Method and system for measuring and calculating urban soil carbon sink based on multi-source data

Abstract

The invention discloses an urban soil carbon sink quantity measuring and calculating method and system based on multi-source data, and relates to the technical field of ecological environment monitoring, and the method comprises the steps: collecting and preprocessing multi-source original data, generating a multi-source feature data cube, analyzing the multi-source feature data cube according to a preset rule, and recognizing potential influence events, forming an event log list; on the basis of a weight generation rule, fusion weights for different data source features are generated for each influenced space grid unit, and a dynamic fusion weight set is formed; performing weighted fusion on the features in the multi-source feature data cube by using the dynamic fusion weight set to generate a context enhanced feature vector; and inputting the situation enhancement feature vector into a pre-trained soil carbon sink basic measurement model, and reasoning to obtain a soil carbon sink preliminary distribution map. According to the method, the weight generation rule is iteratively optimized based on historical data, so that high-precision and self-adaptive measurement and calculation of the soil carbon sink in the urban dynamic environment are realized.

Description

A method and system for measuring urban soil carbon sequestration based on multi-source data Technical Field

[0001] This invention relates to the field of ecological environment monitoring technology, and in particular to a method and system for calculating urban soil carbon sequestration based on multi-source data. Background Technology

[0002] With increasing global attention on climate change and carbon neutrality, accurate measurement of urban soil carbon sinks is becoming increasingly important for assessing regional carbon budgets and formulating climate policies. Under the influence of intensive human activities, urban soil carbon sinks exhibit high spatial heterogeneity and dynamic instability. Traditional methods that rely on limited ground sampling are costly and difficult to capture dynamic changes. Current mainstream technologies integrate multi-source remote sensing data and meteorological data with machine learning models to achieve large-scale soil organic carbon inversion, improving the coverage and efficiency of the measurement.

[0003] Existing technologies face core limitations when dealing with dynamic urban environments: their multi-source data fusion strategies are usually static or globally optimized, and cannot adaptively respond to dynamic changes in data reliability caused by local sudden events. In urban environments, events such as construction projects and extreme weather can instantly change the surface state, causing the information value of different data sources to fluctuate with the type and context of the event. Existing fixed-weight fusion methods cannot accurately assess and dynamically adjust the contribution weights of each data source during the impact of an event, resulting in distorted fusion features. This affects the accuracy and reliability of the model's calculations in the event-affected area, making it difficult for the results to effectively support the precise carbon management needs of highly dynamic cities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for measuring urban soil carbon sink based on multi-source data to solve the problem that existing multi-source data fusion strategies cannot adaptively adjust the weights of each data source according to local sudden events, resulting in insufficient accuracy of carbon sink measurement in dynamic urban environments.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: Firstly, this invention provides a method for measuring urban soil carbon sink based on multi-source data, comprising: collecting and preprocessing multi-source raw data to generate a multi-source feature data cube; analyzing the multi-source feature data cube according to preset rules to identify potential impact events and form an event log list; generating fusion weights for each affected spatial grid cell based on weight generation rules, forming a dynamic fusion weight set; using the dynamic fusion weight set to weightedly fuse features in the multi-source feature data cube to generate a context-enhanced feature vector; inputting the context-enhanced feature vector into a pre-trained basic soil carbon sink measurement model to infer a preliminary soil carbon sink distribution map; acquiring new field measurement data and assimilating and calibrating the preliminary soil carbon sink distribution map using a Kalman filter algorithm to generate an assimilated and calibrated soil carbon sink distribution map; combining the event log list and the dynamic fusion weight set to post-process the assimilated and calibrated soil carbon sink distribution map to generate a final soil carbon sink measurement result map; and iteratively optimizing the weight generation rules based on historical task data.

[0007] As a preferred embodiment of the urban soil carbon sequestration measurement method based on multi-source data described in this invention, the method includes the following steps: acquiring multi-source raw data and preprocessing it to generate a multi-source feature data cube; acquiring optical images, soil temperature and humidity, precipitation data, and green space pedestrian flow records, and aggregating them into an initial aggregation package; performing atmospheric correction on the optical images to generate a first corrected image; performing outlier processing on the soil temperature and humidity to generate standardized sensor data; and spatially resampling and temporally aligning the first corrected image, standardized sensor data, precipitation data, and green space pedestrian flow records to generate vegetation index features, soil moisture features, and environmental meteorological features, and integrating them into a multi-source feature data cube.

[0008] As a preferred embodiment of the urban soil carbon sequestration measurement method based on multi-source data described in this invention, the method involves: analyzing a multi-source feature data cube according to preset rules to identify potential impact events and forming an event log list, including the following steps: extracting daily cumulative precipitation from environmental meteorological features, extracting the normalized vegetation index (NVI) change trend from vegetation index features, and extracting the increase in pedestrian flow per unit time from green space pedestrian records; comparing the daily cumulative precipitation with a first threshold in the preset rules for heavy precipitation events, comparing the NVI change trend with a second threshold in the preset rules for vegetation disturbance events, and comparing the increase in pedestrian flow per unit time with a third threshold in the preset rules for clustered events; identifying spatial units affected by heavy precipitation exceeding the first threshold, spatial units affected by vegetation disturbance exceeding the second threshold, and spatial units affected by clustered events exceeding the third threshold based on the comparison results; generating an event record entry for each identified spatial unit, containing the event type, occurrence timestamp, coordinates of the affected spatial range, and event intensity level; integrating the event record entries of all spatial units and sorting them chronologically to form a structured event log list.

