Machine learning-based inVEST carbon stock confidence rendering method and system

Abstract

This invention discloses a machine learning-based InVEST carbon storage confidence rendering method and system, belonging to the field of map rendering technology. This method obtains the set of vegetation categories predicted by multiple machine learning classifiers for the target pixel coordinates in a thematic map of target carbon storage, calculates the category voting frequency and classification information entropy; normalizes the classification information entropy of all pixels in the region to construct a divergence sub-texture; obtains the baseline carbon storage color value and baseline opacity pre-calculated based on the InVEST model, and uses the extracted classification information entropy to negatively attenuate the baseline opacity in the fragment shader to generate the target opacity; concatenates the baseline carbon storage color value and the target opacity to construct the RGBA vector of the target pixel, generating the target rendering matrix to update the thematic map. This invention maps multi-model classification divergence to visual transparency attenuation, achieving high-fidelity coupled rendering of carbon storage values ​​and underlying confidence levels within the same map.

Description

Technical Field

[0001] This invention relates to the field of map rendering technology, and more specifically, this application relates to an InVEST carbon storage confidence rendering method and system based on machine learning. Background Technology

[0002] In the visualization application scenarios of digital platforms for ecological restoration and carbon sequestration in mining areas, land cover types are typically preprocessed using multi-source spatial classification models. The calculation model then outputs absolute values, which are then used for thematic map coloring in the front-end system. Existing spatial rendering systems generally adopt a unidirectional linear pipeline architecture, focusing on directly converting the final calculation results into deterministic visual layers. However, this conventional rendering mechanism suffers from a serious "deterministic assumption" flaw, completely severing the causal relationship between the objective differences within the underlying data production process and the final front-end graphic presentation pipeline.

[0003] In real and complex surface environments such as mining areas, numerous ecological transition zone boundaries with ambiguous characteristics objectively exist. Due to the complexity of this objective spatial environment, the underlying multi-source spatial classification model will inevitably produce rigid discrepancies in classification results in the transition zone region when extracting the preceding spatial parameters. However, existing conventional technologies, when faced with such precise working conditions, forcibly adhere to the rigid assumption that the data inputs are completely deterministic and mutually independent. The existing rendering pipeline mechanism unilaterally blocks the transmission of this key spatial feature of multi-model classification discrepancies from the underlying graphics pipeline, merely treating the passively received absolute measurement values ​​as error-free "absolute truth values," and performing uniform, high-saturation absolute texture shading accordingly.

[0004] To address the aforementioned issues, this field needs to resolve the visualization distortion problem caused by the inability of existing rendering pipelines to handle and represent the rigid discrepancies in the underlying model under complex transitional conditions in mining areas. Because the existing front-end graphics architecture lacks a cross-domain coupling mechanism to map the rigid contradictions of the underlying discrete classification to graphical visual attributes in a dimensionality-reducing manner, regions containing severe underlying classification errors are forcibly assigned visual rendering weights completely equivalent to high-confidence regions. This single-line, black-box approach directly results in the objectively existing classification discrepancies being completely smoothed out by the front-end's determined smooth coloring, generating a visual "pseudo-high confidence" mapping illusion in the rendered view that masks the actual calculation errors, thus causing severe distortion in platform data representation.

[0005] In summary, existing technologies, when used for spatial visualization in complex surface transition zones of mining areas, forcibly block the transmission of rigid divergence features of the underlying multi-model spatial classification to the front-end graphics pipeline. This results in conventional deterministic rendering mechanisms being unable to couple the measured values ​​with the underlying confidence level within a single view, leading to a technical problem where the visual presentation severely masks the underlying calculation errors. Summary of the Invention

[0006] To address the aforementioned technical issues, this paper provides a machine learning-based InVEST carbon storage confidence rendering method and system, which solves the problems mentioned in the background section.

[0007] In a first aspect, embodiments of this application provide an InVEST carbon storage confidence rendering method based on machine learning, comprising the following steps: obtaining a set of vegetation categories for target pixel coordinates of a target carbon storage thematic map, wherein the set of vegetation categories represents the set of vegetation categories predicted and output by a preset number of independent machine learning classifiers; counting the number of each vegetation category in the set of vegetation categories and dividing it by the total number of vegetation categories in the set of vegetation categories to obtain the category voting frequency of each vegetation category; calculating the classification information entropy of the target pixel coordinates by performing information entropy calculation on the category voting frequency, and normalizing the classification information entropy corresponding to all pixel coordinates within a preset rendering area of ​​the target carbon storage thematic map to construct a two-dimensional single-channel grayscale moment. A bifurcation subtexture in array form is generated; the baseline carbon storage color value and baseline opacity corresponding to the target pixel coordinates are obtained based on the InVEST model and pre-calculated; the classification information entropy corresponding to the target pixel coordinates in the bifurcation subtexture is extracted, and the classification information entropy is negatively attenuated to the baseline opacity in the fragment shader of the graphics rendering pipeline to generate the target opacity; the baseline carbon storage color value is used as RGB channel data, and the target opacity is used as Alpha channel data, and they are concatenated in the fragment shader to construct the target pixel RGBA vector; the target rendering matrix is ​​generated based on the target pixel RGBA vectors corresponding to all pixel coordinates in the preset rendering area, and the target carbon storage thematic map is rendered and updated accordingly.

