosting (XGBoost), which supports custom loss functions and gradient optimization, was selected as the basic nonlinear fitter.
[0157] 2) Quantitative design and threshold calibration of topographic, hydrological and physical penalty items in black soil region.
[0158] In the black soil region, the spatial differentiation of soil pH is highly controlled by topographic and hydrological-driven material transport. In gently sloping hills, surface runoff is intense, and basic ions are easily leached by gravity, resulting in lower soil pH. Conversely, in low-lying catchment areas with higher total water volume (TWI), such as eroded gully bottoms or enclosed depressions, soluble salts tend to accumulate with water flow, leading to a positive pH deviation. Based on these soil-forming physicochemical principles, this embodiment quantifies the hydrophysical penalty loss term (…). ): In the formula, This represents the predicted residuals output by the model iteration. To ensure the adaptability of this constraint mechanism, this embodiment utilizes an inverse function based on eCDF to calculate the dynamic threshold: Set as the 90th percentile (approximately 15°) of the eCDF for the entire region's slope. Set as the 10th percentile of the entire TWI eCDF. Set to the 90th percentile of the TWI eCDF for the entire region. This function limits the model from the algorithmic level from predicting spurious high pH values in extreme leaching areas or anomalous strong acidification in extreme catchment areas.
[0159] 3) Underlying joint optimization and spatial cross-validation of the dual-drive model.
[0160] During the training phase of the XGBoost model, reconstruct the custom joint loss function: The basic data fitting loss uses the Huber loss function (with set parameters). To mitigate the interference of individual extreme outliers in the measured residuals; These are the physical constraint weights. In each gradient descent iteration, the algorithm addresses the above... The first-order gradient (Jacobian) and second-order gradient (Hessian) are calculated, and backpropagation is used to directly intervene in and guide the node splitting gain of the XGBoost decision tree. To avoid "covariate shift" interference caused by the wide-range sample region on the narrow-range sample region, this embodiment strictly implements 5-fold spatial block cross-validation, and determines the optimal learning rate, maximum tree depth, and [other parameters] using a grid search method. Value, until convergence.
[0161] 4) Spatial extrapolation of the high-frequency residual correction map for the entire region.
[0162] A 30m high-resolution environmental covariate raster layer covering the entire black soil region was batch-input into the trained coupled physical constraint XGBoost residual fitting model. The model performed spatial forward extrapolation pixel by pixel, outputting a 30m resolution high-frequency residual correction map covering the entire study area (denoted as...). ).
[0163] This layer not only captures the high-frequency spatial fluctuations of soil pH at the field scale using remote sensing spectral indices, but also undergoes forced optimization by physical penalty terms. The distribution of its predicted extreme values strictly matches the topographic and hydrological evolution of the black soil, providing a local correction layer with clear geological basis for the final multi-scale fusion.
[0164] Step 5: Dynamic credibility assessment and cross-scale fusion reconstruction driven by spatial nonstationarity.
[0165] This embodiment addresses the significant "spatial nonstationarity" in large-scale cross-regional inversion in the black soil region, breaking away from the mechanical paradigm of global equal-weight (1:1) addition in current residual fusion. Based on the dynamic reconstruction of confidence weights for low-frequency and high-frequency information according to geomorphological heterogeneity, it achieves optimal synergy between macroscopic trends and local variations.
[0166] 1) Spatial quantification and normalization of Local Terrain Complexity (TCI).
[0167] Using high-precision DEM data acquired earlier and resampled to 30m, a moving analysis window technique is employed to assess local topographic fragmentation. To match the initial spatial scale of the 250m prior low-frequency map, this embodiment sets the moving window size to [size missing]. Pixel, corresponding to approximately the ground area .
[0168] Calculate the local elevation standard deviation pixel by pixel within this window. ) and local slope variation coefficient ( To eliminate the difference in dimensions, extreme value normalization was performed on the two indicators separately, and then an equal-weighted linear combination was used to generate the comprehensive terrain complexity index. ): In open hinterland, where the terrain is gently undulating, this index... The index approaches 0; however, in transitional zones or areas with severe water erosion, the index increases rapidly, thus enabling a high-precision quantitative characterization of the spatial heterogeneity of the entire black soil region.
[0169] 2) Objective calibration of the adaptive dynamic credibility weighting function.
