Remote sensing ecological parameter fusion method and system based on scale conversion

Abstract

The application discloses a remote sensing ecological parameter fusion method and system based on scale conversion and relates to the technical field of remote sensing data processing. The method comprises the following steps: acquiring multi-source sensing data in a target region and outputting a standardized scale data set; extracting local spatial heterogeneity characteristic values of each grid unit in the standardized scale data set and inputting the local spatial heterogeneity characteristic values into a heterogeneity scale mapping model to generate a dynamically matched scale response function; performing nonlinear scale up-mapping by using the scale response function to obtain an intermediate state characteristic flow subjected to scale deviation correction; constructing an ecological process consistency constraint equation and defining a coupled target function; performing iterative optimization calculation to output a collaborative correction factor, adaptively weighting and correcting the intermediate state characteristic flow until the coupled target function converges; and weighting and fusing the corrected intermediate state characteristic flow in the convergent state and the collaborative correction factor to output a target ecological parameter product. The application effectively improves the precision and physical consistency of scale conversion of multi-source sensing data.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, specifically to a method and system for fusing remote sensing ecological parameters based on scale transformation. Background Technology

[0002] With the rapid development of satellite remote sensing technology, acquiring large-scale ecological parameters (such as NDVI, LAI, and LST) using multi-source sensing data has become a core method for global environmental monitoring. In practical applications, in order to balance high-frequency dynamic monitoring and fine spatial identification, it is usually necessary to perform spatiotemporal fusion of low spatial resolution, high revisit frequency satellite imagery (such as MODIS) and high spatial resolution but long-period imagery (such as Landsat and GF series).

[0003] However, most existing remote sensing fusion algorithms are based on pure statistics or simple linear weighted models, ignoring the nonlinear scaling effects caused by the spatial heterogeneity of the land surface. Due to the high complexity of the spatial distribution of land cover (such as vegetation, soil, and water bodies), low-resolution pixels often contain multiple heterogeneous sub-pixels, which leads to a significant mathematical deviation between the ecological parameters retrieved directly at the low-resolution scale and the arithmetic mean of the high-resolution parameters.

[0004] More seriously, traditional fusion paradigms often treat each ecological parameter as an independent observation indicator. Under complex underlying surfaces or drastic weather fluctuations, this isolated control logic that relies solely on data-driven approaches can lead to physical decoupling or even conflict between the fused parameter products (such as temperature and vegetation indices). This not only fails to accurately reflect the continuity of surface ecological processes, but also results in erroneous decision-making in subsequent ecological assessments and disaster early warning applications due to the "physical imbalance" of input parameters. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for fusing remote sensing ecological parameters based on scale transformation.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses a method for fusing remote sensing ecological parameters based on scale transformation, comprising the following steps:

[0008] Acquire multi-source sensor data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset;

[0009] Extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into a preset heterogeneity scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values;

[0010] The scale response function is used to perform a nonlinear scale up-mapping mapping on the standardized scale dataset to obtain an intermediate feature stream corrected for scale bias.

[0011] An ecological process consistency constraint equation is constructed, and the intermediate state characteristic flow is substituted into the ecological process consistency constraint equation as an input variable. A coupled objective function containing observation data bias term and physical conservation residual term is defined.

[0012] The iterative optimization calculation aims to minimize the coupling objective function, and outputs a collaborative correction factor in each iteration. and utilize the aforementioned collaborative correction factor The intermediate feature stream is adaptively weighted and corrected to generate a corrected intermediate feature stream; the corrected intermediate feature stream is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next iteration until the coupled objective function converges.

[0013] The corrected intermediate-state feature stream obtained in the convergence state is combined with the co-correction factor output from the last iteration. The weighted fusion process is performed to output the target ecological parameter product.

[0014] Secondly, this invention discloses a remote sensing ecological parameter fusion system based on scale transformation, which uses the aforementioned remote sensing ecological parameter fusion method based on scale transformation, including:

[0015] The data perception and alignment module is used to acquire multi-source sensing data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset.

[0016] The scale mapping module is used to extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into a preset heterogeneous scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values.

[0017] The compensation processing module is used to perform a nonlinear scale up-mapping mapping on the standardized scale dataset using the scale response function to obtain an intermediate feature stream corrected for scale bias.

[0018] The module is used to construct the ecological process consistency constraint equation, and the intermediate state feature flow is substituted into the ecological process consistency constraint equation as an input variable. A coupled objective function containing observation data bias term and physical conservation residual term is defined.

[0019] The constraint iteration module is used to perform iterative optimization calculations with the goal of minimizing the coupling objective function, and outputs a collaborative correction factor in each iteration. and utilize the aforementioned collaborative correction factor The intermediate feature stream is adaptively weighted and corrected to generate a corrected intermediate feature stream; the corrected intermediate feature stream is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next iteration until the coupled objective function converges.

[0020] The product generation module is used to combine the corrected intermediate-state feature stream obtained in the convergence state with the co-correction factor output from the last iteration. The weighted fusion process is performed to output the target ecological parameter product.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention introduces local spatial heterogeneity eigenvalues ​​as dynamic adjustment factors, enabling the scale response function to perform nonlinear mapping differently according to the degree of heterogeneity within each pixel. This effectively compensates for scale transformation bias caused by spatial heterogeneity and avoids the problem of large errors in heterogeneous regions in traditional methods.

[0023] 2. This invention incorporates physical mechanisms such as surface energy balance constraints into the optimization objective in the form of residual functions, so that the iterative optimization process not only pursues the fitting accuracy of observation data, but also forces the physical consistency of ecological processes to be satisfied, thereby solving the technical problem that the ecological parameter inversion results contradict the physical mechanisms in traditional methods.

[0024] 3. By constructing an iterative loop, this invention enables the collaborative correction factor to dynamically balance between observation data bias and physical conservation residuals, successively approaching the optimal solution. The final output target ecological parameter product maintains the authenticity of the observation data and conforms to the physical constraints of the ecological process, thereby achieving a synergistic improvement in accuracy and physical consistency in the field of remote sensing ecological parameter inversion. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0026] Figure 1 This is a flowchart of the steps of the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the working principle of the present invention;

[0028] Figure 3 This is a system module connection diagram of the present invention. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] In existing technologies, scale transformation of multi-source sensing data often relies on statistical regression or simple aggregation methods, which struggle to balance the dual constraints of nonlinear bias caused by spatial heterogeneity and the physical consistency of ecological processes. Traditional scale-up methods typically employ fixed function forms or global parameters, failing to dynamically adjust the transformation model based on the local variability of land cover, leading to significant scale biases in highly heterogeneous regions. Furthermore, existing technologies lack a mechanism to effectively integrate physical process models with data-driven methods. This results in ecological parameter products, while fitting well to observational data, may violate objective physical laws such as surface energy balance and water-heat exchange, failing to meet the application requirements of precise ecological monitoring and resource management.