[0009] As a preferred embodiment of the urban soil carbon sink measurement method based on multi-source data described in this invention, the method includes the following steps: Based on the weight generation rules, a fusion weight is generated for each affected spatial grid cell targeting different data source characteristics to form a dynamic fusion weight set. This includes: extracting vegetation index feature values ​​at corresponding locations in the multi-source feature data cube based on spatial grid cells recorded as heavy precipitation events in the event log list; inputting the event type in the event log list, the vegetation index feature values ​​in the multi-source feature data cube, and the combination of the event type (heavy precipitation event) and the vegetation index feature value (high) into the weight generation rules; outputting a fusion weight allocation scheme for vegetation index feature weights, soil moisture feature weights, and environmental meteorological feature weights through the weight generation rules; applying the fusion weight allocation scheme to each spatial grid cell of the heavy precipitation event to generate vegetation index feature fusion weights, soil moisture feature fusion weights, and environmental meteorological feature fusion weights; and collecting the vegetation index feature fusion weights, soil moisture feature fusion weights, and environmental meteorological feature fusion weights of the spatial grid cells to form a dynamic fusion weight set.

[0010] As a preferred embodiment of the urban soil carbon sequestration measurement method based on multi-source data described in this invention, the method involves: weighting and fusing features in a multi-source feature data cube using a dynamic fusion weight set to generate a context-enhanced feature vector, comprising the following steps: extracting vegetation index feature values, soil moisture feature values, and environmental meteorological feature values ​​from the multi-source feature data cube corresponding to spatial grid cells; performing a weighting operation on the vegetation index feature values ​​using vegetation index feature weights to generate weighted vegetation index features; performing a weighting operation on the soil moisture feature values ​​using soil moisture feature weights to generate weighted soil moisture features; performing a weighting operation on the environmental meteorological feature values ​​using environmental meteorological feature weights to generate weighted environmental meteorological features; and combining the weighted vegetation index features, weighted soil moisture features, and weighted environmental meteorological features to generate a context-enhanced feature vector.

[0011] As a preferred embodiment of the urban soil carbon sink measurement method based on multi-source data described in this invention, the following steps are included: inputting context-enhanced feature vectors into a pre-trained basic soil carbon sink measurement model to infer a preliminary distribution map of soil carbon sink; loading a historical multi-source feature data cube and a training dataset of corresponding field-measured soil carbon sink values; training the model on the training dataset using a gradient boosting decision tree algorithm to obtain a pre-trained basic soil carbon sink measurement model; triggering a dynamic inference switch based on the type and intensity of active events in the event log list to selectively switch the pre-trained basic soil carbon sink measurement model between a standard forward propagation path and a feature attention transformation path specifically fine-tuned for different event types; performing differentiated inference on the context-enhanced feature vectors to output an estimated soil organic carbon density value for each spatial grid cell; and interpolating the estimated soil organic carbon density values ​​of the spatial grid cells on the spatial grid to generate a preliminary distribution map of soil carbon sink.

[0012] As a preferred embodiment of the urban soil carbon sink measurement method based on multi-source data described in this invention, the method includes the following steps: acquiring new field measurement data, assimilating and calibrating the preliminary soil carbon sink distribution map based on the Kalman filter algorithm, and generating an assimilated and calibrated soil carbon sink distribution map. These steps include: locating spatial grid cells corresponding to the geographical locations of the new field measurement data in the preliminary soil carbon sink distribution map; comparing the estimated soil organic carbon density located in the preliminary soil carbon sink distribution map with the measured soil organic carbon values ​​from the new field measurement data to obtain residuals; using the residuals to update the state and error covariance of all spatial grid cells in the preliminary soil carbon sink distribution map based on the Kalman filter algorithm; and generating a spatially continuous assimilated and calibrated soil carbon sink distribution map based on the updated state after the Kalman filter algorithm.

[0013] As a preferred embodiment of the urban soil carbon sink measurement method based on multi-source data described in this invention, the method includes the following steps: Post-processing the assimilated and calibrated soil carbon sink distribution map by combining an event log list and a dynamic fusion weight set to generate the final soil carbon sink measurement result map. This includes: identifying the corresponding spatial region in the assimilated and calibrated soil carbon sink distribution map based on the impact range of heavy rainfall events recorded in the event log list; assessing the data reliability level based on the vegetation index characteristic weight change pattern of the dynamic fusion weight set in the corresponding spatial region; superimposing predictive annotations of short-term carbon sink decay on the corresponding spatial region of the assimilated and calibrated soil carbon sink distribution map according to the data reliability level and the characteristics of heavy rainfall events; calculating the confidence interval value of the corresponding spatial region based on the data reliability level to obtain the uncertainty level; and generating the final soil carbon sink measurement result map by superimposing the predictive annotations and the uncertainty level annotations on the assimilated and calibrated soil carbon sink distribution map.

[0014] As a preferred embodiment of the urban soil carbon sequestration measurement method based on multi-source data described in this invention, the method involves iteratively optimizing the weight generation rule based on historical task data, including the following steps: collecting a list of event logs, a dynamic fusion weight set, and corresponding verification error records of the final soil carbon sequestration measurement result map from historical tasks to constitute historical task data; analyzing the historical task data, statistically analyzing the verification errors corresponding to the dynamic fusion weight set used under different combinations of event types and environmental features, identifying event types and environmental feature combinations with verification errors exceeding a threshold, and determining them as low-confidence combinations that need optimization in the weight generation rule; adjusting the mapping logic corresponding to the low-confidence combinations in the weight generation rule, and updating the weight generation rule using the adjusted mapping logic.

[0015] Secondly, this invention provides a system for measuring urban soil carbon sequestration based on multi-source data, comprising: an identification module, which collects and preprocesses multi-source raw data to generate a multi-source feature data cube, analyzes the multi-source feature data cube according to preset rules, identifies potential impact events, and forms an event log list; a weight generation module, which generates fusion weights for each affected spatial grid cell based on weight generation rules, targeting features from different data sources, forming a dynamic fusion weight set; a fusion module, which uses the dynamic fusion weight set to weightedly fuse features in the multi-source feature data cube to generate a context-enhanced feature vector; an inference module, which inputs the context-enhanced feature vector into a pre-trained basic soil carbon sequestration measurement model to infer a preliminary distribution map of soil carbon sequestration; a calibration module, which acquires new field measurement data, performs assimilation calibration on the preliminary distribution map of soil carbon sequestration based on a Kalman filter algorithm, and generates an assimilation calibration soil carbon sequestration distribution map; and an iteration module, which combines the event log list and the dynamic fusion weight set to post-process the assimilation calibration soil carbon sequestration distribution map, generates a final soil carbon sequestration measurement result map, and iteratively optimizes the weight generation rules based on historical task data.

