Spatiotemporal Modeling Method and System for Farmland Carbon Sequestration Remote Sensing Data Based on Machine Learning

Abstract

This invention discloses a method and system for spatiotemporal modeling of farmland carbon sequestration remote sensing data based on machine learning. The method relates to the field of spatiotemporal modeling technology and includes the following steps: completeness quantification, completeness optimization, coupling quantification, and coupling optimization. This invention evaluates the completeness of the collaborative representation using multi-source remote sensing and ground-based measured data, dynamically optimizes extraction frequency and area thresholds, and constructs training samples. It trains a carbon sequestration spatiotemporal model based on machine learning, evaluates the coupling degree of carbon cycle-remote sensing data, dynamically adjusts the pixel alignment window and adaptation resolution, and finally outputs a dynamic spatiotemporal distribution map of farmland carbon sequestration. This improves the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data and solves the problem of low reliability in existing technologies due to the mismatch between remote sensing data representation and actual carbon processes.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal modeling technology, and in particular to a method and system for spatiotemporal modeling of farmland carbon sequestration remote sensing data based on machine learning. Background Technology

[0002] The core of farmland carbon sinks is soil organic carbon and vegetation biomass, which together determine carbon storage. However, current remote sensing inversion relies heavily on optical bands (such as Landsat's red and near-infrared bands) or a single microwave band, which makes it impossible to fully characterize the carbon pool composition. At the same time, farmland has fragmented plot characteristics, but low-to-medium resolution remote sensing data can mask the spatial heterogeneity of soil organic carbon within the plots, making it impossible to accurately quantify the spatiotemporal distribution of small-scale carbon sinks when modeling.

[0003] Remote sensing modeling of farmland carbon sequestration requires the integration of multi-source data, including optical, microwave, and lidar data. However, the spatiotemporal references of different data sources differ. Different satellites have different projection coordinate systems and pixel sizes, which can easily lead to edge effects during data resampling, resulting in distortion of carbon sequestration parameters at field boundaries and spatial reference differences. Different satellites have different transit times, while farmland carbon sequestration-related parameters (such as surface temperature and photosynthetic rate) exhibit intra-diurnal variations. Direct fusion can introduce time difference errors, resulting in temporal reference differences. Current spatiotemporal modeling of farmland carbon sequestration often uses models such as multiple linear regression, which lack characterization of the physicochemical processes of the carbon cycle. This makes it impossible to explain the causes of carbon sequestration differences and has weak generalization ability. Estimation through interpolation or substitution of indicators leads to amplified input errors, ultimately affecting the carbon sequestration simulation results. This results in poor coupling between the carbon sequestration process model and remote sensing data, leading to low reliability of spatiotemporal modeling of farmland carbon sequestration due to the mismatch between remote sensing data representation and real carbon processes. Summary of the Invention

[0004] This application provides a method and system for spatiotemporal modeling of farmland carbon sequestration remote sensing data based on machine learning. This solves the problem in the prior art where the spatiotemporal modeling of farmland carbon sequestration remote sensing data is unreliable due to the mismatch between the remote sensing data representation and the actual carbon process, thereby improving the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data.

[0005] On the one hand, a spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning is provided, including the following steps: acquiring multi-source remote sensing data of the target farmland area within a preset time series, and simultaneously collecting ground-measured farmland carbon sequestration data of the corresponding time series of the target farmland area; obtaining the multi-source data collaborative representation completeness based on the multi-source remote sensing data and the ground-measured farmland carbon sequestration data, which is used to measure the ability of multi-source remote sensing data to represent farmland carbon sequestration parameters; determining whether to perform completeness optimization based on the multi-source data collaborative representation completeness; if so, constructing a training sample dataset for machine learning after completeness optimization; otherwise, directly constructing a training sample dataset for machine learning. The optimization includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold. A machine learning-based training sample dataset is used as input to train a spatiotemporal model of farmland carbon sequestration remote sensing data. The carbon cycle-remote sensing data coupling degree is obtained to measure the coupling degree between the carbon sequestration process model and the remote sensing data during the carbon cycle. Based on the carbon cycle-remote sensing data coupling degree, it is determined whether to perform coupling degree optimization. If so, a spatiotemporal dynamic distribution map of farmland carbon sequestration in the target farmland area is output after coupling degree optimization; otherwise, the spatiotemporal dynamic distribution map of farmland carbon sequestration in the target farmland area is directly output. Coupling degree optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of resolution adaptation.

[0006] On the other hand, a spatiotemporal modeling system for farmland carbon sequestration remote sensing data based on machine learning is provided, including: a completeness quantification module, a completeness optimization module, a coupled quantification module, and a coupledness optimization module. The completeness quantification module acquires multi-source remote sensing data of the target farmland area within a preset time series and simultaneously collects ground-measured farmland carbon sequestration data of the corresponding time series. Based on the multi-source remote sensing data and the ground-measured farmland carbon sequestration data, the completeness of the multi-source data collaborative representation is obtained, which measures the ability of the multi-source remote sensing data to represent farmland carbon sequestration parameters. The completeness optimization module determines whether to perform completeness optimization based on the completeness of the multi-source data collaborative representation. If so, a training sample dataset for machine learning is constructed after completeness optimization; otherwise, a training sample dataset for machine learning is directly constructed. The training sample dataset for machine learning includes completeness optimization, which includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold; the coupling quantification module is used to train the spatiotemporal model of farmland carbon sink remote sensing data with the training sample dataset as input, and obtain the carbon cycle-remote sensing data coupling degree to measure the coupling degree between the carbon sink process model and the remote sensing data during the carbon cycle; the coupling degree optimization module is used to determine whether to perform coupling degree optimization based on the carbon cycle-remote sensing data coupling degree. If so, the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is output after coupling degree optimization; otherwise, the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is directly output. Coupling degree optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of adaptive resolution.

[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0008] 1. By acquiring multi-source remote sensing data of the target farmland area within a preset time series and simultaneously collecting ground-measured farmland carbon sink data of the corresponding time series, the completeness of the multi-source data collaborative representation is obtained based on the multi-source remote sensing data and the ground-measured farmland carbon sink data. This completeness is used to measure the ability of the multi-source remote sensing data to represent farmland carbon sink parameters. The completeness of the multi-source data collaborative representation is used to determine whether to perform completeness optimization. This allows for precise control over the quality of the multi-source data required for farmland carbon sink modeling, avoiding data redundancy and increased processing costs due to over-optimization. Ultimately, this ensures that the quality of the model input data meets the requirements of farmland carbon sink remote sensing modeling. Simultaneously, a balance between data processing accuracy and efficiency is achieved. Based on the training sample dataset of machine learning, a spatiotemporal model of farmland carbon sequestration remote sensing data is trained. The carbon cycle-remote sensing data coupling degree is obtained to measure the coupling degree between the carbon sequestration process model and the remote sensing data during the carbon cycle. The coupling degree of the carbon cycle-remote sensing data is used to determine whether to optimize the coupling degree. While ensuring that the output results can accurately depict the spatiotemporal differences of carbon sequestration (such as nearshore-farshore carbon sequestration gradient, quarterly carbon sequestration changes), a balance between model optimization accuracy and output efficiency is achieved, thereby improving the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data.

[0009] 2. The system determines whether to optimize the completeness of multi-source data collaborative representation. Through dynamic evaluation of the representation capabilities of multi-source remote sensing data, it accurately identifies data gaps or insufficient representation, and intelligently triggers dynamic frequency and area representation threshold adjustments. This ensures high completeness and representativeness of the data input into the machine learning model, significantly improving the accuracy and robustness of the farmland carbon sequestration inversion model. The system also determines whether to dynamically adjust the extraction frequency based on the boundary change rate. By dynamically monitoring the spatiotemporal changes in field boundaries, it intelligently adjusts the boundary extraction frequency, ensuring timely updates of boundary information in areas of drastic change, thus significantly improving the accuracy and timeliness of farmland boundary extraction from remote sensing images. Finally, the system determines whether to dynamically adjust the area representation threshold based on the plot area deviation rate. By real-time monitoring of the deviation between the plot area representation results and the actual area, it intelligently adjusts the area extraction threshold parameters, thereby optimizing the accuracy and stability of plot area extraction in remote sensing images, improving the spatial representation accuracy of farmland carbon sequestration monitoring, and enhancing the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data.