[0008] Secondly, embodiments of this application provide an InVEST carbon storage confidence rendering system based on machine learning, comprising: a vegetation category set acquisition module: used to acquire the vegetation category set of the target pixel coordinates of the target carbon storage thematic map, wherein the vegetation category set represents the set of vegetation categories predicted and output by a preset number of independent machine learning classifiers; a category voting frequency processing module: used to count the number of each vegetation category in the vegetation category set and divide it by the total number of vegetation categories in the vegetation category set to obtain the category voting frequency of each vegetation category; and a divergence subtexture processing module: used to calculate the classification information entropy of the target pixel coordinates by performing information entropy calculation on the category voting frequency, and to normalize the classification information entropy corresponding to all pixel coordinates in the preset rendering area of ​​the target carbon storage thematic map to construct a divergence subtexture in the form of a two-dimensional single-channel grayscale matrix. The system comprises the following modules: a baseline acquisition module (for obtaining baseline carbon storage color values ​​and baseline opacity corresponding to target pixel coordinates pre-calculated based on the InVEST model), a target opacity processing module (for extracting classification information entropy corresponding to target pixel coordinates in the divergence subtexture and negatively attenuating the classification information entropy to the baseline opacity in the fragment shader of the graphics rendering pipeline), a target pixel RGBA vector processing module (for concatenating the baseline carbon storage color values ​​as RGB channel data and the target opacity as Alpha channel data in the fragment shader to construct the target pixel RGBA vector), and a target carbon storage thematic map rendering module (for generating a target rendering matrix based on the target pixel RGBA vectors corresponding to all pixel coordinates within a preset rendering area and updating the target carbon storage thematic map accordingly).

[0009] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned machine learning-based InVEST carbon storage confidence rendering method.

[0010] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0011] 1. This solution quantifies the objective, rigid divergence generated by the underlying multi-source classification model in complex ecological transition zones by acquiring the set of vegetation categories predicted by independent machine learning classifiers and accurately calculating the category voting frequency and classification information entropy. This approach identifies and preserves the uncertain data in the calculation process from the source, avoiding the underlying logical flaw of existing technologies that treat erroneous prediction results as absolute truth values, thus providing reliable confidence data support for subsequent scientific visualization.

[0012] 2. This solution innovatively constructs the normalized classification information entropy into a two-dimensional single-channel grayscale matrix form of a divergence sub-texture. In the fragment shader of the graphics rendering pipeline, the classification information entropy is directly used to perform negative attenuation calculation on the baseline opacity to generate the target opacity. This mechanism directly transforms the abstract algorithmic divergence into an intuitively perceptible Alpha channel transparency attribute, making areas with greater underlying divergence appear more transparent, effectively breaking the misleading visual illusion caused by existing front-end smoothing coloring.

[0013] 3. By using the absolute index of the InVEST model output—the baseline carbon reserve color value—as RGB channel data, and combining it with the target opacity representing the error weight as alpha channel data, these values ​​are seamlessly concatenated in the fragment shader to construct the target pixel's RGBA vector and generate the target rendering matrix. This mechanism allows for the simultaneous presentation of the objective physical measurement of carbon reserves and the spatial confidence level of the data source within a single rendering view. It not only accurately preserves the global ecological distribution characteristics but also intuitively exposes local classification uncertainties, significantly improving the data presentation accuracy of mining area ecological environment monitoring. Attached Figure Description

[0014] Figure 1 A schematic diagram illustrating the steps of the InVEST carbon storage confidence rendering method based on machine learning provided in the embodiments of this application;

[0015] Figure 2 A schematic diagram of the logical flow of the InVEST carbon storage confidence rendering method based on machine learning provided in the embodiments of this application;

[0016] Figure 3 A schematic diagram of the structure of the InVEST carbon storage confidence rendering system based on machine learning provided in the embodiments of this application. Detailed Implementation

[0017] This application's embodiments utilize the InVEST carbon reserve confidence rendering method and system based on machine learning to address the spatial visualization challenges of complex ecological transition zones in mining areas. It solves the technical problem that existing unidirectional deterministic graphics rendering pipelines lack a cross-modal coupling mechanism to map the classification errors of underlying multi-models to visual attributes, resulting in uniform and smooth coloring in a single view severely masking objectively existing discrepancies in underlying spatial calculations, thus creating "pseudo-high confidence" visual deception.

[0018] In complex ecological environments, machine vision and computational models cannot guarantee the absolute truthfulness of attribute determinations for every physical coordinate point. Especially in the border areas of vegetation degradation or in restoration transition zones within mining areas, multiple machine learning algorithms often provide drastically different classification predictions for the same plot of land. This algorithmic divergence, induced by the ambiguity of the objective physical environment, is the core confidence evaluation benchmark in the data chain. However, existing rendering systems completely disregard this, uniformly applying a deterministic color scheme, resulting in visual deception of the data. This solution aims to break down this black box barrier and re-establish a direct causal path from underlying data contradictions to front-end visual representation.

[0019] This approach first delves into the underlying layers of data prediction, moving beyond simply seeking a single final classification result to comprehensively acquire the set of vegetation categories for each target pixel coordinate point on the thematic map of target carbon storage. By statistically analyzing the voting results of various algorithms and converting them into category voting frequencies, this approach exposes the controversies hidden within the algorithms. Subsequently, it uses the principle of information entropy to calculate the category voting frequencies, deriving a precise quantification of the degree of disagreement—the classification information entropy. To enable the front-end graphics engine to understand and handle this abstract statistical bias, this approach further performs global normalization on the classification information entropy corresponding to all pixel coordinate points within the preset rendering area, constructing a sub-texture of the degree of disagreement in the form of a two-dimensional single-channel grayscale matrix. This essentially reduces and solidifies multi-dimensional logical conflicts into a spatial texture mapping table that can directly participate in graphics calculations.