[0170] Based on the credibility decay law, construct a system based on... Exponential dynamic weighting function: In the formula, The dynamic confidence weight coefficients are assigned to the "prior low-frequency map" at the target pixel. To ensure the objectivity and regional adaptability of this function, this embodiment utilizes the 5-fold spatial block cross-validation set (covering 641 measured samples) divided in step four, with minimizing the root mean square error (RMSE) of the global validation set as the optimization objective function, and employs a grid search algorithm to optimize the sensitivity parameters. Adaptive optimization calibration is performed. In the black soil scenario of this embodiment, the optimal parameters are finally calibrated. .
[0171] This function mechanism ensures that in areas with complex terrain, the weights... The exponential drop forces the fusion system to automatically reduce its reliance on the low-frequency prior basis with severe error polarization and instead trust the high-resolution prediction residuals.
[0172] 3) Pixel-level reconstruction output of cross-scale fusion prediction map.
[0173] Based on the above calibration, a global 30m dynamic credibility weight matrix is generated. The 30m prior low-frequency map with high-fidelity physical boundaries generated in step two (denoted as...) ), and the 30m high-frequency residual correction map with coupled hydrophysical constraints generated in step four (denoted as ), Perform pixel-by-pixel dynamic residual compensation fusion: After fusion, a 30m high-resolution cross-scale fusion prediction map of soil pH distribution in black soil farmland covering the entire validation area is output. This map is applicable to flat agricultural areas (…). Without excessive residual compensation, it effectively mitigates the prediction noise caused by soil moisture and roughness interference inherent in high-resolution remote sensing spectra, perfectly inheriting the gradual stability of the macroscopic low-frequency background; while in topographically fragmented areas ( (sharp reduction), then by giving With extremely high compensation weights, it accurately reproduces the details of microscale soil pH variation caused by water migration, fundamentally overcoming the scale error caused by simply magnifying the base map or the anti-physical artifacts caused by simply relying on AI prediction.
[0174] Step Six: Model Accuracy Evaluation and Pixel-Level Spatial Uncertainty Quantification Output.
[0175] This embodiment objectively evaluates the spatial generalization accuracy of the high-resolution cross-scale fusion prediction results of the black soil region generated in step five, and simultaneously quantifies the pixel-level prediction uncertainty, outputting a dual-track data base to support downstream precision agriculture decision-making.
[0176] 1) Accuracy assessment based on spatial block cross-validation.
[0177] Based on the 5-fold spatial blocks defined in the embodiment, the extracted values (denoted as ) of the 30m resolution cross-scale fusion prediction map finally generated in step five at the sample point locations are used. ) and 641 actual ground-measured values (denoted as A one-to-one comparison was performed. The root mean square error (RMSE) of the validation set and the coefficient of determination were calculated. This objectively quantifies the true spatial generalization ability of the fusion system when facing different landform units, and effectively eliminates the "falsely high accuracy" phenomenon caused by spatial autocorrelation effect in the current random cross-validation.
[0178] 2) Pixel-level uncertainty derivation based on quantile loss gradient.
[0179] During the full-domain 30m spatial extrapolation, this embodiment simultaneously activates the quantile regression mechanism of the ensemble model. For the XGBoost architecture, the system instantiates two independent decision tree cluster branches in parallel, setting their objective loss function to an asymmetric quantile loss function, and independently extrapolates and outputs the 5th quantile of the target parameter prediction results. ) and the 95th percentile ( To map the true confidence level after fusion, dynamic confidence weights are introduced simultaneously into the pixel-level 90% prediction width (PIW). Scaling is performed, and the formula is defined as follows: In flat agricultural areas with dense sample distribution and homogeneous micro-topography ( Approaching 1). A value close to 0 indicates extremely high confidence in the model's extrapolation; however, in transition zones with sparse sample points or at the edges of complex erosion gullies with strong scouring, the weights... As the residual variance increases, The value increased significantly. This quantitative indicator intuitively and rigorously characterizes the real spatial uncertainty after cross-scale fusion.
[0180] 3) Output of dual-track spatial digital maps of the black soil region.
[0181] The final output is a dual-track 30m resolution spatial digital product covering the target inversion area. The main layer is a cross-scale fusion prediction map of surface soil pH values, which combines macroscopic physical boundaries with local high-frequency details; the auxiliary layer is a pixel-level aligned... Spatial uncertainty distribution map.
[0182] This dual-track digital base effectively isolates the sources of error from macro trends and local high-frequency variations, directly overcoming the shortcomings of relying solely on low-resolution historical products to guide precision farming. It can be used as a standardized, high-quality underlying spatial parameter, directly input into the farmland plot-level variable fertilizer prescription generation system and the black soil water and soil erosion evolution assessment model.