[0010] 3. The system determines whether to optimize the coupling degree between the carbon cycle process and remote sensing data. By dynamically evaluating the matching degree between the carbon cycle process model and remote sensing data, it intelligently triggers dynamic adjustment of the pixel alignment window and dynamic adjustment of the adaptation resolution. This significantly improves the collaborative representation capability of remote sensing data and the carbon cycle process, ensuring that the final output spatiotemporal distribution map of farmland carbon sinks has higher accuracy and reliability. Furthermore, the system determines whether to dynamically adjust the pixel alignment window based on the pixel scale deviation between remote sensing and carbon cycle data. By real-time detection of the spatial deviation between remote sensing data and the carbon cycle model at the pixel scale, it intelligently adjusts the pixel alignment window. By accurately matching the size and location of the data in the spatial dimension, errors caused by scale mismatch are effectively reduced, improving the accuracy and spatial consistency of farmland carbon sequestration remote sensing inversion. Dynamic resolution adjustment is performed based on the coupling residuals of carbon process and remote sensing data. Real-time analysis of the coupling residuals between the carbon process model and remote sensing data in spatial resolution allows for dynamic adjustment of the remote sensing data resolution to match the scale characteristics of the carbon process model. This reduces signal distortion and information loss caused by resolution mismatch, improves the accuracy and adaptability of carbon flux estimation, and enhances the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 Flowchart of the spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning provided in the embodiments of this application;

[0013] Figure 2 A flowchart illustrating the dynamic adjustment of the area representation threshold for a machine learning-based spatiotemporal modeling method for farmland carbon sequestration remote sensing data, as provided in this application embodiment.

[0014] Figure 3 A flowchart illustrating the dynamic adjustment of the pixel alignment window in the spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning, as provided in the embodiments of this application.

[0015] Figure 4 A schematic diagram of the structure of the spatiotemporal modeling map system for farmland carbon sequestration remote sensing data based on machine learning provided in the embodiments of this application. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] This application provides a machine learning-based method and system for spatiotemporal modeling of farmland carbon sequestration remote sensing data. This addresses the problem of low reliability in existing technologies due to mismatches between remote sensing data representation and actual carbon processes. The method assesses the completeness of the collaborative representation using multi-source remote sensing and ground-based measured data, dynamically optimizes extraction frequency and area thresholds, and constructs training samples. It trains a carbon sequestration spatiotemporal model based on machine learning, evaluates the coupling degree between carbon cycle and remote sensing data, dynamically adjusts the pixel alignment window and adaptation resolution, and finally outputs a dynamic spatiotemporal distribution map of farmland carbon sequestration, thereby improving the reliability of spatiotemporal modeling of farmland carbon sequestration remote sensing data.

[0019] The technical solution in this application is to address the aforementioned problem of low reliability in spatiotemporal modeling of farmland carbon sequestration remote sensing data due to the mismatch between remote sensing data representation and actual carbon processes. The overall approach is as follows:

[0020] By acquiring multi-source remote sensing data of the target farmland area within a preset time series and simultaneously collecting ground-measured farmland carbon sink data of the corresponding time series, the completeness of the multi-source data collaborative representation is obtained based on the multi-source remote sensing data and the ground-measured farmland carbon sink data. The completeness of the multi-source data collaborative representation is then used to determine whether to optimize the completeness. If so, a training sample dataset for machine learning is constructed after optimization; otherwise, a training sample dataset for machine learning is directly constructed. Using the machine learning training sample dataset as input, a spatiotemporal model of farmland carbon sink remote sensing data is trained. The coupling degree between the carbon cycle process and remote sensing data is obtained. The coupling degree is then used to determine whether to optimize the coupling degree. If so, a spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is output after optimization; otherwise, a spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is directly output, thus improving the reliability of spatiotemporal modeling of farmland carbon sink remote sensing data.

[0021] 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.

[0022] This invention provides a spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning, such as... Figure 1The flowchart shown is for a spatiotemporal modeling method for farmland carbon sequestration based on machine learning remote sensing data. The processing flow of this method may include the following steps:

[0023] As the first step in the machine learning-based spatiotemporal modeling method for farmland carbon sequestration remote sensing data, the method acquires multi-source remote sensing data of the target farmland area within a preset time series, and simultaneously collects ground-measured farmland carbon sequestration data of the target farmland area in the corresponding time series. Based on the multi-source remote sensing data and the ground-measured farmland carbon sequestration data, the method obtains the completeness of the multi-source data collaborative representation, which is used to measure the ability of multi-source remote sensing data to represent farmland carbon sequestration parameters.

[0024] It should be noted that the multi-source remote sensing data and ground-measured farmland carbon sink data include boundary position offset, pixel information entropy, and pixel-point spatial matching error. Specifically, the true boundary vector polygons of the fields are obtained through RTK-GPS field measurements. From the fused remote sensing data products (such as a 10m resolution carbon sink distribution map), the field boundary vector polygons are extracted using image segmentation or edge detection algorithms (such as the Canny operator). The boundary position offset is calculated by comparing the true boundary vector polygons with the field boundary vector polygons using the average Hausdorff distance. By preprocessing the remote sensing images, such as radiometric correction and atmospheric correction, the probability distribution of pixels at different gray levels or spectral bands is calculated. The entropy value of each pixel is calculated based on Shannon information entropy theory to obtain the pixel information entropy. The precise geographic coordinates of the points are obtained using RTK-GPS and other equipment. The geographic coordinates of the pixel center point are calculated using the image's georeferenced information (such as affine transformation parameters in the GeoTIFF file). The deviation between the true coordinates and the pixel center coordinates is calculated using the Euclidean distance formula to obtain the pixel-point spatial matching error.

[0025] The offset influence value is obtained by interactively processing the ratio analysis results of the boundary position offset and the offset reference value through the offset influence factor; where the ratio analysis refers to division operation and the interactive operation refers to multiplication operation.

[0026] The information entropy influence value is obtained by interactively processing the analysis results of the ratio of pixel information entropy to information entropy reference value through the information entropy influence factor.

[0027] The influence value of matching error is obtained by interactively processing the analysis results of the proportion of pixel-point spatial matching error to matching error reference value through the matching error influence factor.

[0028] The completeness of the collaborative representation of multi-source data is obtained by coupling the influence values ​​of offset, information entropy, and matching error. Here, coupling refers to the addition operation.

[0029] It should be further noted that the offset influence factor, offset reference value, information entropy influence factor, information entropy reference value, matching error influence factor, and matching error reference value were all obtained from the spatiotemporal modeling database. Specifically, there is a correlation between boundary position offset, pixel information entropy, and pixel-to-point spatial matching error, as follows: These three factors are closely interrelated, jointly reflecting the accuracy and uncertainty of remote sensing data in spatial representation. Boundary position offset measures the geometric deviation between the remotely sensed plot boundary and the actual boundary. Its magnitude directly affects pixel information entropy because boundary blurring or offset can lead to multiple land cover types within a pixel, thus increasing pixel information entropy, i.e., increasing uncertainty. Simultaneously, pixel-to-point spatial matching error reflects the spatial consistency between a specific pixel in the remote sensing data and the measured ground point. A large matching error indicates inaccurate spatial positioning of the remote sensing data, which further exacerbates boundary offset and leads to increased pixel information entropy. Therefore, these three parameters influence each other and together constitute an important indicator system for evaluating the spatial representation accuracy and reliability of remote sensing data. Among them, the boundary position offset is a direct reflection of spatial geometric accuracy, the pixel information entropy is a quantitative indicator of the degree of information mixing, and the pixel-point spatial matching error is a key reflection of spatial positioning accuracy. Together, they determine the applicability and reliability of remote sensing data in farmland carbon sink modeling. Meanwhile, boundary position offset, pixel information entropy, and pixel-to-point spatial matching error are correlated with the completeness of multi-source data collaborative representation, specifically as follows: Boundary position offset, pixel information entropy, and pixel-to-point spatial matching error are positively correlated with the completeness of multi-source data collaborative representation. A larger boundary offset indicates a greater error in the spatial positioning of remote sensing data, leading to significant discrepancies between the extracted field location, shape, and size and the actual situation. High information entropy means that a pixel contains multiple land cover types (such as soil, vegetation, and water bodies), i.e., there is a mixed pixel problem. This makes it difficult for remote sensing data to accurately reflect the spectral characteristics of a single land cover, affecting classification and inversion accuracy. Large pixel-to-point matching error indicates inaccurate spatial positioning of remote sensing data, making the verification results unreliable.