[0020] In the graphics rendering stage, this scheme abandons the traditional single-line shading logic. On one hand, it acquires the baseline carbon storage color value and baseline opacity pre-calculated based on the InVEST model as the base data representing the absolute value of surface carbon sinks. On the other hand, in the fragment shader of the very low-level graphics rendering pipeline, it accurately extracts the classification information entropy corresponding to the divergence subtexture. This scheme cleverly uses the classification information entropy to perform negative attenuation calculation on the baseline opacity, generating a brand-new target opacity. This step is the soul of the entire data reconstruction, which makes spatial areas with huge calculation divergence automatically become transparent and dark, while high-confidence areas remain substantial and full. Finally, the baseline carbon storage color value representing the absolute value is embedded in the RGB channel data, and the target opacity representing the spatial confidence is embedded in the Alpha channel data. These are seamlessly spliced ​​in the fragment shader to construct the target pixel RGBA vector, and the target rendering matrix is ​​generated to update the target carbon storage thematic map.

[0021] The core innovation of this solution lies in its pioneering cross-modal coupled rendering logic that addresses the discrepancies between spatially measured numerical values ​​and underlying computational features. This logic precisely transforms the abstract, multi-model discrete classification contradictions into a transparency decay algorithm natively supported by the graphics pipeline. Without altering the underlying ecological model's structure, it perfectly solves the problem of existing technologies masking computational errors and creating a false high-confidence visual illusion through unidirectional linear rendering. This technological revolution enables mining area environmental monitoring to simultaneously display macroscopic ecological distribution and provide insightful control over the reliability of microscopic algorithms within a single view.

[0022] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0023] like Figure 1 The diagram illustrates the steps of the InVEST carbon storage confidence rendering method based on machine learning provided in this embodiment. The InVEST carbon storage confidence rendering method based on machine learning includes the following steps: For scenarios involving confidence rendering and multimodal graphical visualization of spatial ecological data in complex surface environments and ecological transition zones, existing graphics rendering pipelines generally employ a deterministic unidirectional logical mapping mechanism, directly and forcibly converting the final calculated absolute value into a deterministic visual layer. This severely severs the causal relationship between the objective calculation differences within the underlying data production process and the final physical representation of the graphics. To address this issue, this embodiment provides an InVEST carbon storage confidence rendering method based on machine learning, with the specific execution steps as follows:

[0024] This method obtains the set of vegetation categories for the target pixel coordinates in the thematic map of target carbon storage. The vegetation category set represents the set of vegetation categories predicted by a preset number of independent machine learning classifiers. In complex surface ecological monitoring engineering practices, the preset number typically ranges from 3 to 5, and is obtained by reading the system configuration parameter table. This scheme often employs a combination of algorithms such as random forests, support vector machines, and classification and regression trees to ensure sufficient model heterogeneity and decision complementarity in the underlying spatial parameter extraction stage. In specific application scenarios, this scheme is particularly suitable for typical open-pit mine areas with long-term development. These areas are characterized by long mining histories, significant surface disturbance, and complex transitional zones between vegetation destruction and restoration. Their land cover attributes, such as grassland, shrubland, coniferous forest, and bare soil, are prone to contradictory judgments among different classifiers.

[0025] The number of each vegetation category in the vegetation category set is counted, and then divided by the total number of vegetation categories in the set to obtain the category voting frequency for each vegetation category. The category voting frequency reflects the concentration of the probability distribution of a specific land cover attribute under multiple model evaluations.

[0026] The classification information entropy of the target pixel coordinates is obtained by calculating the information entropy of the category voting frequency. Then, the classification information entropy corresponding to all pixel coordinates within the preset rendering area of ​​the target carbon storage thematic map is normalized to construct a bifurcation subtexture in the form of a two-dimensional single-channel grayscale matrix. The preset rendering area is defined as the set of WebGL viewport bounding box coordinates allocated by the current front-end interactive interface, typically nested within the DOM container of a web visualization display interface built on HTML5 and CSS3. The bifurcation subtexture accurately transforms the abstract multi-model discrete classification contradictions into a texture map table that can be natively parsed by the graphics rendering pipeline, serving as the direct spatial data source for subsequent opacity modulation.

[0027] Obtain the baseline carbon storage color value and baseline opacity corresponding to the target pixel coordinates pre-calculated based on the InVEST model. The baseline carbon storage color value represents the absolute physical measurement index output by the InVEST carbon sink assessment model, which is usually obtained by reading the pre-configured ecological index pseudo-color mapping table; the baseline opacity represents the default completely opaque rendering state, and its value is fixed at 1.0.

[0028] The classification information entropy corresponding to the target pixel coordinates in the divergence subtexture is extracted. In the fragment shader of the graphics rendering pipeline, the classification information entropy is negatively decayed with respect to the baseline opacity to generate the target opacity. The engineering implementation of this step is constrained within the GPU parallel pipeline, ensuring high-concurrency real-time computing performance for tens of millions of pixels. Furthermore, all complex graphics shading calculations are completed under browser client hardware acceleration, eliminating the need for frequent secondary redraw requests to the server and ensuring smooth interactive rendering.