[0016] The beneficial effects of this invention are as follows: It collects and preprocesses optical images, soil temperature and humidity, precipitation, and pedestrian flow data to generate a multi-source feature data cube. Based on preset rules, it identifies human and natural events such as heavy precipitation and vegetation disturbance, forming an event log list. Based on the event log, it dynamically generates fusion weight sets for different data source features for affected grid cells using weight generation rules. These weight sets are then used to weight and fuse features, generating context-enhanced feature vectors, which are input into a pre-trained measurement model. A dynamic inference switch triggered by events switches between standard and dedicated paths, inferring a preliminary carbon sink distribution map. New field measurement data and Kalman filtering are used for assimilation and calibration. The calibrated distribution map is post-processed using the event log and dynamic weight set, with event-related predictive annotations superimposed and uncertainty levels calculated to generate the final result map. The weight generation rules are iteratively optimized based on historical data, thereby achieving high-precision, adaptive measurement of soil carbon sinks in dynamic urban environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0018] Figure 1 is a flowchart of the method for measuring urban soil carbon sink based on multi-source data.

[0019] Figure 2 is a schematic diagram of an urban soil carbon sequestration measurement system based on multi-source data. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0023] Referring to Figures 1 and 2, an embodiment of the present invention is provided, which provides a method for measuring urban soil carbon sink based on multi-source data, including the following steps: S1, collecting multi-source raw data and preprocessing it to generate a multi-source feature data cube.

[0024] S1.1 Acquire optical images, soil temperature and humidity, precipitation data, and green space pedestrian flow records, and aggregate them into an initial aggregation package. Perform atmospheric correction on the optical images to generate the first corrected image, and perform outlier processing on the soil temperature and humidity to generate standardized sensor data.

[0025] Furthermore, optical images are acquired from satellite data sources. These images undergo atmospheric correction to eliminate the effects of atmospheric scattering and absorption, generating a first corrected image. This process improves the accuracy of the surface reflectance of the optical images, laying the foundation for the accurate extraction of subsequent derived features such as vegetation indices. Soil temperature and humidity data are acquired from a sensor network deployed in urban green spaces. The soil temperature and humidity data undergo outlier processing to identify and remove records that deviate from the normal physical range due to momentary sensor failures or transmission interference, generating standardized sensor data. This process ensures the reliability and consistency of ground monitoring data. Precipitation data is acquired from a meteorological data center, and pedestrian flow records in green spaces are obtained from a network platform. The first corrected image, standardized sensor data, precipitation data, and pedestrian flow records in green spaces are then aggregated into an initial aggregation package.

[0026] S1.2. Spatially resample and temporally align the first corrected image, standardized sensor data, precipitation data, and green space pedestrian flow records to generate vegetation index features, soil moisture features, and environmental meteorological features, and integrate them into a multi-source feature data cube.

[0027] Furthermore, the first corrected image, standardized sensor data, precipitation data, and green space pedestrian flow records are spatially resampled and unified to a predefined spatial grid and projection coordinate system to ensure that all data layers are fully aligned in space. At the same time, temporal alignment is performed to unify the timestamps of all data to the same observation reference time. On the unified spatiotemporal reference, vegetation index features are calculated and generated from the first corrected image, soil moisture features are mapped from the standardized sensor data, and environmental meteorological features are calculated and generated by combining precipitation data and pedestrian flow records. The generated vegetation index features, soil moisture features, and environmental meteorological features are integrated into a structured multi-source feature data cube.

[0028] S2. Analyze the multi-source feature data cube according to preset rules, identify potential impact events, and form an event log list.

[0029] S2.1 Extract daily cumulative precipitation from environmental meteorological characteristics, extract the normalized vegetation index change trend from vegetation index characteristics, and extract the increase in pedestrian flow per unit time from green space pedestrian flow records.

[0030] Furthermore, in the environmental and meteorological features of the multi-source feature data cube, the precipitation data is accumulated over time to extract the daily cumulative precipitation from the initial observation time to the current time; in the vegetation index features, the change or rate of change of the normalized vegetation index at the current time compared to the past observation period is obtained to extract the trend of the normalized vegetation index; in the green space pedestrian flow record, the increase in pedestrian flow within a unit time window is calculated to extract the increase in pedestrian flow per unit time.

[0031] S2.2. Compare the daily cumulative precipitation with the first threshold in the preset rules for heavy precipitation events, compare the change trend of the normalized vegetation index with the second threshold in the preset rules for vegetation disturbance events, and compare the increase in the number of people per unit time with the third threshold in the preset rules for gathering events.

[0032] Furthermore, the extracted daily cumulative precipitation is compared with a predefined first threshold used to identify heavy precipitation events to determine whether it has reached or exceeded the threshold; the extracted normalized vegetation index change trend is compared with a predefined second threshold used to identify vegetation disturbance events to determine whether its change magnitude has reached or exceeded the threshold; and the extracted increase in pedestrian traffic per unit time is compared with a predefined third threshold used to identify clustering events to determine whether it has reached or exceeded the threshold, thus correlating continuous observation indicators with discrete event discrimination criteria.

[0033] Specifically, the thresholds in the preset rules are critical values ​​determined based on domain knowledge or historical statistics. For example, the first threshold corresponds to the amount of rainfall that may cause soil erosion or water saturation, the second threshold corresponds to the degree of change in the health status of vegetation, and the third threshold corresponds to the population density that may pose a risk of soil compaction in green spaces. The threshold-based rule discrimination method constructs a mapping bridge from continuous signals to discrete event types, making it possible to automatically and in real time capture various potential disturbances in the urban environment.

[0034] S2.3 Based on the comparison results, identify spatial units affected by heavy precipitation exceeding the first threshold, spatial units affected by vegetation disturbance exceeding the second threshold, and spatial units affected by clustering events exceeding the third threshold.