[0031] To address the aforementioned issues, this study discovered a correlation between local spatial heterogeneity of the Earth's surface and nonlinear bias in scale transformation. By extracting heterogeneity features such as sub-pixel variance and texture contrast, a dynamically matched scale response function can be constructed to achieve adaptive bias correction. Furthermore, it was found that incorporating constraints from ecological processes such as surface energy balance into the optimization objective as residual functions can enforce physical consistency constraints during scale transformation. Further, through iterative optimization to calculate a collaborative correction factor, a dynamic balance mechanism is formed between observational data bias and physically conserved residuals, ultimately achieving synergistic optimization of scale transformation accuracy and physical consistency.

[0032] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0033] Example 1:

[0034] like Figure 1 As shown, the remote sensing ecological parameter fusion method based on scale transformation includes the following steps:

[0035] Acquire multi-source sensor data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset;

[0036] Extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into the preset heterogeneity scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values;

[0037] The scaling response function is used to perform a nonlinear scaling up mapping on the standardized scaling dataset to obtain the scale-biased intermediate feature flow through image correction; the nonlinear radiation bias caused by spatial heterogeneity is compensated by image correction.

[0038] An ecological process consistency constraint equation is constructed, and the intermediate state characteristic flow is substituted into the ecological process consistency constraint equation as an input variable. A coupled objective function containing observation data bias term and physical conservation residual term is defined.

[0039] The iterative optimization calculation aims to minimize the coupling objective function, and outputs the collaborative correction factor in each iteration. And utilize the collaborative correction factor Adaptive weighted correction is applied to the intermediate feature flow to generate a corrected intermediate feature flow; the corrected intermediate feature flow is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next iteration until the coupled objective function converges.

[0040] The corrected intermediate-state feature stream obtained in the convergence state is combined with the co-correction factor output from the last iteration. The weighted fusion process is performed to output the target ecological parameter product.

[0041] like Figure 2 The diagram illustrates the working principle of this application. The system uses a data sensing interface to call hardware storage and communication units composed of satellite remote sensing ground stations or distributed cloud storage arrays to acquire multi-source sensing data within the target area. The data types include optical images, microwave radar data, and digital elevation models (DEMs). In this embodiment, the acquisition of multi-source sensing data relies on multi-dimensional observation platforms such as spaceborne, airborne, and ground stations, covering multiple modes including optical multispectral image sensing, thermal infrared image sensing, and synthetic aperture radar (SAR) active microwave image sensing. Through image sensing technologies with different physical mechanisms, the system simultaneously captures the reflectivity, thermal radiance, and backscattering coefficient information of surface vegetation, providing multi-dimensional and multi-basic raw data support for subsequent cross-scale analysis of ecological parameters. Image sensing includes, but is not limited to, passive optical sensing and active microwave sensing, used to sense the structural and dielectric properties of surface cover, respectively.

[0042] During spatiotemporal registration, the processor utilizes parallel computing arrays (such as GPUs or multi-core CPUs) to perform geometric correction and resampling on data generated by different sensors at different orbital altitudes, imaging times, and projection coordinates. By establishing a unified spatial latitude and longitude grid index and a standardized UTC time reference axis, heterogeneous data are mapped to the same spatiotemporal reference system, eliminating geometric distortions and revisit period differences between sensors, and outputting a standardized scale dataset with physically aligned coordinates.

[0043] Based on a standardized scale dataset, the system extracts local spatial heterogeneity feature values ​​by statistically analyzing the sub-pixel features within each grid cell. Specifically, the system reads the statistical moment distribution, variogram parameters, or information entropy index of pixels within the grid to quantify the spatial distribution complexity of land cover types. Subsequently, the processor inputs the extracted local spatial heterogeneity feature values ​​into a pre-stored heterogeneous scale mapping model in real time. This model dynamically calls its internally stored mapping weight operators based on the strength of heterogeneity to calculate and generate a scale response function that matches the current pixel environment. This scale response function serves as the transformation kernel for subsequent nonlinear transformations, carrying the contribution rate of the pixel's internal structure to the macroscopic observation values.

[0044] The system performs a nonlinear scale-up mapping on the standardized scale dataset using the generated scale response function. Through convolution operations or weighted integral operators, the system aggregates the high spatial resolution raw observation information into low-resolution features. During this process, the scale response function acts as a compensation factor to correct for scaling bias caused by nonlinear effects pixel-by-pixel. After this step, the system generates an intermediate feature stream with consistent physical dimensions and eliminated geometric uncertainties, providing structured input variables for subsequent physical constraint calculations.

[0045] The system constructs ecological process consistency constraint equations by calling a pre-defined ecological physical conservation law logic library. The system substitutes surface physical quantities (such as vegetation index and surface temperature) from the intermediate-state characteristic flow into the equations to simulate ecological physical processes such as energy balance or carbon cycle. A coupled objective function is defined through the logic operation unit. This function consists of two parts at the machine level: first, an observation data deviation term characterizing the degree of deviation between the intermediate-state characteristic flow and the original standardized observation data; and second, a physical conservation residual term characterizing whether the calculation results conform to physical conservation laws (such as energy non-closure). This coupled objective function provides a quantitative error discrimination criterion for subsequent feedback correction.

[0046] The system optimizes by minimizing the coupling objective function and uses a numerical iterative algorithm engine to perform iterative optimization calculations. In each iteration, the system outputs a collaborative correction factor λ based on the current gradient descent direction or heuristic search path. This factor λ is stored at the machine level as a tensor with the same dimension as the intermediate feature flow space. The processor uses this factor λ to adaptively weight and correct the intermediate feature flow, generating a corrected intermediate feature flow through a local residual compensation mechanism. The corrected feature flow is fed back in real-time and re-substituted into the ecological process consistency constraint equation for iterative calculation. The system considers the system to be in a convergent state and freezes the current parameters when the rate of change of the coupling objective function falls below a preset convergence threshold or the number of iterations reaches the hardware-set upper limit.

[0047] During convergence, the system performs global consistency smoothing and feature enhancement on the corrected intermediate feature stream based on the collaborative correction factor λ. The processor converts the fused high-precision ecological parameter data into a standard remote sensing data format (such as GeoTIFF or HDF5) through formatting and encapsulation instructions, outputting the final target ecological parameter product.

[0048] By compensating for biases in the nonlinear push-up process using a scale response function and performing closed-loop iterative corrections in conjunction with ecological process consistency constraint equations, this application addresses the information distortion problem caused by scale effects in multi-source sensor data fusion at the machine implementation level. This application minimizes the coupling between observational data bias and physical conservation residuals, improving the physical and logical rigor and numerical accuracy of ecological parameter products in spatially heterogeneous regions. Through deep coupling of hardware parallel processing and algorithmic iterative feedback, the response cycle for cross-scale data processing is shortened, ensuring that the output products possess high spatiotemporal consistency and representativeness of ecological processes.