[0029] The baseline carbon storage color value is used as the RGB channel data, and the target opacity is used as the alpha channel data. These are concatenated in the fragment shader to construct the target pixel's RGBA vector. The target pixel's RGBA vector fully defines the color and light transmission physical properties of a single pixel in computer graphics.

[0030] A target rendering matrix is ​​generated based on the RGBA vectors of the target pixels corresponding to all pixel coordinates within a preset rendering area, and the target carbon reserve thematic map is rendered and updated accordingly. The target rendering matrix is ​​directly pushed into the graphics card frame buffer object for screen swapping display. The final visualization result is presented to the user in a unified manner through the platform's front-end navigation system, such as a navigation bar that includes the system homepage, mine overview, and dynamic displays in a logical order.

[0031] Figure 2This is a schematic diagram of the logical flow of the InVEST carbon storage confidence rendering method based on machine learning provided in this application embodiment. For spatial visualization of complex surface environments and ecological transition zones, existing unidirectional deterministic graphics rendering mechanisms fail to transmit underlying multi-model classification errors, resulting in uniform smooth coloring that severely masks objectively existing spatial computational discrepancies. This core basic solution transforms the prediction discrepancies of independent machine learning classifiers into classification information entropy and constructs a discrepancy degree sub-texture input graphics rendering pipeline. Within the fragment shader, it innovatively utilizes classification information entropy to perform negative attenuation calculations on the baseline opacity, generating a target opacity containing the confidence gradient. Finally, the baseline carbon storage color value and the target opacity are concatenated into an RGBA vector of the target pixel for rendering. This overall technical chain breaks down the barrier of underlying error propagation, enabling ecological transition zones with greater computational discrepancies to appear visually more transparent. It successfully achieves high-fidelity coupled rendering of the measured absolute values ​​and the underlying spatial assessment confidence, completely eliminating visual deception of pseudo-high confidence and significantly improving the scientific rigor and intuitiveness of ecological monitoring and analysis.

[0032] Furthermore, a detailed explanation of the underlying mathematical dimensions is provided for the specific acquisition process of classification information entropy in the core basic scheme. Rigorous quantification of model divergence is a necessary prerequisite for achieving accurate negative decay calculation, and using the Shannon information entropy formula to measure system disorder is the only optimal solution for assessing the uncertainty of multidimensional discrete variables.

[0033] Specifically, the classification information entropy of the target pixel coordinates is obtained by calculating the information entropy of the category voting frequency. The specific calculation formula strictly follows the information theory standard, as follows:

[0034] ,in, The classification information entropy of the target pixel coordinates. It represents the two-dimensional coordinates corresponding to the target pixel coordinate point, which is used for pixel-level positioning in the image matrix; This represents the total number of vegetation categories contained in the vegetation category set, and its value is derived from a pre-defined list of land use type specifications for the study area. For the first The voting frequency for each vegetation category is strictly limited to a closed interval between 0 and 1. When Pi for a certain category is 0, according to the principle of limiting convergence, the frequency is further limited. =0, to avoid logarithmic operation errors and ensure stable convergence of the information entropy calculation algorithm at the full data boundary.

[0035] Through the above technical solution, this embodiment introduces a rigorous Shannon information entropy calculation formula to quantitatively analyze the category voting frequency, solving the specific technical problem that the discrete classification results of multiple models in complex surface ecological transition zone scenarios are difficult to be quantitatively identified by the graphics system. It realizes the unique technical effect of accurately transforming the abstract multi-source prediction contradictions into classification information entropy features with clear mathematical definitions and continuous numerical spaces, providing a solid data foundation for the continuous transparency degradation of the subsequent graphics rendering pipeline.

[0036] Furthermore, the negative attenuation calculation logic executed within the fragment shader is specifically constructed in depth. The specific process for obtaining the target opacity is as follows:

[0037] Obtain the preset theoretical maximum classification information entropy. The formula for calculating target opacity is:

[0038] ,in, Due to the lack of transparency of the target, As a baseline opacity, For classification information entropy, The pre-defined theoretical maximum classification information entropy, The preset visual degradation coefficient, for The function represents the hardware-level truncation operator built into the fragment shader.

[0039] The preset visual degradation coefficient setting rule is derived from the nonlinear response characteristics of the human eye to transparency perception in the Weber-Fechner law, and the typical engineering value range is 0.8 to 1.5.

[0040] The function forces the calculation result inside the parentheses to be within the physical range of 0 to 1. It is a GPU rendering hardware instruction that ensures that the graphics pipeline will not crash due to color channel overflow when there is abnormally large divergent data input.

[0041] Through the above technical solution, this embodiment constructs a normalized negative decay algorithm that combines the theoretical maximum classification information entropy and the visual degradation coefficient with a hardware-level truncation operator. This solves the specific technical problem that existing graphics systems are prone to rendering overflow or visual gradation that does not conform to the human eye's perception rules when processing extremely large error data. It achieves a unique technical effect of mapping classification information entropy to target opacity that can be directly used by the graphics rendering engine in a smooth, safe and human visual physiological manner.

[0042] Furthermore, while the preceding scheme provides a smooth decay of transparency, in collapse areas where model discrepancies are extremely severe, such as mining area boundaries, simple transparency can easily cause users to overlook the existence of these areas. Therefore, it is necessary to progressively introduce periodic color warnings, which amplify the visual impact of extreme errors.