[0030] As the second step in the spatiotemporal modeling method of farmland carbon sequestration remote sensing data based on machine learning: determine whether to perform completeness optimization based on the completeness of the collaborative representation of multi-source data. If yes, construct a training sample dataset for machine learning after completeness optimization; otherwise, directly construct a training sample dataset for machine learning. Completeness optimization includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold.

[0031] It is important to understand that field boundaries are the core benchmark for defining the spatial range of carbon sink parameters (such as soil organic carbon and bulk density). The rate of change of these boundaries directly determines the spatial matching accuracy between remote sensing data and ground-measured data. Therefore, the extraction frequency of field boundaries can be adjusted based on the rate of boundary change. By dynamically adjusting the extraction frequency to match the rate of boundary change, we can track rapidly changing boundaries in a timely manner through high-frequency observations, ensuring the spatial benchmark consistency between remote sensing data and measured carbon sink parameters. At the same time, we can avoid redundant data collection when the boundaries are stable, balancing data timeliness and acquisition costs, and providing accurate spatial range basis for subsequent remote sensing modeling of farmland carbon sinks.

[0032] Furthermore, if the completeness of the multi-source data collaborative representation is less than or equal to the completeness setting value, no completeness optimization is performed; if the completeness of the multi-source data collaborative representation is greater than the completeness setting value, it is determined whether to perform dynamic adjustment of the extraction frequency based on the boundary change rate. If so, it is determined whether to perform dynamic adjustment of the area representation threshold after dynamic adjustment of the extraction frequency; otherwise, it is determined directly whether to perform dynamic adjustment of the area representation threshold.

[0033] As a further explanation, the specific steps for determining whether to perform dynamic frequency adjustment are as follows:

[0034] If the boundary change rate is greater than the change rate reference value, the harmonic average of the change rate correction and the integrity correction is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency adjustment factor. It is then determined whether the extraction frequency adjustment factor is greater than the set value. By associating the boundary change rate deviation with the harmonic average of the integrity deviation of multi-source data, the extraction frequency adjustment factor is extracted and a threshold judgment is made. This enables precise dynamic control of the field boundary extraction frequency. It captures rapidly changing boundary information in a timely manner through high-frequency extraction to maintain the consistency of the data spatial benchmark, while avoiding redundancy caused by over-extraction in combination with data integrity requirements. Ultimately, it ensures the timeliness and reliability of the boundary data required for subsequent farmland carbon sink remote sensing modeling, laying a spatial benchmark foundation for accurate carbon sink parameter inversion. The change rate correction represents the difference between the boundary change rate and the change rate reference value, and the integrity correction represents the difference between the integrity of the multi-source data collaborative representation and the integrity set value.

[0035] If the rate of change of the boundary is less than or equal to the reference value of the rate of change, no dynamic adjustment of the extraction frequency will be performed. This can ensure that the field boundary data is consistent with the actual spatial benchmark and meet the spatial accuracy requirements of subsequent farmland carbon sequestration remote sensing modeling, while avoiding unnecessary adjustments to the extraction frequency that would lead to redundant acquisition and processing costs of remote sensing data, thus achieving a balance between data acquisition efficiency and modeling accuracy.

[0036] As further explained in detail, the specific steps for determining whether the extraction frequency adjustment factor is greater than the set value of the extraction frequency adjustment factor are as follows:

[0037] If the extraction frequency adjustment factor is greater than the set value, the correction amount of the extraction frequency adjustment factor is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency increase. Based on the current field boundary extraction frequency and the extraction frequency increase, the adjusted field boundary extraction frequency is obtained by superimposing the two. This can accurately quantify the extraction density that needs to be increased in scenarios with rapid changes in field boundaries. This ensures that high-frequency observations can track boundary dynamics in a timely manner to maintain the consistency of the data spatial benchmark. At the same time, the correction amount controls the increase to avoid data redundancy and increased processing costs caused by over-extraction. This provides boundary data support that is both timely and economical for subsequent remote sensing modeling of farmland carbon sequestration. The correction amount of the extraction frequency adjustment factor represents the positive difference between the extraction frequency adjustment factor and the set value of the extraction frequency adjustment factor.

[0038] If the extraction frequency adjustment factor is less than or equal to the set value of the extraction frequency adjustment factor, the comparison value of the extraction frequency adjustment factor is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency reduction amount. Based on the current field boundary extraction frequency and the extraction frequency reduction amount, the reduction processing is performed to obtain the adjusted field boundary extraction frequency. This can accurately quantify the extraction density that needs to be reduced in scenarios where the field boundary changes slowly. This ensures that the boundary data can still match the actual spatial benchmark and meet the accuracy requirements of subsequent farmland carbon sequestration remote sensing modeling. At the same time, by reducing the extraction frequency, unnecessary remote sensing data acquisition and processing costs are reduced, achieving a balance between the economy of data acquisition and the reliability of modeling. The comparison value of the extraction frequency adjustment factor represents the negative difference between the extraction frequency adjustment factor and the set value of the extraction frequency adjustment factor.

[0039] In this embodiment, focusing on the dual requirements of "spatiotemporal accuracy" and "cost-effectiveness" for boundary data in farmland carbon sequestration remote sensing modeling, a boundary data quality assurance system with full-scene coverage and multi-dimensional collaboration was constructed. Specifically, by using the boundary change rate as the core trigger condition and linking the integrity of multi-source data collaborative representation to construct the control logic, the system ultimately achieved full-cycle control of dynamic scene accurate tracking and stable scene cost optimization of field boundary data. This provides timely boundary dynamic response and spatially consistent boundary data support for subsequent spatiotemporal modeling of farmland carbon sequestration remote sensing data, effectively ensuring the accuracy of carbon sequestration parameter (such as soil organic carbon storage and vegetation biomass carbon) inversion.

[0040] It should be noted that the remote sensing inversion shoreline land parcel area representation threshold is the core standard for determining whether remote sensing pixels are included in the target land parcel, while the land parcel area deviation rate refers to the degree of deviation between the remote sensing inverted area and the RTK-GPS measured area, which directly reflects the adaptability of the threshold to the actual land parcel characteristics. Therefore, the area representation threshold can be adjusted based on the land parcel area deviation rate.

[0041] like Figure 2The diagram shows a flowchart of the dynamic adjustment process for the area representation threshold of the spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning, as provided in this application embodiment. The specific logic is as follows: If the plot area deviation rate is greater than the deviation rate reference value, the result of harmonic averaging the deviation rate correction and integrity correction is input into the area representation threshold mapping table for index lookup to obtain the representation threshold adjustment factor. It is then determined whether the representation threshold adjustment factor is greater than the set value. If so, the representation threshold adjustment factor correction is input into the area representation threshold mapping table for index lookup to obtain the reduction amount of the remote sensing inversion shoreline plot area representation threshold. Based on the current... The remote sensing inversion shoreline plot area representation threshold and the reduction amount of the remote sensing inversion shoreline plot area representation threshold are reduced to obtain the adjusted remote sensing inversion shoreline plot area representation threshold. If not, the representation threshold adjustment factor compensation amount is input into the area representation threshold mapping table for index lookup to obtain the increase amount of the remote sensing inversion shoreline plot area representation threshold. The current remote sensing inversion shoreline plot area representation threshold and the increase amount of the remote sensing inversion shoreline plot area representation threshold are superimposed to obtain the adjusted remote sensing inversion shoreline plot area representation threshold. If the plot area deviation rate is less than or equal to the deviation rate reference value, no dynamic adjustment of the area representation threshold is performed.