[0183] Concurrently, to realize the aforementioned high-resolution mapping method driven by spatial nonstationarity, a matching computer execution system was constructed. This system coordinates the operation of the following five core functional modules through low-level instructions to support automated inversion mapping of large-scale spatial information.
[0184] (1) Multi-source heterogeneous data standardization and collaboration module.
[0185] This module is equipped with a data interface and a preprocessing engine, and is configured to: receive coarse-resolution prior products, high-resolution time-series multispectral images, high-precision DEMs, and measured sample point coordinates; perform "optimal bare soil pixel priority" mosaicking of time-series images through a vegetation index joint scoring mechanism to reconstruct a high-resolution bare soil spectral atlas without cloud cover or vegetation interference; and unify the spatial projection benchmark for the entire region, extract multidimensional spectral derived indices and micro-topographic and hydrological attribute features to construct a standardized full-volume high-frequency feature atlas.
[0186] (2) Physical boundary skeleton extraction and guided reconstruction module.
[0187] This module includes a spatial filtering computation unit, which is configured to: extract slope and topographic relief using a high-resolution DEM, construct a "spatial physical boundary skeleton" guide map that reflects the abrupt changes in the actual terrain; perform edge-guided filtering operations based on the least squares cost function and smoothing regularization, and directionally transfer the high-frequency terrain abrupt gradient to the coarse-resolution product, outputting a priori low-frequency map with high-fidelity natural boundaries.
[0188] (3) Residual learning module driven by both physical mechanism and data.
[0189] This module includes a feature matching unit and an integrated machine learning engine, configured to: subtract the actual values of measured sample points from the prior low-frequency map to explicitly extract high-frequency spatial residual labels; use preset percentile thresholds to identify extreme leaching and catchment areas, and construct a piecewise continuous and differentiable physical penalty loss function; by calculating the gradient of this penalty term and backpropagating it, directly intervene in the node splitting optimization process of the tree-based model (such as XGBoost), and then extrapolate the output of a high-frequency residual correction map that conforms to geochemical logic for the entire region.
[0190] (4) Spatial nonstationarity perception and dynamic fusion module.
[0191] This module serves as the central hub for fusion and reconstruction in this system. It is configured to: calculate the local elevation standard deviation and slope variation coefficient based on the DEM moving analysis window to generate the Comprehensive Terrain Complexity Index (TCI); perform exponential decay mapping based on this index and dynamically calibrate the confidence weights pixel by pixel; finally, perform pixel-level weighted fusion of the "prior low-frequency map" and the "high-frequency residual correction map" according to the spatial adaptive weight matrix, and output a high-precision cross-scale fusion prediction map.
[0192] (5) Dual-track uncertainty quantification and mapping output module.
[0193] This module includes spatial validation and variance evaluation units, configured to perform spatial block cross-validation to output the model's true generalization accuracy (such as RMSE and variance). During the full-area extrapolation, an uncertainty assessment mechanism based on quantile loss gradients is activated. The 5th and 95th predicted quantiles are calculated pixel by pixel, and an uncertainty layer with a 90% prediction interval width (PIW) is generated by combining dynamic confidence weights. Finally, the deterministic prediction results and the uncertainty confidence layer are synchronously encapsulated and output in a dual-track manner.
[0194] In summary, the core technologies disclosed herein focus on three key nodes in spatial downscaling inversion, forming an inseparable technological loop. First, at the bottom-level data reconstruction layer, high-precision DEM topographic skeleton features (such as slope and undulation) are used as prior constraints to perform "Guided Image Filtering" on the coarse-resolution attribute base map, eliminating the smoothing effect and mosaic artifacts caused by downscaling from the data structure with extremely low computational cost. Second, at the core residual fitting layer, the limitations of pure data-driven AI are broken, quantifying hydrophysical laws (such as the water catchment / leaching mechanism based on the topographic wetness index TWI) into an asymmetric penalty loss function, forcibly constraining the fitting logic of tree-based machine learning models (such as XGBoost), avoiding abnormal predictions that violate geoscientific common sense. Third, at the top-level spatial fusion layer, a spatially adaptive dynamic weighting function based on Local Topographic Complexity (TCI) is proposed, replacing the traditional global equal-weighted addition, achieving pixel-level site-specific fusion and completely solving the spatial non-stationarity error caused by topographic heterogeneity.
[0195] This disclosure has significant engineering and business application value and broad application and promotion potential. Its application prospects can directly serve the following fields.