[0043] Before constructing the RGBA vector of the target pixel in the fragment shader, the process also includes obtaining a preset divergence threshold. The preset divergence threshold is set based on the prior confidence interval of the historical classification model, and is usually set to 80% of the preset theoretical maximum classification information entropy. It is obtained by reading the rendering engine's security configuration file.

[0044] If the classification information entropy is detected to be greater than or equal to a preset divergence threshold, a high-frequency jitter warning signal is generated in the fragment shader using a unit step function, and the output intensity of the high-frequency jitter warning signal is superimposed on the RGB channel data. Here, the unit step function acts as a hard isolation threshold switch, ensuring that the warning signal is activated only in extremely poor data regions.

[0045] The specific formula for calculating the output strength of the high-frequency jitter warning signal is as follows:

[0046] ,in, The output strength of the high-frequency jitter warning signal. It is a unit step function. To preset the divergence threshold, For classification information entropy, The preset high-frequency oscillation frequency, Represents the sine function. This represents the cosine function.

[0047] This is a preset high-frequency oscillation frequency, the setting of which needs to be adapted to the pixel density of the display device. Typical engineering values ​​are usually set to 10.0 to 50.0. Represents the sine function. The cosine function is represented by the two, which together form a high-frequency cross-grid pattern in a two-dimensional spatial domain.

[0048] Through the above technical solution, this embodiment introduces a high-frequency jitter warning signal generation mechanism based on the unit step function and spatial trigonometric function, which solves the specific technical problem that the lack of sufficient visual warning power in the extreme error area of ​​smooth transparency decay makes it easy for analysts to miss key hidden danger patches. It achieves a unique technical effect of triggering grid jitter rendering warning with strong visual exclusivity in extreme classification divergence areas without changing the distribution topology of the baseline ecological data.

[0049] Furthermore, the focus is on addressing the spatial hierarchy challenges arising from user interactions. In complex ecological monitoring systems, users frequently zoom in and out of their viewpoint, and measures are taken to prevent sampling distortion caused by changes in the viewpoint hierarchy.

[0050] Before extracting the classification information entropy corresponding to the target pixel coordinates in the divergence subtexture, the process also includes: obtaining the screen spatial resolution of the preset rendering area and the original pixel spatial resolution of the divergence subtexture. The screen spatial resolution is defined as the actual geographic meters corresponding to each physical pixel on the current user device screen, and the original pixel spatial resolution is defined as the inherent raster size of the remote sensing inversion data itself.

[0051] If the screen space resolution is detected to be greater than the original pixel space resolution, the default single-point texture sampling instruction is blocked in the graphics rendering pipeline. Blocking the default single-point texture sampling instruction is a routine and necessary technique to prevent the loss of subtle bifurcation features when scaling to large scales.

[0052] Divide the screen spatial resolution by the original pixel spatial resolution to obtain the spatial downsampling ratio; with the target pixel coordinate point as the center, obtain the texture coordinate offset step size of the divergence subtexture per unit pixel. Multiply the spatial downsampling ratio by the texture coordinate offset step size to obtain the coordinate span boundary of the dynamic sampling window, and construct the dynamic sampling window accordingly.

[0053] The process involves obtaining the total number of pixels within the dynamic sampling window; extracting the classification information entropy of all neighboring pixels within the dynamic sampling window; identifying the local maximum entropy value; and counting the number of non-zero entropy values ​​in pixels within the dynamic sampling window. Dividing the non-zero number by the total number of pixels within the dynamic sampling window yields the spatial distribution density. The local maximum entropy value and the spatial distribution density are then weighted and multiplied to generate the macroscopic representation information entropy. The convergence logic of the weighted product calculation is that the macroscopic representation information entropy reaches its peak only when the local maximum entropy value is extremely high and the non-zero entropy values ​​are widely distributed, thus avoiding excessive alarms caused by isolated noise points.

[0054] The macroscopic representation information entropy is used as the classification information entropy corresponding to the extracted target pixel coordinates and output to the fragment shader to perform negative decay calculation.

[0055] Through the above technical solution, this embodiment solves the specific refinement problem that traditional single-point texture sampling causes tiny high-disparity boundaries to be visually swallowed and discarded under large-scale macroscopic zoomed views by constructing a dynamic sampling window based on the comparison between the screen and the original spatial resolution and a macroscopic representation information entropy aggregation mechanism. It achieves the unique technical effect of high-fidelity penetration and transmission of the underlying fine-grained classification error under any level of roaming.

[0056] Furthermore, within the constructed dynamic sampling window, if the information entropy distribution within the region is extremely uneven, extracting only a single macroscopic representation information entropy will result in the loss of heterogeneous structural features within the region. Therefore, it is necessary to further introduce a variance-driven chaotic noise technique.

[0057] Using the baseline carbon storage color value as RGB channel data, the method also includes extracting the spatial variance feature of the classification information entropy of all neighboring pixels within the dynamic sampling window. The spatial variance feature reflects the degree of drastic oscillation in the spatial distribution of the prediction discrepancies between different models within the window.

[0058] If the spatial variance feature is greater than a preset heterogeneity threshold, a significant classification divergence is determined to exist within the current viewport, and the Berlin noise algorithm is invoked. The spatial variance feature is used as the frequency input parameter for the Berlin noise algorithm to generate an undirected granular noise signal with local self-similarity. The preset heterogeneity threshold is typically set between 0.15 and 0.3. As a standard procedural texture generation algorithm in computer graphics, the Berlin noise algorithm uses the spatial variance feature to control the frequency of its gradient vector, resulting in finer and more chaotic noise particles generated in regions with stronger spatial heterogeneity.