[0042] It should be added that the specific steps for determining whether to perform dynamic adjustment of the area representation threshold are as follows:

[0043] If the plot area deviation rate is greater than the deviation rate reference value, the result of harmonic averaging of the deviation rate correction and integrity correction is input into the area characterization threshold mapping table for index lookup to obtain the characterization threshold control factor. It is then determined whether the characterization threshold control factor is greater than the set value, which can accurately trigger the dynamic control decision of the shoreline plot area characterization threshold. This not only addresses the core issue of area inversion deviation to ensure that the plot area extracted by subsequent remote sensing matches the measured area, but also avoids the distortion of carbon sink parameter characterization caused by blind threshold adjustment in combination with data integrity requirements. This provides accurate plot area benchmark data support for subsequent farmland carbon sink remote sensing modeling. The deviation rate correction represents the difference between the plot area deviation rate and the deviation rate reference value.

[0044] As a further specific explanation, determining whether to perform dynamic adjustment of the area representation threshold also includes:

[0045] If the plot area deviation rate is less than or equal to the deviation rate reference value, no dynamic adjustment of the area characterization threshold will be performed. This ensures accurate matching between the shoreline plot area retrieved by remote sensing and the RTK-GPS measured area, avoiding unnecessary threshold adjustments that could lead to disordered plot boundary extraction logic (such as mistakenly including non-target water bodies or roads within the plot area). This reduces the cost of reprocessing remote sensing data and the risk of data deviation in carbon sink modeling, achieving a balance between the stability of plot area data acquisition and the accuracy of carbon sink parameter calculation.

[0046] The specific steps to determine whether the characterization threshold control factor is greater than the set value of the characterization threshold control factor are as follows:

[0047] If the characterization threshold adjustment factor is greater than the set value, the correction amount of the characterization threshold adjustment factor is input into the area characterization threshold mapping table for index lookup to obtain the reduction amount of the remote sensing inversion shoreline plot area characterization threshold. Based on the current remote sensing inversion shoreline plot area characterization threshold and the reduction amount of the remote sensing inversion shoreline plot area characterization threshold, the reduction processing is performed to obtain the adjusted remote sensing inversion shoreline plot area characterization threshold. This can accurately quantify the threshold range that needs to be reduced in scenarios where the plot area deviation rate exceeds the standard. This not only eliminates interference from non-target areas (such as water bodies and roads around narrow plots in polder river mouths) through threshold reduction, ensuring the spatial benchmark accuracy of subsequent carbon sink parameter calculations, but also controls the reduction magnitude through correction amount to avoid excessive threshold reduction leading to the omission of effective plot areas (such as excluding key near-shore carbon sink areas). This provides plot area data support with both accuracy and completeness for farmland carbon sink remote sensing modeling. The correction amount of the characterization threshold adjustment factor represents the positive difference between the characterization threshold adjustment factor and the set value of the characterization threshold adjustment factor.

[0048] If the characterization threshold adjustment factor is less than or equal to the set value, the compensation amount of the characterization threshold adjustment factor is input into the area characterization threshold mapping table for index lookup to obtain the increase in the remote sensing inversion shoreline plot area characterization threshold. Based on the current remote sensing inversion shoreline plot area characterization threshold and the increase in the remote sensing inversion shoreline plot area characterization threshold, the adjusted remote sensing inversion shoreline plot area characterization threshold is obtained. This can accurately quantify the threshold range that needs to be expanded in scenarios with low plot area deviation rates. It avoids the omission of effective shoreline areas (such as the near-shore artificial turf area of ​​the Liangtang River estuary and the natural turf coverage area of ​​the polder river estuary) due to the original threshold being too small, ensuring that the remote sensing inversion area completely covers the measured plot range and ensuring the plot integrity of carbon sink parameters (such as vegetation biomass carbon storage). It also avoids the excessive expansion of the threshold by controlling the increase amount, which would cause non-target areas (such as adjacent roads and scattered water bodies) to be mixed in. This provides plot area benchmark data with both completeness and accuracy for farmland carbon sink remote sensing modeling. The compensation amount of the threshold control factor represents the negative difference between the threshold control factor and the set value of the threshold control factor.

[0049] In this embodiment, focusing on the core requirement of "accurately matching measured data and fully covering the carbon sink area" for farmland carbon sink modeling, a full-process control system of "deviation triggering - comprehensive evaluation - hierarchical regulation" was constructed to achieve dual assurance of the accuracy of farmland area data and the reliability of carbon sink modeling. Ultimately, the goal of "accurate correction of deviations exceeding standards, controllable and stable maintenance of deviations, and complete coverage of effective areas" for shoreline farmland area inversion was achieved, providing a farmland area benchmark with accuracy, completeness, and stability for spatiotemporal modeling of farmland carbon sink remote sensing data.

[0050] The third step in the machine learning-based spatiotemporal modeling method for farmland carbon sequestration remote sensing data is to train a spatiotemporal model of farmland carbon sequestration remote sensing data using a machine learning training sample dataset as input. This process obtains the carbon cycle-remote sensing data coupling degree, which is used to measure the coupling degree between the carbon sequestration process model and the remote sensing data during the carbon cycle.

[0051] It should be noted that carbon cycle-remote sensing data includes the completeness of multi-source data collaborative representation, lidar-optical data terrain matching bias, and remote sensing-carbon cycle pixel scale bias. Among these, the registration between lidar and optical imagery is usually achieved through matching ground control points or feature points. During the registration process, there is a spatial offset between the lidar point cloud and the optical imagery, which will produce terrain matching bias. This bias can be obtained by calculating the geometric offset between the two. The remote sensing-carbon cycle pixel scale bias is obtained by extending ground-measured data (such as sample plot surveys) to the remote sensing pixel scale using machine learning methods.

[0052] The specific steps to obtain the coupling degree of the carbon cycle process and remote sensing data are as follows:

[0053] The integrity impact value is obtained by interactively processing the results of the proportion analysis of the integrity of multi-source data collaborative characterization and the integrity reference value through the integrity impact factor; where the proportion analysis refers to the division operation and the interactive processing refers to the multiplication operation.

[0054] The influence value of terrain matching deviation is obtained by interactively processing the analysis results of the ratio of terrain matching deviation to the reference value of LiDAR-optical data through terrain matching deviation influence factor.

[0055] The scale deviation impact value was obtained by interactively processing the results of the analysis of the proportion of scale deviation and scale deviation reference value of remote sensing carbon cycle pixels by the scale deviation impact factor.

[0056] The carbon cycle process-remote sensing data coupling degree is obtained by coupling the integrity influence value, the terrain matching deviation influence value, and the scale deviation influence value. Here, coupling refers to the additive operation.