[0049] This application further proposes that, in the aforementioned scale-transformation-based remote sensing ecological parameter fusion method, extracting the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset is a logical prerequisite for achieving nonlinear scale-up mapping. Specifically, the system retrieves high spatial resolution image data from the standardized scale dataset through the data sensing and alignment module. And low-resolution pixel grid data. In one embodiment, the high spatial resolution image data It can be derived from raw band data from domestically produced Gaofen series satellites or Sentinel-2 satellite. The processor traverses the image data using a spatial topology mapping operator. Based on a preset spatial resampling benchmark, the sub-pixel set within the corresponding low-to-medium resolution pixel grid data (such as MODIS or China Central Satellite Imagery) is precisely located, and the pixel value of each sub-pixel within that set is obtained one by one. .

[0050] To quantify the structural complexity within a pixel, the system calculates pixel values ​​using a sliding window protocol. Local variance within the preset moving window Preferably, the size of the movable window can be preset according to the degree of fragmentation of the land cover type. to For example, in the forest edge region, the window size is selected as a pixel. The pixel size is optimized to balance detail capture and computational efficiency. Simultaneously, the system utilizes the Gray-Level Co-occurrence Matrix (GLCM) algorithm to extract texture contrast features. This is used to characterize the spatial arrangement logic and contrast differences among sub-pixels. Essentially, this process involves refined image recognition and pattern classification of high-resolution sub-pixel sets: local variance is used to identify the boundaries between homogeneous, flat regions and heterogeneous, fragmented regions, while texture contrast is used to identify the spatial arrangement patterns of different land cover types such as vegetation patches, bare soil surfaces, and water bodies. Through this image recognition logic, the system accurately locates key spatial positions requiring nonlinear scale compensation, avoiding the misidentification problem of traditional methods that "generalize from limited evidence" in heterogeneous regions.

[0051] The processor then performs a linear or non-linear weighted mapping on the local variance and texture contrast features, outputting the final local spatial heterogeneity feature values. The calculation formula is shown in the following example:

[0052] ;

[0053] in, and This is a weighting coefficient, typically ranging from 0.3 to 0.7. ;

[0054] and These are the global maxima of the local variance and the global maxima of the texture contrast features within the current processing region, used to perform normalization processing.

[0055] The obtained local spatial heterogeneity eigenvalues The value of is limited to the interval [0,1]. The closer the value is to 1, the stronger the spatial heterogeneity within the grid cell.

[0056] In the construction and pre-training phase of the heterogeneous scale mapping model, the system employs a massive number of high- and low-resolution sample pairs for supervised learning. The training dataset typically contains standardized remote sensing patches from different geographical regions and phenological periods, with a preferred sample size of at least 50,000 pixels. Training conditions are set as follows: the Adam optimizer is used, the initial learning rate is set between 0.0001 and 0.001, the batch size is 64, the training iterations are set to 200 to 500 epochs, and a Dropout mechanism (probability 0.2) is introduced to prevent overfitting. By executing the backpropagation algorithm on high-performance computing units (such as computing nodes equipped with NVIDIA RTX 4090 processors), the model can accurately fit the local spatial heterogeneity feature values. The nonlinear mapping relationship between the parameters and the scaling response function.

[0057] One specific implementation of the heterogeneity scaling model is a fully connected neural network with one hidden layer. The input layer of this network has one neuron, corresponding to the normalized local spatial heterogeneity feature value H; the output layer has one neuron, corresponding to the model curvature operator required for the scaling response function. The estimated value is [value missing]. The number of neurons in the hidden layer can be selected between 8 and 32, preferably 16, and the ReLU activation function is used. Those skilled in the art should understand that the above neural network structure is only an example and can be replaced by other regression models, such as random forest or support vector machine.

[0058] By introducing dual feature constraints of local variance and texture contrast, this application achieves an accurate characterization of the spatial heterogeneity of complex surface landscapes, providing highly reliable prior weights for subsequent nonlinear scale-up. This not only enhances the system's ability to perceive hidden features in complex geographical environments but also strongly supports the accuracy of scale bias correction, ensuring the dual consistency of the generated intermediate feature stream in both spatial detail and physical logic, and significantly reducing systematic inversion errors caused by scale effects.

[0059] This application further proposes that, based on the above-mentioned extraction of local spatial heterogeneity feature values, this application further realizes the cross-dimensional transformation from spatial statistical features to physical response functions by establishing a mathematical analytical model. Specifically, the construction logic of the scale response function strictly follows the physical constraints of the second-order Taylor expansion, in order to quantitatively compensate for the nonlinear scale bias caused by surface inhomogeneity.

[0060] In one embodiment, the scale response function is defined in functional form as follows:

[0061] ;

[0062] in, These are low-resolution observations in a standardized scale dataset, representing the output of physical quantities at the macroscopic pixel scale.

[0063] It is a pre-defined scale transformation fundamental function, usually defined by a physical inversion model (such as a radiative transfer model or LAI inversion operator);

[0064] The mathematical expectation (mean) of high spatial resolution image data within the corresponding grid represents the fundamental contribution under an ideal homogeneous state.

[0065] It is defined as a predefined model curvature operator, whose physical significance lies in characterizing the fundamental functions. The degree of nonlinear sensitivity to changes in surface parameters.

[0066] To achieve precise coupling between the scale response function and the complexity of the surface, the system executes dynamic matching logic, that is, based on the real-time extracted local spatial heterogeneity feature values... The numerical value is dynamically adjusted by the model curvature operator according to a preset mapping table. The value of .

[0067] The pre-defined mapping relationship is mathematically defined as a function. This makes the model curvature operator The eigenvalue H of local spatial heterogeneity satisfies the following relationship:

[0068] ;

[0069] in, It is a parameterized sigmoid function. The sigmoid function was chosen because it can capture the heterogeneous eigenvalues ​​of the input local space. (The range is typically in the range of [0,1]) Smoothly mapped to a bounded output range, which conforms to the physical intuition that curvature changes should be gentle and finite;

[0070] k, a, b, c are the training parameters of the model. k controls the magnitude of curvature change, a and b control the slope and offset of the S-shaped function, and c is the curvature reference value.

[0071] Dynamic adjustment refers to the model adjusting its behavior according to different H values ​​via a function. Calculate the corresponding The value of .