[0059] Undirected particle noise signals are used as confidence warning masks and superimposed onto the baseline carbon reserve color value using a linear interpolation algorithm to obtain the updated baseline carbon reserve color value. The mixing factor of the linear interpolation algorithm is typically set to 0.3 to 0.5 to avoid noise completely covering the original background color.

[0060] Through the above technical solution, this embodiment generates undirected granular noise signals by extracting spatial variance features to drive the Berlin noise algorithm and mixes them with the reference color through linear interpolation. This solves the problem of the inability to intuitively express the specific refinement of the extremely fragmented and heterogeneous spatial topology of multiple model divergences within the macroscopic aggregation window, and achieves the unique technical effect of using fractal chaotic textures as a natural metaphor for the underlying data to predict the state of violent structural oscillations.

[0061] Furthermore, considering that vegetation degradation in mining areas should have ecological and physical inertia over a long time series, it is necessary to superimpose the temporal continuity constraint with the spatial constraint of the preceding scheme to solve the problem of dynamic confidence assessment when the algorithm repeatedly jumps between years.

[0062] After generating the target opacity, the process also includes: obtaining the historical category voting frequency sequence of the target cell coordinates within a preset continuous time window. In practical applications, the preset continuous time window typically relies on a native JavaScript time series carousel component integrated into the front-end platform, which obtains the long-term evolution span node sequence currently anchored by the user.

[0063] Calculate the temporal variance of the historical category voting frequency sequence; if the temporal variance exceeds a preset temporal stability threshold, then the target pixel coordinate point is determined to exhibit classification oscillation characteristics. The preset temporal stability threshold is derived from the empirical blocking value of the Markov chain steady-state transition probability matrix.

[0064] The time series penalty coefficient is calculated based on the ratio of the time series fluctuation variance to a preset time series stability threshold. This ratio calculation forms a dynamic penalty mechanism, and the convergence logic limits the maximum penalty coefficient to no more than 2.0.

[0065] The target opacity is divided by the timing penalty factor to generate a secondary decay opacity; this secondary decay opacity replaces the target opacity and is temporarily stored as alpha channel data in the rendering queue. This temporary storage operation is performed by the graphics API's memory stack buffer.

[0066] Through the above technical solution, this embodiment introduces a temporal fluctuation variance evaluation mechanism within a preset continuous time window and generates a temporal penalty coefficient for secondary opacity decay. This solves the specific and detailed technical problem of the underlying classification model in long-term ecological monitoring giving anti-physical intuition jump evaluations between adjacent years due to the fragility of the decision boundary. It achieves the unique technical effect of using the physical inertia of the ecological time dimension to dynamically verify the classification error of single-period static space and suppress cross-period rendering.

[0067] Furthermore, after generating the secondary attenuation opacity, the process also includes: acquiring a digital elevation model matrix aligned with the target carbon reserve thematic map in spatial coordinates. The digital elevation model matrix is ​​a key three-dimensional data source characterizing the true undulating topography of the mining area.

[0068] The neighborhood elevation array of the target pixel coordinates is extracted from the digital elevation model matrix, and the terrain slope angle corresponding to the target pixel coordinates is calculated using the spatial gradient algorithm. The spatial gradient algorithm is implemented by taking the partial derivative of the normal vector of the 3D point cloud.

[0069] If the detected temporal fluctuation variance exceeds the temporal stability threshold and the terrain slope angle exceeds the preset steep slope threshold, the classification oscillation feature is determined to be observation noise induced by terrain shadows, and a preset penalty amplification factor is extracted. The preset steep slope threshold is consistently set between 35 and 45 degrees; the preset penalty amplification factor is usually greater than 1.0. The physical mechanism of this logic closely matches the characteristics of extremely deep pits and steep spoil heaps left by open-pit mining activities. Since steep slopes are prone to spectral distortion due to changes in the solar incidence angle, the temporal oscillation of the model in this case must be spurious noise, and the transparency penalty must be amplified.

[0070] If the detected temporal fluctuation variance is greater than the temporal stability threshold and the terrain slope angle is less than or equal to the steep slope threshold, the classification oscillation feature is determined to be a real change in the physical surface, and a preset penalty suppression factor is extracted. The preset penalty suppression factor is strictly limited to a value between 0 and 1, and is used to reverse the erroneous penalties generated by the above scheme.

[0071] The temporary time-series penalty coefficients are multiplied by the penalty amplification factor or the penalty suppression factor to generate the terrain coupling penalty coefficients. The target opacity is divided by the terrain coupling penalty coefficients to generate the third-order attenuation opacity, which is then used as the updated Alpha channel data input into the subsequent stitching and construction steps.

[0072] Through the above technical solution, this embodiment introduces a digital elevation model to extract the terrain slope angle and temporal fluctuation variance for three-dimensional spatiotemporal orthogonal cross-validation. This solves the specific technical problem that relying solely on the time dimension variance judgment can easily lead to the misjudgment of real high-frequency mechanical excavation on a flat working surface as model algorithm error and result in excessive transparency penalties. It achieves the unique technical effect of accurately separating optical shadow observation noise from the real ecological structure evolution from the physical geoscience causal mechanism.