[0057] It should be further noted that the completeness impact factor, completeness reference value, terrain matching deviation impact factor, terrain matching deviation reference value, scale deviation impact factor, and scale deviation reference value were all obtained from the spatiotemporal modeling database. Specifically, there is a correlation between the completeness of multi-source data collaborative representation, the terrain matching deviation of lidar-optical data, and the pixel scale deviation of remote sensing-carbon cycle, as follows: there is a close mutual influence among the completeness of multi-source data collaborative representation, the terrain matching deviation of lidar-optical data, and the pixel scale deviation of remote sensing-carbon cycle, which together determine the comprehensive representation capability of remote sensing data in farmland carbon sink modeling. The completeness of multi-source data collaborative representation refers to the degree of matching between different remote sensing data sources (such as optical, microwave, and lidar) in terms of spatiotemporal reference, physical meaning, and model coupling. Its level directly affects the terrain matching deviation between lidar and optical data, as this deviation stems from the geometric offset during spatial registration between lidar point clouds and optical images. Inconsistencies in projection coordinate systems, pixel sizes, and terrain undulation processing among multi-source data will increase terrain matching deviation, thus improving the completeness of collaborative representation. Simultaneously, the pixel scale deviation between remote sensing and carbon cycle reflects the mismatch between the remote sensing pixel scale and the carbon cycle process scale. A higher level of multi-source data collaborative representation completeness, for example, if pixel scale effects are not considered during data fusion, will exacerbate pixel scale deviation, leading to distortion in the spatial distribution of carbon sink parameters and affecting the accurate modeling of the carbon cycle process. These three factors collectively determine the reliability and applicability of multi-source remote sensing data in farmland carbon sink monitoring. Meanwhile, there is a positive correlation between the completeness of multi-source data collaborative representation, the terrain matching deviation between lidar and optical data, and the pixel scale deviation between remote sensing and carbon cycle data, and the coupling degree between carbon cycle process and remote sensing data. Specifically: the greater the completeness of multi-source data collaborative representation, the worse the matching degree of different remote sensing data sources (such as optical, microwave, lidar, etc.) in terms of spatiotemporal reference, physical meaning, and model coupling; the greater the terrain matching deviation between lidar and optical data, the less accurate the spatial alignment between lidar and optical imagery, resulting in serious errors in the extraction of terrain parameters such as terrain undulation, slope, and aspect; the greater the pixel scale deviation between remote sensing and carbon cycle data, the less accurately remote sensing pixels can capture the detailed features of the carbon cycle process, such as small-scale carbon flux changes and vegetation structure heterogeneity.

[0058] The fourth step in the machine learning-based spatiotemporal modeling method for farmland carbon sequestration remote sensing data is to determine whether to perform coupling optimization based on the coupling degree between the carbon cycle process and remote sensing data. If so, the spatiotemporal dynamic distribution map of farmland carbon sequestration in the target farmland area is output after coupling optimization. If not, the spatiotemporal dynamic distribution map of farmland carbon sequestration in the target farmland area is output directly. Coupling optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of adaptive resolution.

[0059] Furthermore, if the coupling degree between the carbon cycle process and remote sensing data is less than or equal to the set coupling degree value, no coupling degree optimization is performed; if the coupling degree between the carbon cycle process and remote sensing data is greater than the set coupling degree value, it is determined whether to perform dynamic adjustment of the pixel alignment window based on the pixel scale deviation between remote sensing and carbon cycle. If yes, it is determined whether to perform dynamic adjustment of the adaptation resolution after dynamic adjustment of the pixel alignment window; otherwise, it is determined directly whether to perform dynamic adjustment of the adaptation resolution.

[0060] It should be noted that the multi-source data cell alignment window is the core tool for achieving cell-level spatial matching of multi-source remote sensing data such as lidar, optical, and microwave. Its size needs to be precisely adapted to the actual spatial scale of the carbon cycle process. The remote sensing-carbon cycle cell scale deviation (the degree of mismatch between the remote sensing cell scale and the carbon cycle micro-plot scale) directly reflects the adaptation defects of the current alignment window to the spatial characteristics of the carbon cycle process. Therefore, the cell alignment window can be adjusted based on this deviation.

[0061] like Figure 3 The diagram shows a flowchart of the dynamic adjustment process for the pixel alignment window in the spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning provided in this application embodiment. The specific logic is as follows: If the remote sensing-carbon cycle pixel scale deviation is less than or equal to the scale deviation reference value, then no dynamic adjustment of the pixel alignment window is performed; if the remote sensing-carbon cycle pixel scale deviation is greater than the scale deviation reference value, then the scale deviation correction amount and coupling degree correction amount are input into the pixel alignment window mapping table for index lookup to obtain the alignment window adjustment factor. It is then determined whether the alignment window adjustment factor is greater than the set value of the alignment window adjustment factor. If so, the alignment window adjustment factor is adjusted. The sub-correction value is input into the cell alignment window mapping table for index lookup to obtain the reduction amount of the multi-source data cell alignment window width. Based on the current multi-source data cell alignment window and the reduction amount of the multi-source data cell alignment window width, a reduction process is performed to obtain the adjusted multi-source data cell alignment window. If not, the alignment window adjustment factor is input into the cell alignment window mapping table for index lookup to obtain the expansion amount of the multi-source data cell alignment window width. Based on the current multi-source data cell alignment window and the expansion amount of the multi-source data cell alignment window width, a superposition process is performed to obtain the adjusted multi-source data cell alignment window.

[0062] As a further explanation, the specific steps for determining whether to perform dynamic adjustment of the pixel alignment window are as follows:

[0063] If the remote sensing-carbon cycle pixel scale deviation is less than or equal to the scale deviation reference value, then no dynamic adjustment of the pixel alignment window is performed. This ensures that the current pixel alignment window size is accurately matched with the actual spatial scale of the carbon cycle process, and that the pixel-level alignment error of multi-source data meets the requirements. This, in turn, ensures that the inversion deviation rate of carbon sink parameters (such as soil organic carbon and bulk density) and the coupling degree of the carbon cycle process-remote sensing data meet the accuracy requirements of farmland carbon sink modeling. At the same time, it avoids unnecessary window adjustments that could cause alignment logic disorder (such as mis-splitting of carbon cycle homogeneous micro-plots or mixing in heterogeneous regional data), reduces the computational cost and time loss of multi-source data realignment, and achieves a balance between data processing efficiency and carbon sink modeling reliability.

[0064] If the scale deviation of remote sensing-carbon cycle pixels is greater than the scale deviation reference value, the scale deviation correction and coupling correction are input into the pixel alignment window mapping table for index lookup to obtain the alignment window adjustment factor. It is then determined whether the alignment window adjustment factor is greater than the set value, enabling accurate assessment of the pixel alignment window adjustment needs. This avoids blind window adjustment based solely on scale deviation (such as over-adjustment ignoring scenarios where coupling meets standards), and uses dual-factor collaborative analysis to pinpoint the core problem of "scale mismatch + insufficient coupling," providing a scientific basis for subsequent targeted window size adjustments. This ensures that the adjusted window matches the carbon cycle micro-plot scale and eliminates pixel scale deviation, laying a reliable spatial foundation for accurate multi-source data alignment and farmland carbon sink parameter inversion. The scale deviation correction represents the difference between the remote sensing-carbon cycle pixel scale deviation and the scale deviation reference value, while the coupling correction represents the difference between the carbon cycle process-remote sensing data coupling degree and the set coupling degree value.

[0065] If the alignment window adjustment factor is greater than the set value, the alignment window adjustment factor correction amount is input into the pixel alignment window mapping table for index lookup to obtain the reduction amount of the multi-source data pixel alignment window width. Based on the current multi-source data pixel alignment window and the reduction amount of the multi-source data pixel alignment window width, a reduction process is performed to obtain the adjusted multi-source data pixel alignment window. This can accurately quantify the window width that needs to be reduced in scenarios where the remote sensing-carbon cycle pixel scale deviation exceeds the standard. This is achieved by reducing the window size to accurately match the carbon cycle micro-plot scale, avoiding confusion in carbon sink parameter inversion caused by the original large window covering heterogeneous areas. At the same time, the reduction magnitude is controlled by the correction amount to prevent excessive window reduction from causing data noise interference (such as a decrease in the accuracy of lidar topographic data). This ensures that the multi-source data pixel-level alignment error and the coupling degree of the carbon cycle process-remote sensing data both meet the requirements of farmland carbon sink modeling, providing spatially consistent multi-source data support for accurate carbon sink parameter inversion. The alignment window adjustment factor correction amount represents the positive difference between the alignment window adjustment factor and the set value of the alignment window adjustment factor.