[0196] (1) Precision agriculture and farmland quality protection: It can generate high-fidelity digital base maps of soil physicochemical properties (such as soil pH value, organic matter, etc.) with a resolution of better than 30m, providing reliable data support for farmland prescription map generation, variable fertilization operations and black soil protection projects.
[0197] (2) Ecological environment monitoring and degradation assessment: By accurately depicting the spatial distribution details of physicochemical parameters, it supports the monitoring of soil acidification, secondary salinization and land degradation evolution at the watershed scale, and provides a quantitative decision-making basis for ecological fragile area restoration projects.
[0198] (3) Earth System Science and Surface Process Simulation: Overcoming the shortcomings of traditional input parameters that violate physical laws, it serves as a high-quality pre-input parameter for Earth system models, hydrological and carbon cycle models, effectively reducing the uncertainty of large-scale ecological simulation.
[0199] (4) Large-scale high-precision digital soil base update: effectively revitalize the existing massive amount of coarse-resolution historical survey data, and achieve seamless conversion to high-resolution level through the cross-scale collaborative architecture of the present invention, supporting the fine mapping and spatiotemporal dynamic update of the results of the new round of soil resource survey.
[0200] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.
[0201] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.
[0202] Figure 3 A block diagram of a surface parametric mapping apparatus 300 based on adaptive fusion of physical constraints and spatial residuals according to an embodiment of the present disclosure is shown. Figure 3 As shown, the device 300 includes: The acquisition module 310 is used to acquire multi-scale, multi-source heterogeneous spatial data of the target inversion area and perform standardized collaborative preprocessing. The multi-scale, multi-source heterogeneous spatial data includes a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. The reconstruction module 320 is used to reconstruct the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map base to generate a prior low-frequency map. The explicit extraction module 330 is used to explicitly extract high-frequency spatial residual labels based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset. The residual correction module 340 is used to learn and correct high-frequency residuals by coupling hydrophysical constraints, and to generate high-frequency residual correction maps based on high-frequency spatial residual labels and preprocessed high-resolution temporal multispectral image sets. The fusion module 350 is used for dynamic credibility assessment driven by spatial nonstationarity, dynamically adjusting the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map.
[0203] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0204] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0205] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.
[0206] Figure 4 A block diagram of an exemplary electronic device 400 capable of implementing embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.
[0207] Electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in ROM 402 or a computer program loaded into RAM 403 from storage unit 408. RAM 403 can also store various programs and data required for the operation of electronic device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. I / O interface 405 is also connected to bus 404.
[0208] Multiple components in electronic device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of displays, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows electronic device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.
[0209] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 100 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of method 100 described above may be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform method 100 by any other suitable means (e.g., by means of firmware).
[0210] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0211] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0212] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0213] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including voice input, speech input, or tactile input).
[0214] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.
[0215] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.
[0216] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.
[0217] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A method for mapping earth parameters based on adaptive fusion of physical constraints and spatial residuals, characterized in that, include: Acquire multi-scale, multi-source, heterogeneous spatial data of the target inversion region and perform standardized collaborative preprocessing; The multi-scale, multi-source heterogeneous spatial data includes a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground-measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. The preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map base are reconstructed to generate a prior low-frequency map. Based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset, high-frequency spatial residual labels are explicitly extracted; By learning and correcting high-frequency residuals coupled with hydrophysical constraints, a high-frequency residual correction map is generated based on the high-frequency spatial residual labels and the preprocessed high-resolution temporal multispectral image set. Based on the dynamic credibility assessment driven by spatial nonstationarity, the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map is dynamically adjusted to generate a cross-scale fusion prediction map.
2. The method according to claim 1, characterized in that, The acquisition of multi-scale, multi-source, heterogeneous spatial data of the target inversion region, and the standardization and collaborative preprocessing, include: Acquire multi-scale, multi-source heterogeneous spatial data of the target inversion area, including a high-resolution digital elevation model, a coarse-resolution prior low-frequency map base, a ground measured sample point surface parameter dataset, and a high-resolution time-series multispectral image set. Based on the time-series bare soil quality score, the high-resolution time-series multispectral image set is reconstructed to generate a high-resolution bare soil spectral base map; the scoring mechanism constraints of the time-series bare soil quality score include: the score is negatively correlated with the normalized vegetation index of the image, and positively correlated with the normalized burn ratio and the normalized differential bare soil index. Using the high-resolution bare soil spectral base map as the target spatial reference, spatial reference unification and environmental variable extraction are performed on the high-resolution digital elevation model, the coarse-resolution prior low-frequency map base, and the ground measured sample point surface parameter dataset, and a high-resolution environmental covariate atlas is generated.