[0073] Figure 3This is a schematic diagram of the InVEST carbon storage confidence rendering system based on machine learning provided in an embodiment of this application. The InVEST carbon storage confidence rendering system based on machine learning includes: a vegetation category set acquisition module: used to acquire the vegetation category set of the target pixel coordinates of the target carbon storage thematic map; the vegetation category set represents the set of vegetation categories predicted and output by a preset number of independent machine learning classifiers; a category voting frequency processing module: used to count the number of each vegetation category in the vegetation category set and divide it by the total number of vegetation categories in the vegetation category set to obtain the category voting frequency of each vegetation category; and a divergence subtexture processing module: used to calculate the classification information entropy of the target pixel coordinates by performing information entropy calculation on the category voting frequency, and normalize the classification information entropy corresponding to all pixel coordinates within a preset rendering area of ​​the target carbon storage thematic map to construct a two-dimensional single-channel... The system consists of: a grayscale matrix-based subtexture for bifurcation; a baseline acquisition module for acquiring baseline carbon storage color values ​​and baseline opacity corresponding to target pixel coordinates pre-calculated based on the InVEST model; a target opacity processing module for extracting classification information entropy corresponding to target pixel coordinates from the bifurcation subtexture, and negatively attenuating the classification information entropy with the baseline opacity in the fragment shader of the graphics rendering pipeline to generate target opacity; a target pixel RGBA vector processing module for concatenating the baseline carbon storage color value as RGB channel data and the target opacity as Alpha channel data in the fragment shader to construct the target pixel RGBA vector; and a target carbon storage thematic map rendering module for generating a target rendering matrix based on the target pixel RGBA vectors corresponding to all pixel coordinates within a preset rendering area, and rendering and updating the target carbon storage thematic map accordingly.

[0074] This system can be deployed on servers with parallel computing capabilities and web rendering clients. In actual deployment, the system is built on a browser / server architecture. The server side utilizes the IIS server built into the Windows system for environment setup and local / external network web page publishing, and sends data resources to client browsers via the HTTP protocol without requiring the installation of special software.

[0075] This application also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the InVEST carbon storage confidence rendering method based on machine learning.

[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0081] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A machine learning-based InVEST carbon storage confidence rendering method, characterized in that, Includes the following steps: Obtain the vegetation category set of the target pixel coordinates of the target carbon storage thematic map, wherein the vegetation category set represents the set of vegetation categories predicted by a preset number of independent machine learning classifiers; The number of each vegetation category in the vegetation category set is counted, and then divided by the total number of vegetation categories in the vegetation category set to obtain the category voting frequency of each vegetation category; The classification information entropy of the target pixel coordinate point is obtained by calculating the information entropy of the category voting frequency, and the classification information entropy of all pixel coordinate points in the preset rendering area of ​​the target carbon storage thematic map is normalized to construct a two-dimensional single-channel gray-scale matrix form of the divergence subtexture. Obtain the baseline carbon storage color value and baseline opacity corresponding to the target pixel coordinate point pre-calculated based on the InVEST model; Extract the classification information entropy corresponding to the target pixel coordinates in the divergence subtexture, and in the fragment shader of the graphics rendering pipeline, perform negative attenuation calculation on the baseline opacity to generate the target opacity. The baseline carbon storage color value is used as RGB channel data, and the target opacity is used as Alpha channel data. They are concatenated in the fragment shader to construct the target pixel RGBA vector. A target rendering matrix is ​​generated based on the RGBA vectors of the target pixels corresponding to all pixel coordinates within the preset rendering area, and the target carbon storage thematic map is rendered and updated accordingly.

2. The InVEST carbon storage confidence rendering method based on machine learning according to claim 1, characterized in that, The specific process for obtaining the classification information entropy is as follows: The classification information entropy of the target pixel coordinate point is obtained by calculating the information entropy of the category voting frequency. The specific calculation formula satisfies: ,in, The classification information entropy of the target pixel coordinates. This represents the two-dimensional coordinates corresponding to the target pixel coordinate point. This represents the total number of vegetation categories contained in the set of vegetation categories. For the first Category voting frequency for each vegetation category.

3. The InVEST carbon storage confidence rendering method based on machine learning according to claim 2, characterized in that, The specific process for obtaining the target opacity is as follows: Obtain the preset theoretical maximum classification information entropy; ,in, Due to the lack of transparency of the target, As a baseline opacity, For classification information entropy, The pre-defined theoretical maximum classification information entropy, The preset visual degradation coefficient, The saturate function represents the hardware-level truncation operator built into the fragment shader.

4. The InVEST carbon storage confidence rendering method based on machine learning according to claim 2, characterized in that, Before concatenating the fragments into an RGBA vector for the target pixel in the fragment shader, the following steps are also included: Obtain the preset divergence threshold; If the classification information entropy is detected to be greater than or equal to the preset divergence threshold, a high-frequency jitter warning signal is generated in the fragment shader by using a unit step function, and the output intensity of the high-frequency jitter warning signal is superimposed onto the RGB channel data. The specific formula for calculating the output strength of the high-frequency jitter warning signal is as follows: ,in, The output strength of the high-frequency jitter warning signal. It is a unit step function. To preset the divergence threshold, For classification information entropy, The preset high-frequency oscillation frequency, Represents the sine function. This represents the cosine function.