[0072] Preferably, in a typical vegetation cover inversion scenario, the system's preset mapping relationship can be defined as a piecewise linear function or an exponential decay function. For example, when the local spatial heterogeneity eigenvalues When the eigenvalues ​​are in the low heterogeneity interval [0, 0.2], the curvature operator takes a small value to maintain a linear approximation; while when the eigenvalues ​​are in the local spatial heterogeneity interval... When the value exceeds a preset threshold of 0.6 (e.g., in karst landforms or urban-rural transition zones), the system automatically increases the weight coefficient of the curvature operator to enhance the second-order term's ability to correct nonlinear deviations.

[0073] To further enhance the robustness of the technical solution under extremely complex surface conditions, as a preferred embodiment, the system can call multiple heterogeneous scale mapping models with different structures in parallel (such as linear regression models, radial basis function networks, and the aforementioned fully connected neural networks) to generate corresponding candidate values ​​for the curvature operator $f''k(E[P_H])$. The system then performs weighted fusion of the models based on their historical performance confidence levels under the current local spatial heterogeneity feature value H. The formula for calculating the fused curvature operator is as follows:

[0074] ;

[0075] Where K is the total number of models participating in the fusion;

[0076] Let the fusion weights of the k-th model satisfy the following condition: Fusion weights The method for determining the fusion weight is as follows: based on verified high-reliability reference data (such as ground-based measured points or high-confidence historical inversion results) within the current pixel neighborhood, the prediction error of each model within that neighborhood is calculated, and the smaller the error, the higher the fusion weight is assigned. This model fusion mechanism effectively avoids the systematic bias that may occur in a single model under specific geographical conditions, and enhances the adaptability and stability of the scale response function in the nonlinear mapping process.

[0077] By introducing a second-order Taylor expansion term, this application transforms the scale bias, which is difficult to observe directly, into a product term of a computable curvature operator and heterogeneous eigenvalues. This analytical modeling logic effectively breaks the deadlock of poor adaptability of traditional empirical models to complex surfaces, enabling the system to generate customized response functions for each observation grid. By dynamically adjusting the curvature operator, this invention significantly improves the system's accuracy in capturing extreme nonlinear inversion processes, ensuring that the intermediate feature flow retains the true physical and statistical properties of the surface to the maximum extent during the conversion from high resolution to low resolution, thus providing mathematical support for the high fidelity of subsequent ecological parameter fusion.

[0078] This application further proposes that, based on the analytical mapping of the aforementioned scale response function, it further realizes the transformation from simple data fitting to deep fusion driven by physical mechanisms by introducing hard constraints of physical laws. Specifically, the ecological process consistency constraint equation constructed by the system aims to ensure that the fused parameters still follow the basic geophysical conservation laws on a macroscopic scale. In one embodiment, the constraint equation includes at least a surface energy balance (SEB) constraint term, the construction logic of which strictly follows the principle of energy conservation, that is, the net radiant energy received by the surface should maintain a dynamic balance with the heat flux terms consumed.

[0079] The system first obtains the net radiation from the standardized scale dataset through the data sensing interface. With surface temperature Data. Subsequently, the processor invokes a preset evapotranspiration physics operator (such as the Penman-Monteith formula or a single-layer source model), based on the leaf area index in the intermediate-state feature flow. (Leaf Area Index) and current meteorological driving factors are used to estimate latent heat flux in real time. Simultaneously, the system combines surface parameters (cover type, soil moisture, etc.) from the standardized scale dataset and utilizes vegetation structure information provided by intermediate state feature flow to calculate soil heat flux. With sensible heat flux .

[0080] To quantify the physical and logical deviations of various ecological parameters, the system constructs a residual function of the following form:

[0081] ;

[0082] In this physical model, the units of measurement for each flux term are uniformly set to [unit not specified]. Ideally, if the intermediate feature stream generated by the fusion perfectly conforms to physical laws, then the residual... It should approach 0. However, due to the nonlinear error of remote sensing inversion and sensor noise, the residual... It typically exhibits a non-zero fluctuation value. The system will use this residual function The physical residual term, as part of the coupled objective function, is used to constrain the iterative optimization process. Preferably, in practical engineering implementation, the system sets the effective triggering threshold range of the physical residual to be [range missing]. If the calculated residual If the range is exceeded (such as in extremely arid or highly heterogeneous mountainous areas), then the current intermediate characteristic flow is determined to have a significant physical and logical conflict.

[0083] By transforming the surface energy balance equation into a mathematical penalty term, the problem of "inversion parameter non-closure" commonly found in traditional statistical fusion methods is effectively solved. The system can use physical laws to perform "logical auditing" of intermediate characteristic flows. This constraint mechanism, which introduces prior physical knowledge, ensures that the generated ecological parameter products not only numerically approximate the observed values ​​but also possess rigorous scientific validity in the ecological physical evolution process. This underlying mechanism guarantees the practical value of the product in high-reliability application scenarios such as climate simulation and precise water resource management.

[0084] This application further proposes that, within the constraint framework of the aforementioned physical conservation residual term, this application achieves deep coupling between observation data-driven and physical logic constraints by constructing a nonlinear optimization coupling objective function. Specifically, the system-defined coupling objective function... The aim is to obtain the optimal ecological parameter correction operator by minimizing the bias and physical non-closure between multi-source data. The formula for calculating this coupling objective function is defined as follows:

[0085] ;

[0086] in, and These are the observation confidence weight and the physical constraint penalty factor, respectively.

[0087] These are the ecological parameter estimates generated by the forward model based on the intermediate state feature flow at the current time.

[0088] These are the actual observed parameters in the standardized scale dataset;

[0089] It characterizes the fitting accuracy of the fusion result to the original observation data;

[0090] Based on normalized vegetation index (Normalized Difference Vegetation Index) and surface temperature The ecological process consistency residual function constructed by (Land Surface Temperature) is transformed into physical penalty cost in the form of a squared term.

[0091] In one embodiment, the observation confidence weight Physical constraint penalty factor The selection of [the appropriate sensor type] is usually based on sensor noise assessment and uncertainty analysis of the physical model. For example, in regions with high signal-to-noise ratios for satellite sensors, [the appropriate sensor type] is preferred. The value is set between 0.6 and 0.8; however, in areas with dramatic changes in ecological mechanisms (such as arid and semi-arid transition zones), the value is adjusted accordingly. The range of values ​​is up to This is to enhance the effect of physical conservation laws on suppressing data noise.

[0092] In the specific implementation logic, the processor constructs a Lagrange function by introducing Lagrange multipliers, transforming the physically constrained optimization problem into an unconstrained dual problem, and then iteratively solves it using the gradient descent algorithm. Preferably, in a typical computational task, the system sets the initial learning rate to be... The momentum factor is 0.9, and the coupled objective function is calculated in each iteration. Partial derivatives with respect to the intermediate state characteristic flow. Iterative optimization calculates the final output co-correction factor. This collaborative correction factor Stored in tensor form, it quantitatively represents the correction vector of each spatial grid cell under the current physical constraints.