[0035] Furthermore, based on the comparison results, all spatial units with daily cumulative precipitation exceeding the first threshold are identified in geospatial data. These units are classified as spatial units affected by heavy precipitation. All spatial units with normalized vegetation index (NVI) changes exceeding the second threshold are identified. These units are classified as spatial units affected by vegetation disturbance. All spatial units with per-unit-time increase in pedestrian traffic exceeding the third threshold are identified. These units are classified as spatial units affected by clustered events. This achieves aggregation from points or pixels that meet the conditions to affected areas with clear geographical scope, enabling precise location of the spatial range of events, rather than just giving a judgment on whether an overall event has occurred. For example, a heavy precipitation event may only affect a local area of ​​a city, accurately delineating the set of affected grid units.

[0036] S2.4 Generate an event record entry for each identified spatial unit, containing the event type, occurrence timestamp, affected spatial range coordinates, and event intensity level. Integrate the event record entries of all spatial units and sort them in chronological order to form a structured event log list.

[0037] Furthermore, for each identified spatial unit affected by heavy precipitation, vegetation disturbance, or clustering events, a structured event record entry is created. Each event record entry includes the following fields: an event type field indicating whether it is a heavy precipitation event, a vegetation disturbance event, or a clustering event; an occurrence timestamp field recording the specific time when the unit was identified as meeting the threshold conditions; an impact spatial range coordinate field recording the unique coordinate identifier of the unit in the preset spatial grid; and an event intensity level field that can be graded based on the magnitude of exceeding the threshold, for example, according to the percentage of daily cumulative precipitation exceeding the first threshold. Finally, all generated event record entries from different spatial units are sorted and merged according to the order of their occurrence timestamps to form a structured, time-ordered event log list.

[0038] S3. Based on the weight generation rules, generate fusion weights for each affected spatial grid cell that are tailored to different data source characteristics, forming a dynamic fusion weight set.

[0039] S3.1. Based on the spatial grid cells recorded as heavy precipitation events in the event log list, extract the vegetation index feature values ​​of the corresponding locations in the multi-source feature data cube.

[0040] Furthermore, the event log list is read, and all entries with the event type field recorded as heavy precipitation events are filtered out. From the entries, each specific spatial grid cell identified by the coordinate field of the affected spatial range is obtained. Based on the coordinates of these spatial grid cells, the vegetation index feature values ​​of the corresponding spatial locations are queried and extracted from the multi-source feature data cube.

[0041] Specifically, by establishing a direct link between events and the specific environmental conditions of the affected areas, and by precisely navigating the event log list, the geographical locations affected by specific events can be quickly located, and key environmental characterization data of these locations at the time of the event can be obtained from the data cube. For example, the vegetation index characteristic value reflects the density or health status of the vegetation at that location, which is an important basis for judging the impact of the event.

[0042] S3.2 Input the event type from the event log list, the vegetation index feature value in the multi-source feature data cube, and the combination of the event type (heavy precipitation event) and the vegetation index feature value (high) into the weight generation rule.

[0043] Furthermore, the event type field corresponding to the heavy precipitation event entries selected from the event log list is combined with the vegetation index feature value of the corresponding spatial grid unit to form an input combination. At the same time, the weight generation rule predefines several standard scenario combinations, such as the combination of event type heavy precipitation event and high vegetation index feature value, and matches or maps the current actual input combination with the standard scenario combination in the weight generation rule.

[0044] Specifically, the core logic entry point for dynamic weight generation couples discrete event types with continuous environmental characteristic values ​​(such as high, medium, and low vegetation indices) to form a composite context description. This composite context more accurately portrays the changing background of data reliability than a single event type. For example, the impact of heavy precipitation events on the reliability of optical remote sensing data differs drastically depending on whether they occur in densely vegetated areas or bare soil areas; this cannot be distinguished by event type alone. By inputting composite conditions such as heavy precipitation events plus high vegetation indices, the decision-making of the weight generation rules can be refined to a more precise context level, thereby achieving a more refined assessment of data reliability.

[0045] S3.3 Output the fusion weight allocation scheme of vegetation index feature weight, soil moisture feature weight and environmental meteorological feature weight through weight generation rules.

[0046] Furthermore, upon receiving an input combination consisting of event type and vegetation index feature values, the weight generation rule outputs a corresponding fusion weight allocation scheme based on predefined mapping logic. This scheme explicitly specifies the weight values ​​or weight adjustment coefficients that should be assigned to the three types of data source features—vegetation index feature weights, soil moisture feature weights, and environmental meteorological feature weights—in the specific context of the current identification. The weight generation rule is essentially a lookup table or a set of decision logics, and its mapping relationship is pre-defined based on domain knowledge or historical data analysis.

[0047] Specifically, for example, in the case of a combination of heavy rainfall events and high vegetation index feature values, the rule may output a scheme to reduce the weight of vegetation index features and increase the weight of soil moisture features. This solidifies the complex data reliability judgment logic into an executable set of rules, making the dynamic weight allocation process automated, interpretable, and consistent. The rule outputs not a single weight, but a complete weight allocation scheme for different features, which reflects the differentiated assessment of the relative importance of different data sources in specific contexts.

[0048] S3.4 Apply the fusion weight allocation scheme to generate vegetation index feature fusion weight, soil moisture feature fusion weight, and environmental meteorological feature fusion weight for each spatial grid cell of the heavy precipitation event.

[0049] Furthermore, for each spatial grid cell identified in the event log list as being affected by a heavy precipitation event, a fusion weight allocation scheme is applied to assign each cell a vegetation index feature fusion weight, a soil moisture feature fusion weight, and an environmental meteorological feature fusion weight. This implements the global weight allocation scheme to each affected spatial location, achieving personalized weight configuration at the spatial granularity.

[0050] Specifically, the weight set obtained for each spatial unit is derived based on its own event attributes (heavy precipitation) and local environmental conditions (vegetation index value), which ensures the relevance and rationality of the weight allocation. For example, within the area affected by a heavy precipitation event, units with high vegetation cover and units with low vegetation cover may be assigned different values ​​for their vegetation index feature weights, achieving fine-tuning based on unit-level context.