5. The InVEST carbon storage confidence rendering method based on machine learning according to claim 1, characterized in that, Before extracting the classification information entropy corresponding to the target pixel coordinates in the divergence subtexture, the following is also included: Obtain the screen space resolution of the preset rendering area, and the original pixel space resolution of the divergence subtexture; If the screen space resolution is detected to be greater than the original pixel space resolution, the default single-point texture sampling instruction is blocked in the graphics rendering pipeline. Divide the screen spatial resolution by the original pixel spatial resolution to obtain the spatial downsampling ratio; take the target pixel coordinate point as the center to obtain the texture coordinate offset step size of the divergence subtexture per unit pixel; Multiply the spatial downsampling ratio by the texture coordinate offset step size to obtain the coordinate span boundary of the dynamic sampling window, and construct the dynamic sampling window accordingly. Obtain the total number of pixels within the dynamic sampling window; Extract the classification information entropy of all neighboring pixels within the dynamic sampling window, filter out the local maximum entropy value, and count the number of non-zero entropy values ​​of pixels within the dynamic sampling window. Divide the non-zero number by the total number of pixels in the dynamic sampling window to obtain the spatial distribution density; calculate the macroscopic representation information entropy by performing a weighted product of the local maximum entropy value and the spatial distribution density. The macroscopic representation information entropy is used as the classification information entropy corresponding to the extracted target pixel coordinates and output to the fragment shader to perform negative decay calculation.

6. The InVEST carbon storage confidence rendering method based on machine learning according to claim 5, characterized in that, Using the baseline carbon reserves color value as RGB channel data, it also includes: Extract the spatial variance features of the classification information entropy of all neighboring pixels within the dynamic sampling window; If the spatial variance feature is greater than the preset heterogeneity threshold, it is determined that there is a significant classification divergence in the current window, and the Berlin noise algorithm is called. The spatial variance feature is used as the frequency input parameter of the Berlin noise algorithm to generate an undirected granular noise signal with local self-similarity. The undirected particle noise signal is used as a confidence warning mask and superimposed onto the baseline carbon reserve color value through a linear interpolation algorithm to obtain the updated baseline carbon reserve color value.

7. The InVEST carbon storage confidence rendering method based on machine learning according to claim 6, characterized in that, After generating the target opacity, the following is also included: Obtain the historical category voting frequency sequence of the target cell coordinates within a preset continuous time window; Calculate the time-series variance of the historical category voting frequency sequence; If the variance of the temporal fluctuation is greater than the preset temporal stability threshold, then the target pixel coordinate point is determined to have classification oscillation characteristics. The time series penalty coefficient is calculated based on the ratio of the time series fluctuation variance to the preset time series stability threshold. Divide the target opacity by the time penalty coefficient to generate a secondary decay opacity; Replace the target opacity with the secondary decay opacity and temporarily store it in the rendering queue as Alpha channel data.

8. The InVEST carbon storage confidence rendering method based on machine learning according to claim 7, characterized in that, After generating the secondary decay opacity, the following is also included: Obtain a digital elevation model matrix aligned with the thematic map of the target carbon storage in spatial coordinates; The neighborhood elevation array of the target pixel coordinate point is extracted from the digital elevation model matrix, and the terrain slope angle corresponding to the target pixel coordinate point is calculated using the spatial gradient algorithm. If the detected time series fluctuation variance is greater than the time series stability threshold and the terrain slope angle is greater than the preset steep slope threshold, then the classification oscillation feature is determined to be observation noise induced by terrain shadow, and the preset penalty amplification factor is extracted. If the detected temporal fluctuation variance is greater than the temporal stability threshold and the terrain slope angle is less than or equal to the steep slope threshold, then the classification oscillation feature is determined to be a real change in the physical surface, and the preset penalty suppression factor is extracted. The temporary time-series penalty coefficients are multiplied by the penalty amplification factor or the penalty suppression factor to generate the terrain coupling penalty coefficients. Divide the target opacity by the terrain coupling penalty coefficient to generate a third-degree attenuation opacity, which is then used as the updated Alpha channel data in the subsequent stitching and construction steps.

9. A machine learning-based InVEST carbon storage confidence rendering system, characterized in that, include: Vegetation category set acquisition module: used to acquire the vegetation category set of the target pixel coordinates of the target carbon storage thematic map, wherein the vegetation category set represents the set of vegetation categories predicted by a preset number of independent machine learning classifiers; Category voting frequency processing module: used to count the number of each vegetation category in the vegetation category set, and divide it by the total number of vegetation categories in the vegetation category set to obtain the category voting frequency of each vegetation category; The divergence subtexture processing module is used to calculate the classification information entropy of the target pixel coordinates by calculating the information entropy of the category voting frequency, and to normalize the classification information entropy of all pixel coordinates in the preset rendering area of ​​the target carbon storage thematic map, and construct a divergence subtexture in the form of a two-dimensional single-channel grayscale matrix. The benchmark acquisition module is used to acquire the benchmark carbon storage color value and benchmark opacity corresponding to the target pixel coordinate point pre-calculated based on the InVEST model. Target opacity processing module: used to extract the classification information entropy corresponding to the target pixel coordinates in the divergence subtexture. In the fragment shader of the graphics rendering pipeline, the classification information entropy is negatively attenuated to the baseline opacity to generate the target opacity. The target pixel RGBA vector processing module is used to construct the target pixel RGBA vector by concatenating the base carbon storage color value as RGB channel data and the target opacity as Alpha channel data in the fragment shader. The target carbon reserves thematic map rendering module is used to generate a target rendering matrix based on the RGBA vectors of the target pixels corresponding to all pixel coordinates within a preset rendering area, and to render and update the target carbon reserves thematic map accordingly.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images