[0066] As a further explanation, determining whether the alignment window adjustment factor is greater than the set value of the alignment window adjustment factor also includes:

[0067] If the alignment window adjustment factor is less than or equal to the set value, the alignment window adjustment factor comparison value is input into the pixel alignment window mapping table for index lookup to obtain the expansion amount of the multi-source data pixel alignment window width. Based on the current multi-source data pixel alignment window and the expansion amount of the multi-source data pixel alignment window width, the adjusted multi-source data pixel alignment window is obtained. This can accurately quantify the window width that needs to be expanded in scenarios where the remote sensing-carbon cycle pixel scale deviation is small but data consistency needs to be optimized. By expanding the window, the adjusted window can integrate more homogeneous pixel information (such as lidar topographic data and optical remote sensing vegetation index data), reduce noise interference from single pixels, and avoid over-expansion that would cause the window to cover heterogeneous carbon cycle micro-plots. This ensures that the adjusted window still matches the actual scale of the carbon cycle process, maintains the coupling degree of the carbon cycle process-remote sensing data and the alignment accuracy of multi-source data pixels, and provides a multi-source data alignment foundation with both data stability and spatial adaptability for farmland carbon sink parameter inversion. The alignment window adjustment factor comparison value represents the negative difference between the alignment window adjustment factor and the set value of the alignment window adjustment factor.

[0068] In this embodiment, focusing on the core requirements of "spatial consistency of multi-source data" and "scale adaptability of carbon cycle" for farmland carbon sink modeling, a full-scenario precise control system of "deviation triggering - dual-factor evaluation - hierarchical regulation" was constructed to achieve deep coupling between multi-source data alignment accuracy and carbon sink modeling reliability. Ultimately, the system achieves full-cycle management of the pixel alignment window, including "efficient maintenance of stable scenarios, precise control of deviant scenarios, and optimized integration of noisy scenarios." This ensures that the pixel-level alignment of multi-source data adapts to the actual spatial scale of the carbon cycle process while maintaining high data consistency, providing a reliable spatial foundation for accurate inversion of farmland carbon sink parameters and efficient coupling of carbon cycle and remote sensing data.

[0069] It should be noted that the carbon cycle process scale adaptation resolution is a key scale parameter that connects carbon cycle mechanism models (such as organic matter decomposition and vegetation carbon sequestration models) with remote sensing data. It needs to be accurately matched with the spatial heterogeneity characteristics of the carbon cycle process. The carbon process-remote sensing data coupling residual represents the deviation between the simulated values ​​of the carbon cycle model and the remote sensing inversion values. It directly reflects the dual adaptation defects of the current adaptation resolution with the carbon cycle process and remote sensing data. Therefore, the carbon cycle process scale adaptation resolution can be adjusted based on this coupling residual.

[0070] As a further explanation, the specific steps for determining whether to perform dynamic resolution adjustment are as follows:

[0071] If the carbon process-remote sensing data coupling residual is less than or equal to the coupling residual reference value, then no dynamic adjustment of the adaptation resolution is performed. This ensures that the current adaptation resolution matches both the actual spatial scale of the carbon cycle process and the pixel scale of the multi-source remote sensing data. This ensures that the deviation rate between the simulated values ​​of the carbon cycle model and the remote sensing inversion values ​​meets the accuracy requirements for farmland carbon sink modeling. At the same time, it avoids unnecessary resolution adjustments that could lead to problems such as carbon cycle model parameter reconstruction (e.g., recalibrating the organic matter decomposition rate coefficient) and repeated resampling of remote sensing data. This reduces data processing costs and the risk of modeling process disorder, achieving a balance between the stability of carbon cycle process modeling and the synergy of remote sensing data.

[0072] If the carbon process-remote sensing data coupling residual is greater than the coupling residual reference value, the coupling residual correction and coupling degree correction are input into the adaptation resolution mapping table for index lookup to obtain the resolution adjustment factor. It is then determined whether the resolution adjustment factor is greater than the set value of the resolution adjustment factor. This enables accurate judgment of the carbon cycle process scale adaptation resolution adjustment needs. This avoids blind resolution adjustment based solely on the coupling residual (such as over-adjustment in scenarios where the coupling degree meets the standard), and also uses dual-factor collaborative analysis to locate the core contradiction of "residual exceeding the standard + insufficient coupling degree". This provides a scientific basis for subsequent targeted resolution adjustments, ensuring that the adjusted resolution can reduce the simulated value of the carbon cycle model. The coupling residual correction represents the difference between the carbon process-remote sensing data coupling residual and the coupling residual reference value.

[0073] As a further explanation, the specific steps to determine whether the resolution control factor is greater than the set value of the resolution control factor are as follows:

[0074] If the resolution adjustment factor is greater than the set value, the correction amount of the resolution adjustment factor is input into the adaptive resolution mapping table for index lookup to obtain the carbon cycle process scale adaptive resolution up-adjustment amount. Based on the current carbon cycle process scale adaptive resolution and the carbon cycle process scale adaptive resolution up-adjustment amount, the adjusted carbon cycle process scale adaptive resolution is obtained. This up-adjustment ensures that the adjusted resolution can accurately characterize the micro-scale heterogeneity of the carbon cycle, avoiding excessive coupling residuals caused by the original low-resolution smoothing heterogeneity. At the same time, the correction amount controls the up-adjustment magnitude to prevent excessive resolution increase from causing a disconnect with the scale of multi-source remote sensing data and increasing data resampling errors. This ensures that the adjusted resolution is both adapted to the spatial characteristics of the carbon cycle process and coordinated with the pixel scale of remote sensing data, providing a reliable basis for accurate inversion of farmland carbon sink parameters and carbon cycle modeling with "process-data" scale matching. The resolution adjustment factor correction amount represents the positive difference between the resolution adjustment factor and the set value of the resolution adjustment factor.

[0075] If the resolution adjustment factor is less than or equal to the set value, the reference value of the resolution adjustment factor is input into the adaptive resolution mapping table for index lookup to obtain the carbon cycle process scale-adapted resolution reduction amount. Based on the current carbon cycle process scale-adapted resolution and the carbon cycle process scale-adapted resolution reduction amount, a reduction process is performed to obtain the adjusted carbon cycle process scale-adapted resolution. This can accurately quantify the resolution accuracy that needs to be reduced in scenarios where the carbon process-remote sensing data coupling residual is small but data synergy needs to be optimized. This means that by reducing the resolution, the adjusted resolution is coordinated with the pixel scale of multi-source remote sensing data, reducing frequent resampling errors caused by the mismatch between high resolution and remote sensing data scale. At the same time, the reference value controls the reduction magnitude to avoid excessive resolution reduction and smooth the micro-scale heterogeneity of the carbon cycle. This ensures that the adjusted resolution can maintain the spatial feature characterization accuracy of the carbon cycle process and improve the coupling degree of the carbon cycle process-remote sensing data, providing a scale basis for balancing "process characterization accuracy" and "data synergy" for farmland carbon sink modeling. The reference value of the resolution adjustment factor represents the negative difference between the resolution adjustment factor and the set value of the resolution adjustment factor.

[0076] In this embodiment, addressing the dual requirements of "accurate characterization of the carbon cycle process" and "efficient collaboration of remote sensing data" in farmland carbon sink modeling, a full-scenario scale optimization system of "residual triggering - dual-factor evaluation - hierarchical regulation" was constructed to achieve a deep unification of carbon cycle modeling accuracy and multi-source data synergy. Ultimately, it achieved full-cycle management of "stable scene maintenance and efficient operation at adaptable resolution, accurate optimization of residual exceedances, and dynamic balance of data collaboration," ensuring that the carbon cycle process characterization accurately reflects spatial heterogeneity while maintaining high synergy with remote sensing data, providing reliable scale support for farmland carbon sink parameter inversion and deep coupling of carbon cycle and remote sensing data.