[0051] S3.5. Aggregate the vegetation index feature fusion weights, soil moisture feature fusion weights, and environmental meteorological feature fusion weights of spatial grid units to form a dynamic fusion weight set.

[0052] Furthermore, after assigning weights to each spatial grid cell affected by the heavy precipitation event, the fusion weights of vegetation index features, soil moisture features, and environmental meteorological features generated by all these spatial grid cells are collected and organized into a structured dataset corresponding to the spatial grid index, namely, a dynamic fusion weight set. This dynamic fusion weight set not only contains the weight values ​​themselves but is also associated with each specific grid cell through the spatial index.

[0053] Specifically, a global, spatially queryable weight lookup table is generated. The dynamic fusion weight set is the final output, decoupling the event-driven, context-aware dynamic weighting process from the subsequent static data fusion process. This allows the fusion operation to be executed efficiently and in batches, while ensuring that the data fusion at each spatial location follows a weighting strategy tailored to that location.

[0054] S4. Use the dynamic fusion weight set to perform weighted fusion of features in the multi-source feature data cube to generate context-enhanced feature vectors.

[0055] S4.1 Extract vegetation index feature values, soil moisture feature values, and environmental meteorological feature values ​​from the multi-source feature data cube.

[0056] Furthermore, for each spatial grid cell that needs to be processed, based on the coordinate index of the cell, vegetation index feature values, soil moisture feature values, and environmental meteorological feature values ​​are read and extracted from the corresponding spatial location of the multi-source feature data cube. From the dynamic fusion weight set, based on the coordinate index of the same spatial grid cell, the vegetation index feature fusion weight, soil moisture feature fusion weight, and environmental meteorological feature fusion weight generated in advance for that cell are read and extracted.

[0057] Specifically, the original feature data and context-aware weight data are precisely aligned and paired in space to prepare the corresponding operands and coefficients for subsequent weighting operations. This relies on the consistency of the spatial grid definition between the multi-source feature data cube and the dynamically fused weight set, ensuring that the feature value at each location corresponds to the weight customized for that location. This is the foundation for realizing spatially heterogeneous weighted fusion.

[0058] S4.2. Use vegetation index feature weights to perform a weighted operation on the vegetation index feature values ​​to generate a weighted vegetation index feature. Use soil moisture feature weights to perform a weighted operation on the soil moisture feature values ​​to generate a weighted soil moisture feature. Use environmental meteorological feature weights to perform a weighted operation on the environmental meteorological feature values ​​to generate a weighted environmental meteorological feature.

[0059] Furthermore, the vegetation index feature fusion weights extracted from the dynamic fusion weight set are used as multipliers and multiplied with the vegetation index feature values ​​extracted from the multi-source feature data cube. The result is the weighted vegetation index feature of the unit. Similarly, the soil moisture feature fusion weights extracted from the dynamic fusion weight set are used as multipliers and multiplied with the soil moisture feature values ​​extracted from the multi-source feature data cube. The result is the weighted soil moisture feature of the unit. Finally, the environmental meteorological feature fusion weights extracted from the dynamic fusion weight set are used as multipliers and multiplied with the environmental meteorological feature values ​​extracted from the multi-source feature data cube. The result is the weighted environmental meteorological feature of the unit.

[0060] Specifically, dynamically generated fusion weights reflecting event and contextual characteristics are applied to the original feature values ​​to perform a context-aware feature transformation, assigning different importance coefficients to the features of each spatial unit. For example, in areas affected by heavy rainfall and with dense vegetation, the fusion weight of vegetation index features may be lower, while the fusion weight of soil moisture features may be higher. This results in soil moisture information being enhanced and vegetation index information being relatively suppressed in the weighted feature vector, thus more accurately reflecting the fact that soil moisture is a more reliable indicator of carbon sinks in this context. The rule-based, interpretable weighting operation essentially encodes domain knowledge (the reliability differences of different data sources in a specific context) into the feature representation.

[0061] S4.3 Combine weighted vegetation index features, weighted soil moisture features, and weighted environmental meteorological features to generate a context-enhanced feature vector.

[0062] Furthermore, for each spatial grid cell, the three scalar values ​​obtained after weighting operations—weighted vegetation index, weighted soil moisture, and weighted environmental meteorological features—are connected or combined in a predetermined order to form a multidimensional feature vector. This vector is defined as the context-enhanced feature vector of the spatial grid cell. For example, the three weighted feature values ​​are arranged sequentially in the order of vegetation index, soil moisture, and environmental meteorological features to form a three-dimensional vector.

[0063] Specifically, a new feature representation that integrates multi-source information and is context-aware weighted is constructed to replace the original multi-source feature data. The generated context-enhanced feature vector not only contains information from the original data, but also incorporates judgments on the current local events and environmental state through dynamic weights. The feature representation serves as the input to the subsequent soil carbon sequestration basic measurement model, enabling the model to directly infer based on the information that has been optimized for reliability and obtain prediction results.

[0064] S5. Input the context-enhanced feature vector into the pre-trained soil carbon sink basic measurement model, and infer the preliminary distribution map of soil carbon sink.

[0065] S5.1 Load the historical multi-source feature data cube and the corresponding field measurement soil carbon sink value training dataset, and use the gradient boosting decision tree algorithm to train on the training dataset to obtain the pre-trained basic measurement model of soil carbon sink.

[0066] Furthermore, a training dataset is constructed using a stored historical multi-source feature data cube and field-measured soil carbon sink data that are strictly paired with it in time and space. Each sample in this training dataset consists of a feature vector composed of vegetation index features, soil moisture features, and environmental meteorological features of the historical multi-source feature data cube at a certain historical moment and spatial location, and the measured soil organic carbon density value obtained at the same location at the same moment through field sampling and laboratory analysis as a label. The gradient boosting decision tree algorithm is used, with the feature vector as input and the measured soil organic carbon density value as the learning target, to iteratively train the algorithm until the model prediction error converges to an acceptable level, thereby obtaining a pre-trained basic measurement model of soil carbon sink that can map from the feature vector to the soil organic carbon density estimate.