[0093] By transforming the complex problem of ecological parameter fusion into a rigorous mathematical optimization process, this application effectively avoids the subjective bias caused by manual parameter tuning in traditional fusion algorithms. This is achieved by observing the confidence weights. Physical constraint penalty factor The system employs a dynamic weighting mechanism, enabling it to adaptively adjust the contribution weights of "data-driven" and "physical constraints" based on the characteristics of different underlying surfaces. This gradient descent-based closed-loop optimization logic ensures the effectiveness of the collaborative correction factor. It can accurately point to the most reasonable numerical solution space in terms of physical logic, thereby quickly locking the optimal state in complex nonlinear inversion paths, and providing deterministic algorithmic support for the generation of the final high-precision ecological parameter products.

[0094] Furthermore, in the gradient descent optimization process of the aforementioned coupled objective function, the system achieves automated closed-loop control of the iterative process by establishing a precise convergence determination mechanism. Specifically, the core logic of the convergence determination lies in monitoring the collaborative correction factor. Numerical stability under continuous iteration step size ensures that the optimization process can accurately lock onto the optimal solution range that coincides with the physical logic and observation data.

[0095] At the end of each iteration, the processor first calculates the collaborative correction factor generated at the current moment in real time. Compared to the previous moment The Euclidean distance between them. This distance characterizes the displacement magnitude of the model in the parameter space, and its calculation formula is defined as:

[0096] ;

[0097] Where n is the total number of grid cells.

[0098] The system then applied the Euclidean distance With the preset convergence threshold Numerical logic comparisons are performed. Preferably, the convergence threshold is determined based on the sensitivity differences of different ecological parameters (such as LAI and NDVI). The value range is usually set within to Between, preferably For example, when performing high-precision fusion of Leaf Area Index (LAI), the system will use a convergence threshold. Set as This is to ensure that even minor physical and logical deviations can be detected.

[0099] If determining Euclidean distance If the system logic control unit outputs a "convergence" command, it determines that the current intermediate state feature flow has reached an equilibrium state between physical conservation and observation fitting, and immediately triggers the subsequent weighted fusion logic. If the Euclidean distance... Still greater than or equal to the convergence threshold If the current search has not yet reached the extreme point, the system determines that it will continue to execute the next iteration loop, dynamically updating the values ​​of the Lagrange multipliers based on the gradient direction. Furthermore, to prevent oscillations and non-convergence on complex underlying surfaces (such as high-altitude areas with heavy cloud cover), the system also sets a maximum iteration limit. (Usually 50 to 100 times), as a fallback protection against forced convergence.

[0100] By introducing Euclidean distance as a "hard metric" for convergence, this application effectively solves the common problems of "overfitting" or "insufficient search" in nonlinear optimization. Through precise control of the iteration depth, the system can maximize the saving of hardware computing resources and shorten the production cycle while ensuring the physical consistency of ecological parameters. This adaptive termination mechanism based on numerical stability ensures that the final output collaborative correction factor has high reliability and spatial smoothness, providing a robust logical gate for generating remote sensing ecological products with operational value and avoiding systematic deviations caused by manual intervention in algorithm termination.

[0101] This application further proposes that, in determining the collaborative correction factor... After reaching convergence, the system performs nonlinear compensation for the intermediate state feature flow through refined matching of spatiotemporal features. Specifically, this application utilizes a collaborative correction factor. The adaptive weighted correction aims to address the local feature distortion problem caused by surface fragmentation in remote sensing inversion. The system first retrieves the local spatial heterogeneity feature values ​​extracted in the preceding steps. The two-dimensional spatial distribution is used as a basis to construct a spatial weight matrix that is negatively correlated with it. .

[0102] In one embodiment, the spatial weight matrix The construction logic follows the physical principle that "the higher the heterogeneity of a region, the lower the spatial representativeness of its observations, and the corresponding correction weights should be reduced accordingly." Preferably, the spatial weight matrix... The elements in The negative correlation function can be defined as follows, with the formula as follows:

[0103] ;

[0104] in, The preset sensitivity adjustment coefficient typically ranges from 0.5 to 2.0 (e.g., 1.5 in hilly areas with complex terrain). In a specific embodiment, for the leaf area index (LAI) fusion task in the fragmented landform region of the Loess Plateau, the sensitivity adjustment coefficient is... The preferred value is 1.2.

[0105] The aforementioned negative correlation function ensures the heterogeneity of eigenvalues ​​in the local space. For highly heterogeneous pixels approaching 1 (such as urban-rural building transition zones), the weight w decays rapidly, thereby suppressing noise superposition caused by insufficient spatial representativeness in the region.

[0106] Subsequently, the processor performs high-dimensional tensor operations, including the co-correction factor. With spatial weight matrix The system performs a Hadamard product operation. This product logic ensures precise alignment of the correction operation in the spatial dimension, achieving pixel-by-pixel residual compensation for the intermediate feature stream. Specifically, the system incrementally adjusts the initial inversion values ​​using a weighted correction factor to generate the corrected intermediate feature stream. In the system implementation path, this process involves element-wise transformations of large-scale matrices, typically completed within a single instruction cycle by calling underlying linear algebra libraries (such as high-performance tensor computation engines).

[0107] By introducing a weighting mechanism negatively correlated with spatial heterogeneity, this application effectively overcomes the risk of local distortion caused by the "one-size-fits-all" global correction in traditional fusion methods. This is achieved through a spatial weighting matrix. By dynamically adjusting the correction intensity, the system can perform strong compensation in homogeneous regions to improve consistency, while performing weak compensation in heterogeneous regions to preserve local details. Local compensation achieved through dot product operations ensures high fidelity of the fused product during cross-scale transformations, improves the numerical stability and spatial representativeness of the parametric product under complex underlying surfaces, and provides more realistic and reliable underlying data support for subsequent ecological process simulations.

[0108] After outputting the target ecological parameter products, this application constructs a closed-loop self-calibration logic from "product output" to "model evolution" by introducing a ground truth feedback mechanism. Specifically, the system performs a product consistency verification step, which aims to conduct a final logical review and accuracy traceability of the space-based inversion results using ground-based observation data.

[0109] The system acquires calibration data from ground-based measurement stations deployed within the target area via a remote communication interface or a dedicated data gateway. In one embodiment, this calibration data originates from globally available flux observation towers (such as FLUXNET) or regional automatic weather stations, and its observational indicators cover the measured leaf area index, land surface temperature, and four radiation components. The processor uses spatial interpolation or nearest neighbor matching algorithms to map the discrete station calibration data to the corresponding target ecological parameter product grid cells.