[0077] like Figure 4The diagram illustrates the structure of a machine learning-based spatiotemporal modeling system for farmland carbon sequestration remote sensing data, as provided in this embodiment of the invention. This system includes: a completeness quantification module, a completeness optimization module, a coupled quantification module, and a coupledness optimization module. The completeness quantification module acquires multi-source remote sensing data of the target farmland area within a preset time series and simultaneously collects ground-measured farmland carbon sequestration data of the corresponding time series. Based on the multi-source remote sensing data and the ground-measured farmland carbon sequestration data, it obtains the completeness of the multi-source data collaborative representation, which measures the ability of the multi-source remote sensing data to represent farmland carbon sequestration parameters. The completeness optimization module determines whether to perform completeness optimization based on the completeness of the multi-source data collaborative representation. If so, it constructs a completeness model for machine learning after the completeness optimization. If no training sample dataset is available, a training sample dataset for machine learning is directly constructed. Completeness optimization includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold. Coupling quantification module: used as input of the machine learning training sample dataset to train the spatiotemporal model of farmland carbon sink remote sensing data, obtain carbon cycle-remote sensing data coupling degree to measure the coupling degree between the carbon sink process model and remote sensing data during the carbon cycle. Coupling degree optimization module: used to determine whether to perform coupling degree optimization based on the carbon cycle-remote sensing data coupling degree. If yes, output the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area after coupling degree optimization. If no, directly output the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area. Coupling degree optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of adaptive resolution.

[0078] In this embodiment, by constructing four core modules—complete quantification and optimization, coupled quantification and optimization—a progressive dual-optimization closed loop is formed, ultimately achieving high precision, high mechanistic consistency, and strong generalization capability in farmland carbon sink remote sensing monitoring. Specifically, the complete quantification module first extracts a collaborative representation completeness index from multi-source remote sensing and ground-based measured data to objectively assess the data foundation's ability to represent carbon sink parameters. The completeness optimization module dynamically adjusts the data extraction frequency and area representation threshold based on this index, ensuring the sufficiency and representativeness of input features in the spatiotemporal dimensions through iterative optimization, thus improving data quality from the source. Furthermore, the coupled quantification module trains a spatiotemporal model on a high-quality dataset and calculates its mechanistic coupling degree with the carbon cycle process model, evaluating the physical rationality of the model's output. The coupling degree optimization module dynamically adjusts the pixel alignment window and adaptation resolution accordingly, deeply integrating remote sensing inversion and process mechanisms to ensure that the model is not only data-driven but also conforms to ecological and physiological laws. Ultimately, the spatiotemporal dynamic distribution map of farmland carbon sinks output by the system combines the advantages of large-area coverage of remote sensing technology, the high-precision nonlinear fitting capability of machine learning, and the mechanistic interpretability of the process model. It effectively solves the core problems of insufficient data representation, large prediction bias caused by lack of mechanistic constraints, and weak generalization ability in traditional remote sensing carbon sink models, and provides reliable technical support for precision agricultural management and carbon sink accounting.

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning, characterized in that, Includes the following steps: Acquire multi-source remote sensing data of the target farmland area within a preset time series, and simultaneously collect ground-measured farmland carbon sink data of the target farmland area within the corresponding time series. Based on the multi-source remote sensing data and the ground-measured farmland carbon sink data, obtain the completeness of the multi-source data collaborative representation, which is used to measure the ability of multi-source remote sensing data to represent farmland carbon sink parameters. Whether to perform completeness optimization is determined based on the completeness of the collaborative representation of multi-source data. If yes, a training sample dataset for machine learning is constructed after completeness optimization. If no, a training sample dataset for machine learning is constructed directly. The completeness optimization includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold. Using a training sample dataset based on machine learning as input, a spatiotemporal model of farmland carbon sink remote sensing data is trained. The carbon cycle-remote sensing data coupling degree is obtained, which is used to measure the coupling degree between the carbon sink process model and the remote sensing data in the carbon cycle process. The coupling degree of the carbon cycle process and remote sensing data is used to determine whether to perform coupling degree optimization. If so, the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is output after coupling degree optimization. If not, the spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area is output directly. The coupling degree optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of adaptive resolution. Multi-source remote sensing data and ground-measured farmland carbon sink data include boundary position offset, pixel information entropy, and pixel-point spatial matching error; The offset influence value is obtained by interactively processing the analysis results of the ratio of boundary position offset to offset reference value through offset influence factor. The information entropy influence value is obtained by interactively processing the analysis results of the ratio of pixel information entropy to information entropy reference value through information entropy influence factor; The influence value of matching error is obtained by interactively processing the analysis results of the proportion of pixel-point spatial matching error to matching error reference value through the matching error influence factor. The completeness of the collaborative representation of multi-source data is obtained by coupling the influence values ​​of offset, information entropy, and matching error. If the completeness of the multi-source data collaborative representation is less than or equal to the completeness setting value, then no completeness optimization will be performed; If the completeness of the multi-source data collaborative representation is greater than the completeness set value, then it is determined whether to perform dynamic adjustment of the extraction frequency based on the boundary change rate. If yes, then it is determined whether to perform dynamic adjustment of the area representation threshold after dynamic adjustment of the extraction frequency. If no, then it is determined whether to perform dynamic adjustment of the area representation threshold directly. The carbon cycle-remote sensing data includes the completeness of multi-source data collaborative characterization, the terrain matching deviation of lidar-optical data, and the pixel scale deviation of remote sensing-carbon cycle. The specific steps for obtaining the carbon cycle process-remote sensing data coupling degree are as follows: The integrity impact value is obtained by interactively processing the results of the analysis of the proportion of the integrity of multi-source data collaborative characterization and the integrity reference value through the integrity impact factor; The influence value of terrain matching deviation is obtained by interactively processing the analysis results of the ratio of terrain matching deviation to the reference value of LiDAR-optical data through terrain matching deviation influence factor. The scale deviation impact value was obtained by interactively processing the results of the analysis of the proportion of scale deviation and scale deviation reference value of remote sensing carbon cycle pixels by the scale deviation impact factor. The coupling degree of carbon cycle process-remote sensing data is obtained by coupling the integrity influence value, the terrain matching deviation influence value and the scale deviation influence value. If the coupling degree between the carbon cycle process and remote sensing data is less than or equal to the coupling degree setting value, then no coupling degree optimization will be performed; If the coupling degree between the carbon cycle process and remote sensing data is greater than the set coupling degree, then it is determined whether to perform dynamic adjustment of the pixel alignment window based on the pixel scale deviation between remote sensing and carbon cycle. If yes, then it is determined whether to perform dynamic adjustment of the adaptation resolution after the dynamic adjustment of the pixel alignment window. If no, then it is determined directly whether to perform dynamic adjustment of the adaptation resolution.

2. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 1, characterized in that, The specific steps for determining whether to perform dynamic adjustment of the extraction frequency are as follows: If the boundary change rate is greater than the change rate reference value, the result of the harmonic averaging of the change rate correction and the integrity correction is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency adjustment factor. It is then determined whether the extraction frequency adjustment factor is greater than the extraction frequency adjustment factor set value. The change rate correction is used to characterize the degree of deviation between the boundary change rate and the change rate reference value, and the integrity correction is used to characterize the degree of deviation between the integrity of the multi-source data collaborative characterization and the integrity set value. If the boundary change rate is less than or equal to the change rate reference value, then no dynamic frequency control will be performed.

3. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 2, characterized in that, The specific steps for determining whether the extraction frequency adjustment factor is greater than the set value of the extraction frequency adjustment factor are as follows: If the extraction frequency adjustment factor is greater than the set value of the extraction frequency adjustment factor, the correction amount of the extraction frequency adjustment factor is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency increase amount. The current field boundary extraction frequency and the extraction frequency increase amount are superimposed to obtain the adjusted field boundary extraction frequency. The correction amount of the extraction frequency adjustment factor is used to characterize the degree of positive deviation between the extraction frequency adjustment factor and the set value of the extraction frequency adjustment factor. If the extraction frequency adjustment factor is less than or equal to the set value of the extraction frequency adjustment factor, the control quantity of the extraction frequency adjustment factor is input into the extraction frequency mapping table for index lookup to obtain the extraction frequency reduction amount. Based on the current field boundary extraction frequency and the extraction frequency reduction amount, the reduction process is performed to obtain the adjusted field boundary extraction frequency. The control quantity of the extraction frequency adjustment factor is used to characterize the degree of negative deviation between the extraction frequency adjustment factor and the set value of the extraction frequency adjustment factor. The specific steps for determining whether to perform dynamic adjustment of the area representation threshold are as follows: If the plot area deviation rate is greater than the deviation rate reference value, the result of harmonic averaging of the deviation rate correction amount and the integrity correction amount is input into the area characterization threshold mapping table for index lookup to obtain the characterization threshold control factor. It is then determined whether the characterization threshold control factor is greater than the characterization threshold control factor set value. The deviation rate correction amount is used to characterize the degree of deviation between the plot area deviation rate and the deviation rate reference value.

4. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 3, characterized in that, The determination of whether to perform dynamic adjustment of the area representation threshold also includes: If the plot area deviation rate is less than or equal to the deviation rate reference value, no dynamic adjustment of the area characterization threshold will be performed. The specific steps for determining whether the characterization threshold control factor is greater than the set value of the characterization threshold control factor are as follows: If the characterization threshold adjustment factor is greater than the characterization threshold adjustment factor set value, the characterization threshold adjustment factor correction amount is input into the area characterization threshold mapping table for index lookup to obtain the remote sensing inversion shoreline plot area characterization threshold reduction amount. Based on the current remote sensing inversion shoreline plot area characterization threshold and the remote sensing inversion shoreline plot area characterization threshold reduction amount, a reduction process is performed to obtain the adjusted remote sensing inversion shoreline plot area characterization threshold. The characterization threshold adjustment factor correction amount is used to characterize the degree of positive deviation between the threshold adjustment factor and the characterization threshold adjustment factor set value. If the characterization threshold adjustment factor is less than or equal to the characterization threshold adjustment factor set value, the compensation amount of the characterization threshold adjustment factor is input into the area characterization threshold mapping table for index lookup to obtain the increase in the remote sensing inverted coastal land area characterization threshold. Based on the current remote sensing inverted coastal land area characterization threshold and the increase in the remote sensing inverted coastal land area characterization threshold, the adjusted remote sensing inverted coastal land area characterization threshold is obtained. The compensation amount of the characterization threshold adjustment factor is used to characterize the degree of negative deviation between the threshold adjustment factor and the characterization threshold adjustment factor set value.

5. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 1, characterized in that, The specific steps for determining whether to perform dynamic adjustment of the pixel alignment window are as follows: If the remote sensing-carbon cycle pixel scale deviation is less than or equal to the scale deviation reference value, then no dynamic adjustment of the pixel alignment window will be performed. If the scale deviation of the remote sensing-carbon cycle pixel is greater than the scale deviation reference value, the scale deviation correction amount and the coupling degree correction amount are input into the pixel alignment window mapping table for index lookup to obtain the alignment window adjustment factor. It is then determined whether the alignment window adjustment factor is greater than the alignment window adjustment factor set value. The scale deviation correction amount is used to characterize the degree of deviation between the remote sensing-carbon cycle pixel scale deviation and the scale deviation reference value, and the coupling degree correction amount is used to characterize the degree of deviation between the carbon cycle process-remote sensing data coupling degree and the coupling degree set value. If the alignment window adjustment factor is greater than the alignment window adjustment factor setting value, the alignment window adjustment factor correction amount is input into the pixel alignment window mapping table for index lookup to obtain the reduction amount of the multi-source data pixel alignment window width. Based on the current multi-source data pixel alignment window and the reduction amount of the multi-source data pixel alignment window width, a reduction process is performed to obtain the adjusted multi-source data pixel alignment window. The alignment window adjustment factor correction amount is used to characterize the degree of positive deviation between the alignment window adjustment factor and the alignment window adjustment factor setting value.

6. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 5, characterized in that, The step of determining whether the alignment window adjustment factor is greater than the set value of the alignment window adjustment factor also includes: If the alignment window adjustment factor is less than or equal to the alignment window adjustment factor setting value, the alignment window adjustment factor reference value is input into the cell alignment window mapping table for index lookup to obtain the width expansion of the multi-source data cell alignment window. The current multi-source data cell alignment window and the width expansion of the multi-source data cell alignment window are superimposed to obtain the adjusted multi-source data cell alignment window. The alignment window adjustment factor reference value is used to characterize the degree of negative deviation between the alignment window adjustment factor and the alignment window adjustment factor setting value. The specific steps for determining whether to perform dynamic resolution adjustment are as follows: If the carbon process-remote sensing data coupling residual is less than or equal to the coupling residual reference value, then no dynamic adjustment of the adaptive resolution will be performed. If the carbon process-remote sensing data coupling residual is greater than the coupling residual reference value, the coupling residual correction amount and the coupling degree correction amount are input into the adaptive resolution mapping table for index lookup to obtain the resolution adjustment factor. It is then determined whether the resolution adjustment factor is greater than the resolution adjustment factor set value. The coupling residual correction amount is used to characterize the degree of deviation between the carbon process-remote sensing data coupling residual and the coupling residual reference value.

7. The spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning according to claim 6, characterized in that, The specific steps for determining whether the resolution control factor is greater than the set value are as follows: If the resolution control factor is greater than the resolution control factor set value, the resolution control factor correction amount is input into the adaptation resolution mapping table for index lookup to obtain the carbon cycle process scale adaptation resolution up-adjustment amount. The current carbon cycle process scale adaptation resolution and the carbon cycle process scale adaptation resolution up-adjustment amount are superimposed to obtain the adjusted carbon cycle process scale adaptation resolution. The resolution control factor correction amount is used to characterize the positive deviation of the resolution control factor from the resolution control factor set value. If the resolution control factor is less than or equal to the resolution control factor set value, the resolution control factor control quantity is input into the adaptive resolution mapping table for index lookup to obtain the carbon cycle process scale adaptive resolution reduction amount. Based on the current carbon cycle process scale adaptive resolution and the carbon cycle process scale adaptive resolution reduction amount, a reduction process is performed to obtain the adjusted carbon cycle process scale adaptive resolution. The resolution control factor control quantity is used to characterize the degree of negative deviation between the resolution control factor and the resolution control factor set value.

8. A spatiotemporal modeling system for farmland carbon sequestration remote sensing data based on machine learning, wherein the system applies the spatiotemporal modeling method for farmland carbon sequestration remote sensing data based on machine learning as described in any one of claims 1-7, characterized in that... include: Complete metric module, completeness optimization module, coupling metric module, and coupling optimization module: The complete quantification module is used to acquire multi-source remote sensing data of the target farmland area within a preset time series, and simultaneously collect ground-measured farmland carbon sink data of the target farmland area in the corresponding time series. Based on the multi-source remote sensing data and the ground-measured farmland carbon sink data, the completeness of the multi-source data collaborative representation is obtained, which is used to measure the ability of multi-source remote sensing data to represent farmland carbon sink parameters. The completeness optimization module is used to determine whether to perform completeness optimization based on the completeness of the collaborative representation of multi-source data. If yes, a training sample dataset for machine learning is constructed after completeness optimization. If no, a training sample dataset for machine learning is directly constructed. The completeness optimization includes dynamic adjustment of extraction frequency and dynamic adjustment of area representation threshold. The coupling quantification module is used to train a spatiotemporal model of farmland carbon sink remote sensing data based on a training sample dataset of machine learning as input, and to obtain the carbon cycle-remote sensing data coupling degree to measure the coupling degree between the carbon sink process model and the remote sensing data during the carbon cycle process. The coupling optimization module is used to determine whether to perform coupling optimization based on the coupling degree of the carbon cycle process and remote sensing data. If so, it outputs a spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area after coupling optimization. If not, it directly outputs a spatiotemporal dynamic distribution map of farmland carbon sink in the target farmland area. The coupling optimization includes dynamic adjustment of pixel alignment window and dynamic adjustment of adaptive resolution.

Images