[0067] Specifically, a benchmark prediction model with strong nonlinear fitting capabilities is constructed. This model learns the complex relationship between features and carbon sinks from historical data, providing a foundation for subsequent inference. The gradient boosting decision tree algorithm can effectively handle the interactions between features and resist overfitting, making it suitable as a basic measurement model.

[0068] S5.2. Based on the type and intensity of active events in the event log list, a dynamic inference switch is triggered to control the pre-trained soil carbon sink basic measurement model to selectively switch between the standard forward propagation path and the feature attention transformation path that is specially fine-tuned for each type of event. Differential inference is performed on the context-enhanced feature vector, and the estimated value of soil organic carbon density for each spatial grid cell is output.

[0069] Furthermore, the event log list is queried to check if there are any active event records in the current area to be measured. If so, a dynamic inference switch is triggered based on the event type and intensity level recorded in the event log list. The dynamic inference switch controls the computation flow inside the pre-trained soil carbon sink basic measurement model according to the event type, selectively switching it from the standard forward propagation path to a backup inference path that has been fine-tuned for the parameters of this type of event. The backup path is usually implemented by adjusting the feature attention mechanism or some network layer parameters on the basic model to better capture the special patterns of the impact of this type of event on carbon sink. The selected path is used to perform forward propagation on the input context-enhanced feature vector, and the soil organic carbon density estimate of the corresponding spatial grid cell is output.

[0070] Specifically, it breaks away from the static, single-path inference of traditional models and introduces an event-conditional dynamic inference mechanism. Instead of treating all inputs the same, it adaptively selects or adjusts its internal attention or processing preferences based on the presence and type of a specific event. For example, when the event log indicates a heavy rainfall event, the dynamic inference switch guides the model to activate a fine-tuning path that focuses more on extracting information from weighted soil moisture features, because soil moisture is a more reliable indicator of short-term carbon sink changes in this context. For clustered events, it may activate a path that focuses more on human activity features, making the same base model resilient to different disturbance scenarios. This improves the prediction accuracy and adaptability in non-stationary, event-disturbed urban environments without the need to train completely independent models for each scenario.

[0071] S5.3. Based on the estimated soil organic carbon density of the spatial grid cells, interpolation is performed on the spatial grid to generate a preliminary distribution map of soil carbon sink.

[0072] Furthermore, after obtaining the soil organic carbon density estimates for all spatial grid cells, the center point coordinates of each spatial grid cell are used as location points, and the estimated soil organic carbon density of that cell is used as the attribute value of that point. On the spatial grid framework of the entire study area, spatial interpolation methods, such as Kriging interpolation or inverse distance weighted interpolation, are used to spatially estimate the attribute values ​​of grid locations that are not directly calculated or have no estimated values ​​due to missing data, thereby generating a preliminary spatially continuous distribution map of soil carbon sinks covering the entire study area.

[0073] Specifically, discrete point-like or grid-like estimation results are converted into continuous spatial distribution surfaces, facilitating visualization analysis and subsequent spatial processing. Spatial interpolation methods utilize the first law of geography, which states that spatially proximate objects are more similar than those that are far apart. This allows for the reasonable inference of values ​​for unknown areas based on the estimated values ​​of known points, forming a complete spatial distribution pattern map.

[0074] S6. Obtain new field measurement data, and perform assimilation calibration on the preliminary distribution map of soil carbon sink based on the Kalman filter algorithm to generate an assimilation calibration soil carbon sink distribution map.

[0075] S6.1. In the preliminary distribution map of soil carbon sink, locate the spatial grid cells corresponding to the geographical locations of the new field measurement data.

[0076] Furthermore, the new field measurement data includes a set of sampling point information with precise geographic coordinates and corresponding measured values ​​of soil organic carbon. Based on the geographic coordinates of these sampling points, they are matched to the same spatial reference system and grid division system on which the preliminary distribution map of soil carbon sink is based. The spatial grid cell where each sampling point is located is determined, and the unique identifier or index of these spatial grid cells is recorded. A spatial correspondence between high-precision point-based measured data and continuous areal estimation results is established. Through precise geographic registration, it is ensured that subsequent comparisons and calibrations are performed in the same spatial location. This is a prerequisite for the effective implementation of data assimilation.

[0077] S6.2. Compare the estimated soil organic carbon density located in the preliminary distribution map of soil carbon sink with the measured soil organic carbon value from the new field measurement data to obtain the residual.

[0078] Furthermore, for each successfully located sampling point, the estimated soil organic carbon density of that cell is obtained from the preliminary distribution map of soil carbon sink within its spatial grid cell. Simultaneously, the measured soil organic carbon value for the sampling point is obtained from new field measurement data. The difference between the measured soil organic carbon value and the estimated soil organic carbon density is calculated; this difference is the residual for that point. The process of obtaining the residual essentially involves comparing the model's predictions with the true ground observations, quantifying the model's uncertainty or bias at a specific location. For example, the model might overestimate or underestimate carbon storage at a certain location due to localized soil types or unmonitored human activities; the residual represents this bias.

[0079] S6.3. Based on the Kalman filter algorithm, the residual is used to update the state and error covariance of all spatial grid cells of the preliminary distribution map of soil carbon sink. According to the updated state of the Kalman filter algorithm, a spatially continuous assimilation calibration soil carbon sink distribution map is generated.

[0080] Furthermore, the preliminary distribution map of soil carbon sink is considered as a dynamic system state containing the estimated soil organic carbon density of all spatial grid cells and their estimation error covariance matrix. The residuals of each sampling point are used as observation information. The update equation of the Kalman filter algorithm is applied, and these residual observation information are used to globally update the state vector (i.e., the estimated soil organic carbon density) and its error covariance matrix of all spatial grid cells. The Kalman filter update process not only directly corrects the estimated value of the cell where the sampling point is located, but also propagates the correction information to neighboring and even more distant spatial grid cells through the spatial correlation of the error covariance matrix. After the Kalman filter update is completed, the original soil organic carbon density estimate is replaced with the updated state vector (i.e., the corrected soil organic carbon density estimate) to form a new, spatially continuous carbon density distribution, which is the assimilation and calibration soil carbon sink distribution map.