[0110] The processor performs a comparison operation to calculate the absolute deviation between the target ecological parameter product and the calibration data. In one embodiment, the system sets a preset confidence interval. This range is typically set based on industry standards or historical error curves of the sensor. For example, for Leaf Area Index (LAI) products, the confidence interval for its absolute deviation is preferably set as follows: If the calculated absolute deviation exceeds the confidence interval, the system determines that the current fusion model has a systematic drift under a specific land surface type.

[0111] The system defines the deviation feature vector (including deviation magnitude, geographic coordinates, and phenological characteristics) that exceeds the confidence interval as the loss signal. The processor calls a preset parameter optimization algorithm (such as a nonlinear least squares method based on the Gauss-Newton method or a particle swarm optimization algorithm) to adjust the internal weight parameters of the heterogeneous scale mapping model in reverse. In the machine implementation path, the system uses minimizing the loss signal as the coupling objective function and achieves adaptive evolution of the model parameters by iteratively fine-tuning the neuron weights or scale transformation kernels within the model.

[0112] By introducing real-time feedback from ground-based measured data, this application effectively addresses the common problems of "model degradation" and "spatiotemporal limitations" in remote sensing inversion. By converting bias into a loss signal and performing inverse optimization, the system can continuously correct its internal parameters according to changes in different geographical environments, thus enabling the heterogeneous scale mapping model to possess self-learning and incremental update capabilities. This closed-loop verification mechanism ensures that the output ecological parameter products possess long-term temporal stability and extremely high industry calibration accuracy, significantly enhancing the system's inversion robustness in the face of complex climate change and extreme geographical environments.

[0113] This application enables precise fine-tuning of scale mapping parameters using measured true values, significantly reducing the residual bias between the fused product and ground observations. This adaptive verification mechanism ensures that the ecological parameter products maintain extremely high consistency across different phenological periods and geographical zones, improving the operational calibration quality of the products and providing self-healing computing power support for high-precision ecological environment monitoring.

[0114] This application further proposes that, in the acquisition stage of the aforementioned multi-source sensing data, this application further eliminates the interference of complex surface geometry on radiation observations by introducing elevation dimension constraints. Specifically, the multi-source sensing data, in addition to covering optical and microwave bands, also includes digital elevation model (DEM) data of the same region, to support refined terrain correction logic.

[0115] During the spatiotemporal registration process, the processor invokes a preset terrain radiation transfer model and extracts the slope and aspect characteristic parameters of the target grid using digital elevation model (DEM) data. In one embodiment, the system calculates the tilt angle and azimuth orientation of each pixel relative to the horizontal plane by performing second-order difference operations on the DEM raster. Preferably, the processor combines the geometric parameters of the sun-synchronous orbit (such as the solar altitude angle and solar azimuth angle) to construct a pixel-by-pixel cosine correction operator to perform terrain correction on multi-source sensing data.

[0116] Specifically, the machine implementation of terrain correction aims to eliminate variations in effective illumination area caused by slope, as well as shading or backlighting effects caused by aspect. In one embodiment, the system employs a C-Correction (C-Correction) or SCS+C model, incorporating empirical terrain coefficients to correct for non-uniform irradiance biases caused by mountainous terrain. For example, in mountainous areas with slopes greater than 15°, the system can automatically identify and compensate for weak signals on shaded slopes, ensuring comparable reflectance characteristics for similar features in different terrain locations within a standardized scale dataset.

[0117] By incorporating 3D terrain features into the spatiotemporal registration process, this application effectively solves the "same object, different spectrum" phenomenon existing in traditional planar registration methods. Through DEM-assisted irradiance correction, the system can remove terrain-induced spurious heterogeneity from the underlying raw observations, thereby extracting local spatial heterogeneity feature values. It can more purely reflect the structural characteristics of the land cover itself. This preprocessing mechanism with terrain perception capabilities provides a cleaner data foundation for subsequent high-precision scale transformation, and improves the robustness of the ecological parameter fusion model in complex geographical units such as hills and mountains.

[0118] This invention achieves a leap from "horizontal projection" to "stereoscopic correction" of remote sensing data by introducing terrain correction technology assisted by digital elevation models. It effectively eliminates irradiance bias caused by complex terrain and unifies the physical benchmark of multi-source data in mountainous environments. This terrain-aware preprocessing logic significantly enhances the physical consistency of standardized scale datasets and reduces computational distortion of subsequent scale response functions in terrain-undulating regions, laying a solid geometric and physical foundation for generating high-precision ecological parameter products with terrain robustness.

[0119] The following is a specific example of a remote sensing ecological parameter fusion method based on scale transformation:

[0120] Select the core treatment area for karst rocky desertification in the Yunnan-Guizhou Plateau (1200 Using the vegetation growing season (July 2025) as an application scenario, the system first acquires multi-source sensing data within the study area: 1-meter resolution Gaofen-7 optical imagery, 300-meter resolution Sentinel-3 LAI inversion product, 30-meter resolution ASTER DEM data, and synchronous measured data from 12 ground flux observation towers. Geometric fine correction, resampling, and C-model topographic correction are performed on the multi-source data to eliminate irradiance bias in areas with slopes greater than 15°, outputting a 30-meter resolution standardized scale dataset under a unified UTM projection.

[0121] Traverse the set of sub-pixels within a 30-meter grid corresponding to the Gaofen-7 image in the standardized dataset, and calculate the local variance of sub-pixel NDVI under a 3×3 moving window. Texture contrast features T_c=0.61 were extracted using the gray-level co-occurrence matrix, and normalized weighted mapping was performed with weights ω1=0.4 and ω2=0.6 to obtain local spatial heterogeneity feature values. This indicates strong heterogeneity in the surface cover of this karst peak-cluster-valley area. The H value is input into a pre-trained heterogeneous scale mapping model, and dynamic matching yields a model curvature operator f''=0.82. A scale response function is constructed based on a second-order Taylor expansion. Substitute the high-resolution LAI mean The low-resolution mapping value after bias correction is calculated. The intermediate-state feature flow was obtained by nonlinear scaling up, and the initial scale bias was reduced from 18.2% in the traditional linear aggregation method to 6.7%.

[0122] Construct ecological process consistency constraint equations, and input the net radiation from the standardized dataset. The surface temperature LST is 29.6℃, and the latent heat flux LE is estimated to be 262 based on the intermediate state LAI. The soil heat flux G=41 was calculated based on the surface cover parameters. sensible heat flux The physical conservation residual term is obtained. Define the coupling objective function. Set the observation confidence weight α=0.6 and the physical constraint penalty factor β=0.4, and substitute them into the initial fusion estimate. Observed values The initial objective function value is calculated. .