3. The method according to claim 1, characterized in that, The process of reconstructing the preprocessed high-resolution digital elevation model and the coarse-resolution prior low-frequency map basis to generate the prior low-frequency map includes: Using spatial micro-topography analysis algorithms, based on a preprocessed high-resolution digital elevation model, the slope and topographic relief of the target inversion area are extracted pixel by pixel. The slope and terrain undulation are subjected to extreme value normalization and then linearly fused with equal weights into a physical boundary skeleton matrix. Preliminary resampling is performed on the preprocessed coarse-resolution prior low-frequency map substrate; Based on the coarse-resolution prior low-frequency map base after preliminary resampling and the physical boundary skeleton matrix, terrain-guided low-frequency reconstruction based on physical boundary skeleton constraints is performed to generate a prior low-frequency map.
4. The method according to claim 2, characterized in that, The explicit extraction of high-frequency spatial residual labels based on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset includes: Multi-source data point topology matching and value extraction are performed on the prior low-frequency map and the preprocessed ground measured sample point surface parameter dataset to generate prior low-frequency predicted values and actual observed values corresponding to the spatial locations of each ground measured sample point. Based on the prior low-frequency predicted values and true observed values, explicit spatial residual calculations are performed on all spatially matched sample points one by one to generate the corresponding true spatial residuals of the land surface parameters. Based on the true spatial residuals of the surface parameters and the high-resolution environmental covariate atlas, high-frequency spatial residual labels are constructed.
5. The method according to claim 2, characterized in that, The high-frequency residual learning and correction through coupled hydrophysical constraints, generating a high-frequency residual correction map based on the high-frequency spatial residual labels and the preprocessed high-resolution temporal multispectral image set, includes: Using the high-frequency spatial residual label as the target dependent variable and the high-resolution environmental covariate map as the input independent variable matrix, a basic nonlinear fitting framework is constructed based on a tree-based ensemble algorithm with a preset loss function. Construct a piecewise, continuously differentiable physical penalty loss function for extreme terrain conditions; Construct a total loss function based on the preset Huber loss function and the physical penalty loss function; The model is trained based on the aforementioned basic nonlinear fitting framework until the total loss function converges, and the physical constraint residual fitting model is output. The high-resolution environmental covariates of all effective pixels within the target inversion area are input into the physical constraint residual fitting model, and a high-frequency residual correction map is output.
6. The method according to claim 5, characterized in that, The tree-based ensemble algorithm for the preset loss function includes at least one of extreme gradient boosting trees, tree-based models that support custom joint gradient optimization, or physically inspired neural network frameworks.
7. The method according to claim 1, characterized in that, The dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined linearly with equal weights to generate a comprehensive terrain complexity index. An adaptive dynamic credibility weight matrix is constructed based on the decay function of the comprehensive terrain complexity index. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
8. The method according to claim 1, characterized in that, The dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, the local elevation standard deviation and local slope variation coefficient are calculated pixel by pixel. The local elevation standard deviation and local slope variation coefficient are normalized for extreme values and then combined linearly with equal weights to generate a comprehensive terrain complexity index. Based on the S-shaped nonlinear mapping function, and according to the decay function of the comprehensive terrain complexity index, an adaptive dynamic credibility weight matrix is constructed. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
9. The method according to claim 1, characterized in that, The dynamic reliability assessment based on spatial nonstationarity, which dynamically adjusts the cross-scale fusion reconstruction ratio of the prior low-frequency map and the high-frequency residual correction map to generate a cross-scale fusion prediction map, includes: Based on the preprocessed high-resolution digital elevation model, local verification error and local landform features are calculated pixel by pixel. Using geographic weighted regression or spatial variable coefficient models, with local validation error as the dependent variable and local topographic features as the independent variable, an adaptive dynamic confidence weight matrix is implicitly solved and output through local matrix operations. Based on the adaptive dynamic confidence weight matrix, the prior low-frequency map and the high-frequency residual correction map are dynamically fused at the pixel level to generate a cross-scale fused prediction map.
10. The method according to any one of claims 1 to 9, characterized in that, The method further includes: The reliability of the cross-scale fusion prediction map is evaluated, and the uncertainty at the spatial pixel level is quantified using an ensemble learning mechanism to generate a dual-track spatial digital layer with both deterministic and uncertain aspects.