[0081] Specifically, the classic Kalman filter algorithm, which assimilates time-series data, is applied to the optimization of spatial data, achieving point-to-area calibration. Utilizing the mathematical framework of state estimation and error covariance update in Kalman filtering, the uncertainty information and corrections from a few high-precision point observations are optimally propagated and integrated into the continuous areal estimation of the entire region through a spatial correlation structure. For example, if the model overestimates a certain sampling point, Kalman filtering will not only lower the estimate at that point but also appropriately lower the estimates of the surrounding area according to the preset spatial covariance function and the distance decay law. This makes the calibration results spatially smooth and reasonable, fully utilizing the spatial autocorrelation of soil properties. It efficiently improves the global accuracy of the entire distribution map with limited field measurement data, resolving the contradiction between the sparsity of field sampling and the need for comprehensive regional monitoring.

[0082] S7. Combining the event log list and the dynamic fusion weight set, post-process the assimilation calibration soil carbon sink distribution map to generate the final soil carbon sink measurement result map.

[0083] S7.1 Based on the impact range of heavy precipitation events recorded in the event log list, identify the corresponding spatial regions in the assimilated and calibrated soil carbon sink distribution map.

[0084] Furthermore, the event log list is read, and records with the event type field being "heavy precipitation event" are retrieved. From these records, the coordinate field of the affected spatial range is extracted. These coordinates define the set of all spatial grid cells affected by the heavy precipitation event. Based on the extracted coordinate set, these corresponding spatial grid cells are located and identified in the assimilation and calibration soil carbon sink distribution map. The set of these identified spatial grid cells constitutes the corresponding spatial region affected by the heavy precipitation event.

[0085] S7.2. Based on the dynamic fusion weight set’s vegetation index characteristic weight change pattern in the corresponding spatial region, assess the data reliability level. Based on the data reliability level and the characteristics of heavy precipitation events, overlay the predictive label of short-term carbon sink decay in the corresponding spatial region of the assimilated and calibrated soil carbon sink distribution map.

[0086] Furthermore, for each spatial grid cell within the corresponding spatial region, the dynamic fusion weight set is queried to obtain the vegetation index feature fusion weight value assigned to that cell in the most recent weight generation process; the distribution or change pattern of vegetation index feature fusion weights of all cells in the region is analyzed, for example, assessing the degree or range of the general decrease in weight values; based on the decrease pattern of vegetation index feature fusion weights, the relative reliability level of optical remote sensing data in the region during the measurement process is assessed, because a decrease in weights usually means that the data source is judged to be less reliable in the current context; combined with the ecological process characteristic that heavy precipitation events can lead to short-term soil hypoxia, which may inhibit microbial activity and slow down carbon decomposition, a predictive label in the form of text or legend, such as short-term carbon sink decay, is overlaid on the corresponding spatial region affected by heavy precipitation events in the assimilation and calibration soil carbon sink distribution map to indicate to users that the carbon sink function of the region may be negatively affected by the event in the short term.

[0087] Specifically, the data reliability assessment information contained in the dynamic fusion weights is combined with knowledge of the ecological effects of specific events to enhance the interpretability of the measurement results. It not only outputs a numerical result but also qualitative or semi-quantitative predictions about how the result might be affected by recent events. For example, if dynamic weights show that the weights of optical data in a certain area have been significantly reduced, suggesting that cloud cover or canopy water accumulation has severely impacted the observation quality, then overlaying short-term carbon sink attenuation labels in this case is both an inference based on ecological knowledge and a supplementary explanation for how data uncertainty might cause the measured values ​​to deviate from the true trend, thus enhancing the decision support value of the results.

[0088] S7.3 Calculate the confidence interval value of the corresponding spatial region based on the data reliability level to obtain the uncertainty level.

[0089] Furthermore, for the corresponding spatial region, the uncertainty level is calculated. The calculation of the uncertainty level is based on an expression. equal Multiply Multiply by the reciprocal negative The power of, plus Multiply by one and subtract R. Where, It is the data convergence, which is obtained by calculating the variance or dispersion of each feature weight in the dynamic fusion weight set within the region. The larger the variance, the greater the disagreement in trust judgment of different data sources during fusion, the lower the data convergence, and the greater the contribution of uncertainty. Event distance represents the time interval between the current moment and the end time of the most recent heavy rainfall event affecting the region recorded in the event log list. The greater the contribution of uncertainty, the more its impact decays exponentially. This refers to the result validation rate. By querying historical task data, it statistically analyzes the consistency ratio between past assimilation calibration results and subsequent more accurate validation data under similar heavy precipitation events and similar data convergence scenarios. The lower the ratio, the worse the historical performance and the higher the current uncertainty. It is the data convergence adjustment coefficient. This is the result verification rate adjustment coefficient, used to balance the contribution weights of both factors to the total uncertainty. It is calculated using this expression. The value represents the level of uncertainty for that spatial region.

[0090] Specifically, a multi-factor-driven, quantifiable uncertainty assessment model was constructed. Instead of employing simple subjective assignment or a single indicator, it decomposes uncertainty into three sources with clear physical or procedural significance: the internal consistency of the data fusion process, the timeliness of external interference events, and so on. and experiential performance in similar historical contexts. For example, even if the weight adjustment is large during data fusion, leading to... The value is low, but if the event occurred a long time ago, it could lead to... The value is large, and the verification rate for similar cases in history is high. High, ultimate uncertainty It might not be high, or conversely, if the event has just occurred... Even if the data is consistent, the value is small. A higher value may also increase uncertainty. The assessment method, which comprehensively considers real-time data status, event dynamics, and historical experience, is more comprehensive in reflecting the true confidence level of carbon sink measurement in the dynamic urban environment than assessments based solely on model prediction variance or sampling density. This makes the output uncertainty level information more meaningful.