[0123] The gradient descent iteration is initiated with the objective of minimizing the coupling objective function. After 7 iterations, the Euclidean distance between two adjacent outputs of the co-correction factor λ is... Less than the preset convergence threshold The convergence state was determined, and the post-convergence collaborative correction factor λ = 0.18 was obtained. A negative correlation spatial weight matrix was constructed based on the heterogeneity eigenvalues ​​H, and the sensitivity adjustment coefficient γ = 1.2 was used to calculate the corresponding grid weights. After multiplying λ with the weight matrix, pixel-by-pixel residual compensation is performed on the intermediate state feature flow. After weighted fusion, the final 30-meter resolution LAI product is output. Verified by ground measurement points, the fusion result effectively eliminates the numerical distortion phenomenon common in traditional methods in karst peak-cluster valley areas, improves the parameter inversion accuracy of vegetation cover edge areas, and the output LAI product is strictly self-consistent with the surface energy balance equation in terms of physical logic. It can be directly used as input parameters for regional carbon sink estimation and hydrological process simulation without additional consistency correction.

[0124] This embodiment fully demonstrates the application process of the method in complex karst landform areas: nonlinear scale bias is corrected by a heterogeneity-driven dynamic scale response function, and iterative optimization is achieved by combining the physical constraints of ecological processes. The final output LAI product not only retains the spatial detail features of high-resolution images, but also fully conforms to the physical laws of surface energy balance. It can accurately identify the differences in vegetation restoration of different rocky desertification control models at the small watershed scale. Compared with traditional fusion products, it significantly reduces the numerical distortion of fragmented landform areas such as peak clusters and valleys. It can be directly used as input parameters for regional carbon sink estimation and hydrological process simulation without additional consistency correction, effectively supporting the dynamic evaluation of rocky desertification control effectiveness and the precise implementation of ecological restoration measures.

[0125] Example 2:

[0126] like Figure 3 As shown, the scale-transformation-based remote sensing ecological parameter fusion system, using the aforementioned scale-transformation-based remote sensing ecological parameter fusion method, includes:

[0127] The data perception and alignment module is used to acquire multi-source sensing data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset.

[0128] The scale mapping module is used to extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into the preset heterogeneous scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values.

[0129] The compensation processing module is used to perform nonlinear scale up-mapping mapping on the standardized scale dataset using the scale response function to obtain the intermediate feature flow corrected for scale bias.

[0130] The module is used to construct the ecological process consistency constraint equation. The intermediate state feature flow is substituted into the ecological process consistency constraint equation as an input variable, and a coupled objective function containing the observation data bias term and the physical conservation residual term is defined.

[0131] The constraint iteration module is used to perform iterative optimization calculations with the objective function of minimizing coupling, and outputs a collaborative correction factor in each iteration. And utilize the collaborative correction factor Adaptive weighted correction is applied to the intermediate feature flow to generate a corrected intermediate feature flow; the corrected intermediate feature flow is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next iteration until the coupled objective function converges.

[0132] The product generation module is used to combine the corrected intermediate-state feature stream obtained in the convergence state with the co-correction factor output from the last iteration. The weighted fusion process is performed to output the target ecological parameter product.

[0133] Specifically, the data sensing and alignment module consists of a distributed satellite receiving terminal and a high-speed I / O interface, responsible for real-time retrieval of multi-source sensing data (such as Landsat-8, MODIS, DEM, etc.) of the target area. On the hardware side, this module utilizes a multi-channel fiber optic array to achieve concurrent throughput of Tb-level data and calls a parallel processing engine to perform automatic orthorectification based on feature point matching, outputting a standardized scale dataset in a unified projected coordinate system.

[0134] The scale mapping module is deployed in the central processing unit (CPU) of the high-performance computing node. It extracts the local spatial heterogeneity feature values ​​of each grid cell by executing a spatial autocorrelation algorithm. This module is connected to a pre-set heterogeneous scale mapping model storage unit via a bus. Based on the heterogeneity features calculated in real time, it dynamically retrieves and generates the scale response function from a fixed lookup table, thereby decoupling the software-defined parameters from the hardware computing logic.

[0135] The core of the compensation processing module consists of a graphics processing unit (GPU) array, which utilizes the GPU's CUDA Cores to perform nonlinear scale-up mapping on the entire field mesh. This module implements pixel-by-pixel convolution operations of high-order Taylor series through hardware acceleration, transforming the raw observation data into an intermediate feature stream corrected for scale bias.

[0136] The construction module includes a physical rule logic library, which defines constraint equations such as surface energy balance through an instruction set. This module loads the intermediate state feature stream into a high-speed cache (L3 cache) and constructs a coupled objective function containing observation bias terms and physical residual terms online according to preset physical conservation laws.

[0137] The constraint iteration module, serving as the algorithmic core of the system, performs gradient descent optimization using a dedicated digital signal processor (DSP) or AI operator accelerator card. This module iteratively calculates the cooperative correction factor through an iterative logic control unit. The intermediate state feature flow is adaptively weighted and corrected. The iterative process is monitored at the hardware level by a convergence determination circuit until the Euclidean distance reaches a preset threshold.

[0138] The product generation module is responsible for the final feature fusion and file packaging. On the hardware side, a large-scale storage array (RAID 6) is used for multi-threaded writing. The corrected feature stream and correction factor are bit-depth aligned and weighted to output the target ecological parameter product that conforms to international remote sensing format standards.

[0139] Taking the leaf area index (LAI) fusion of a fragmented landform area in the Loess Plateau as an example: The system first acquires 10-meter resolution Gaofen-6 imagery and 500-meter resolution MODIS data through the data perception and alignment module. The scale mapping module identifies the area as having crisscrossing gullies and extremely uneven vegetation distribution, and calculates the spatial heterogeneity characteristic values. The value was 0.72. Subsequently, the compensation processing module called the quadratic curvature operator to generate a response function, performing a nonlinear up-propagation on the MODIS data. The construction module, based on the energy balance equation, discovered a correlation between the estimated latent heat flux and net radiation in this region. The physical residuals. The constraint iteration module then starts, and after 12 gradient descent iterations, the collaborative correction factor is determined. exist Convergence with improved accuracy. Ultimately, the product generation module outputs a high-quality 30-meter resolution LAI dataset that eliminates scale-effects and conforms to the law of energy conservation.

[0140] This system achieves high-precision correction of remote sensing ecological parameters during cross-scale transformation by deeply integrating hardware-accelerated computing power with physical constraint algorithms. The system effectively eliminates systematic inversion errors caused by the heterogeneity of the Earth's surface. By introducing hardware-level convergence monitoring and physical residual auditing, the system not only improves the processing efficiency of heterogeneous data fusion but also ensures the physical and logical rigor of the output products under complex terrain conditions.

[0141] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for fusing remote sensing ecological parameters based on scale transformation, characterized in that: Includes the following steps: Acquire multi-source sensor data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset; Extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into a preset heterogeneity scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values; The scale response function is used to perform a nonlinear scale up-mapping mapping on the standardized scale dataset to obtain an intermediate feature stream corrected for scale bias. An ecological process consistency constraint equation is constructed, and the intermediate state characteristic flow is substituted into the ecological process consistency constraint equation as an input variable. A coupled objective function containing observation data bias term and physical conservation residual term is defined. The iterative optimization calculation aims to minimize the coupling objective function, and outputs a collaborative correction factor in each iteration. And utilize the aforementioned collaborative correction factor The intermediate feature stream is adaptively weighted and corrected to generate a corrected intermediate feature stream; the corrected intermediate feature stream is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next round of iteration until the coupled objective function converges. The corrected intermediate-state feature stream obtained in the convergence state is combined with the co-correction factor output from the last iteration. The target ecological parameter product is output by performing weighted fusion processing.

2. The remote sensing ecological parameter fusion method based on scale transformation according to claim 1, characterized in that: Extracting the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, including: Obtain the normalized scale dataset, wherein the normalized scale dataset includes high spatial resolution image data. and low-resolution pixel grid data; Traversing the high spatial resolution image data Locate the set of sub-pixels within the corresponding low-to-medium resolution pixel grid data, and obtain the pixel value of each sub-pixel in the set of sub-pixels. ; Calculate the pixel value Local variance within the preset moving window And use the gray-level co-occurrence matrix to extract texture contrast features. ; Regarding the local variance With the texture contrast features Perform a weighted mapping and output the local spatial heterogeneity feature value H.

3. The remote sensing ecological parameter fusion method based on scale transformation according to claim 1, characterized in that: The scaling response function is constructed based on a second-order Taylor series expansion, and its functional form is defined as follows: ； in, These are the low-resolution observations in the standardized scale dataset; It is a fundamental function for scale transformation. The mean value of high spatial resolution image data; The preset model curvature operator; Generating a scale response function that dynamically matches the local spatial heterogeneity eigenvalues ​​means, based on the local spatial heterogeneity eigenvalues... The numerical value is dynamically adjusted according to a preset mapping relationship by the curvature operator of the model. The value of .

4. The method for fusing remote sensing ecological parameters based on scale transformation according to claim 1, characterized in that: The ecological process consistency constraint equation includes at least a surface energy balance constraint term, which is constructed through the following logic: Obtain the surface parameters and net radiation from the standardized scale dataset. and ground temperature data; Based on the leaf area index in the intermediate characteristic flow Estimating latent heat flux ; Soil heat flux is estimated based on the surface parameters and the intermediate characteristic flow. and sensible heat flux ; Constructing the residual function The residual function This is the physical conservation residual term.

5. The method for fusing remote sensing ecological parameters based on scale transformation according to claim 1, characterized in that: The formula for calculating the coupling objective function is as follows: ； in, and These are the observation confidence weight and the physical constraint penalty factor, respectively. These are the estimated ecological parameters generated based on the current intermediate state feature flow. These are the observation parameters in the standardized scale dataset; Based on normalized vegetation index Leaf area index and surface temperature Constructed ecological process consistency residual function; The iterative optimization calculation introduces Lagrange multipliers to the coupled objective function. Perform gradient descent to solve for the cooperative correction factor and output it. .

6. The method for fusing remote sensing ecological parameters based on scale transformation according to claim 1, characterized in that: The convergence criterion for the coupled objective function is as follows: Calculate the cooperative correction factor generated in two consecutive iterations. The Euclidean distance; If the Euclidean distance is less than the preset convergence threshold ϵ, or the number of iterations reaches the preset maximum number of iterations threshold, then the convergence state is determined and the weighted fusion logic is triggered; if neither condition is met, then the iteration loop continues.

7. The remote sensing ecological parameter fusion method based on scale transformation according to claim 1, characterized in that: Using the aforementioned collaborative correction factor The intermediate-state feature stream is adaptively weighted and corrected, and the specific steps include: Based on the local spatial heterogeneity characteristic value The spatial distribution of the local spatial heterogeneity feature values ​​is constructed. The spatial weight matrix W of negative correlation, where regions with higher heterogeneity have lower weights; The collaborative correction factor A dot product operation is performed with the spatial weight matrix W to perform local residual compensation on the intermediate feature flow, generating a corrected intermediate feature flow.

8. The method for fusing remote sensing ecological parameters based on scale transformation according to claim 1, characterized in that: Following the output of the target ecological parameter product, a product consistency verification step is also included, including: Acquire calibration data from ground-based measuring stations within the target area; Compare the absolute deviations between the target ecological parameter product and the calibration data; If the absolute deviation exceeds a preset confidence interval, the deviation feature vector is used as a loss signal, and the internal parameters of the heterogeneous scaling model are adjusted by a preset parameter optimization algorithm to minimize the loss signal.

9. The method for fusing remote sensing ecological parameters based on scale transformation according to claim 1, characterized in that: The multi-source sensing data also includes digital elevation model data, and the spatiotemporal registration step further includes: The digital elevation model data is used to perform terrain correction on the multi-source sensing data to eliminate irradiance deviations caused by slope and aspect.

10. A remote sensing ecological parameter fusion system based on scale transformation, characterized in that: The method for fusing remote sensing ecological parameters based on scale transformation as described in any one of claims 1 to 9 includes: The data perception and alignment module is used to acquire multi-source sensing data within the target area and perform spatiotemporal registration, outputting a standardized scale dataset. The scale mapping module is used to extract the local spatial heterogeneity feature values ​​of each grid cell in the standardized scale dataset, and input the local spatial heterogeneity feature values ​​into a preset heterogeneous scale mapping model to generate a scale response function that dynamically matches the local spatial heterogeneity feature values. The compensation processing module is used to perform a nonlinear scale up-mapping mapping on the standardized scale dataset using the scale response function to obtain an intermediate feature stream corrected for scale bias. The module is used to construct the ecological process consistency constraint equation, and the intermediate state feature flow is substituted into the ecological process consistency constraint equation as an input variable. A coupled objective function containing observation data bias term and physical conservation residual term is defined. The constraint iteration module is used to perform iterative optimization calculations with the goal of minimizing the coupling objective function, and outputs a collaborative correction factor in each iteration. and utilize the aforementioned collaborative correction factor The intermediate feature stream is adaptively weighted and corrected to generate a corrected intermediate feature stream; the corrected intermediate feature stream is then used as an updated input variable and substituted back into the ecological process consistency constraint equation for the next iteration until the coupled objective function converges. The product generation module is used to combine the corrected intermediate-state feature stream obtained in the convergence state with the co-correction factor output from the last iteration. The weighted fusion process is performed to output the target ecological parameter product.

Images