[0091] The expression for the level of uncertainty is:

[0092] in, The level of uncertainty is... For data convergence, For event distance, For the result validation rate, This is the data convergence adjustment coefficient. This is the result verification rate adjustment coefficient.

[0093] S7.4. The assimilation calibration soil carbon sink distribution map with superimposed predictive annotations and annotation uncertainty levels is used to generate the final soil carbon sink measurement result map.

[0094] Furthermore, based on the assimilation-calibrated soil carbon sink distribution map that has already been overlaid with predictive annotations for short-term carbon sink attenuation, uncertainty level values ​​or level labels are further overlaid on the corresponding spatial areas. This comprehensive map, which simultaneously contains spatial distribution information of carbon sink, predictive annotation information affected by events, and uncertainty level information for each region, is generated as the final soil carbon sink measurement result map. It integrates the core measurement results, process interpretation information, and quality assessment information into one comprehensive result. The soil carbon sink measurement result map not only tells users how much carbon sink there is, but also indicates which areas need to pay attention to the potential impact of events and how confident they are in the results. This achieves an improvement from providing a single data product to providing comprehensive decision support information.

[0095] S8. Iteratively optimize the weight generation rules based on historical task data.

[0096] S8.1 Collect the event log list, dynamic fusion weight set and corresponding final soil carbon sequestration measurement result map verification error record from the historical tasks to form historical task data.

[0097] Furthermore, structured data generated from each complete measurement task in the past are collected and archived, including event log lists corresponding to each task, recording all event types, spatial ranges, and temporal information identified during task execution; dynamic fusion weight sets generated for the affected areas in each task, recording the fusion weights assigned to the features of each data source under different scenarios; and the final soil carbon sequestration measurement result map generated in each task. Verification error records generated by subsequent verification of this result map through more intensive or more precise independent field measurements are also obtained. These event log lists, dynamic fusion weight sets, and verification error records, which have time sequence and task identification, are aligned and correlated to form a historical task data set for rule optimization. A historical knowledge base with a complete scenario-decision-result chain is established. The event log lists describe the external scenario at the time, the dynamic fusion weight sets reflect the data fusion decisions made under that scenario, and the verification errors quantify the accuracy of the final results under that decision.

[0098] S8.2 Analyze historical task data, statistically analyze the verification error corresponding to the dynamic fusion weight set used under different event types and environmental feature combinations, identify event types and environmental feature combinations with verification errors higher than the threshold, and determine them as low-confidence combinations that need to be optimized in the weight generation rules.

[0099] Furthermore, the historical task data is analyzed. For each historical task data point, firstly, based on its event log list and the application records of the dynamic fusion weight set, the specific event types and environmental feature combinations actually applied in the task, as well as the corresponding dynamic fusion weights, are extracted. Then, the verification error records of the final soil carbon sequestration measurement results map associated with the task are queried to obtain the error index for the whole or a specific region. The historical data is grouped according to the event type and environmental feature combination, and the dynamic fusion weight modes used under each combination are statistically analyzed. The average verification error or other error aggregation indexes under these weight modes are compared with a pre-set performance threshold. Event types and environmental feature combinations with average verification errors that are consistently or significantly higher than the threshold are identified as low-reliability combinations where the current mapping logic in the weight generation rule may be inaccurate or unsuitable and needs to be optimized.

[0100] Specifically, the performance of the weight generation rule was post-evaluated using historical feedback data, establishing an empirically based rule effectiveness evaluation closed loop. For example, the analysis may have found that under the combination of heavy precipitation events and high vegetation index characteristic values, the strategy of significantly reducing optical weights was used many times in historical tasks, but the final verification error was generally high. This suggests that the decision to significantly reduce optical weights in the current rule may be too aggressive or needs to be adjusted, thus marking it as a low-confidence combination.

[0101] S8.3 Adjust the mapping logic corresponding to low-confidence combinations in the weight generation rule, and update the weight generation rule using the adjusted mapping logic.

[0102] Furthermore, for each low-confidence combination, the mapping logic corresponding to that combination in the weight generation rule is adjusted. The adjustment can be based on in-depth analysis of historical task data under that combination. For example, analyzing the verification error distribution of historical tasks using different weight modes under that combination, and finding a weight mode with relatively low error as an adjustment reference; or, combining domain knowledge to modify the original mapping logic. The specific content of the adjustment can be modifying the specific values ​​or proportional relationships of the output vegetation index feature weight, soil moisture feature weight, and environmental meteorological feature weight under that combination. After completing the adjustment of the mapping logic for all low-confidence combinations, the adjusted new mapping logic is used to cover or update the original weight generation rule, forming an optimized new version of the weight generation rule.

[0103] Specifically, it completes an iterative optimization loop from problem identification to problem solving. The weight generation rule is not a fixed formula, but a living rule set that can be improved by learning from historical experience. The optimization process is goal-oriented, directly aiming to reduce the measurement error in similar situations in the future. For example, for the combination of heavy precipitation events and low confidence with high vegetation index feature values, the optimization may adjust the rule from significantly reducing the optical weight to moderately reducing the optical weight, while slightly increasing the radar weight. This new rule will be used in future measurement tasks. The iterative optimization mechanism enables the adaptability and robustness of the entire measurement method to be continuously enhanced with the accumulation of experience. Especially when dealing with the endless new situations and events in the urban environment, it can gradually correct the shortcomings of the initial rule and achieve continuous performance improvement.

[0104] This embodiment also provides a system for measuring urban soil carbon sink based on multi-source data, including: an identification module, which collects and preprocesses multi-source raw data to generate a multi-source feature data cube, analyzes the multi-source feature data cube according to preset rules, identifies potential impact events, and forms an event log list; a weight generation module, which generates fusion weights for each affected spatial grid cell based on weight generation rules, forming a dynamic fusion weight set; a fusion module, which uses the dynamic fusion weight set to weight and fuse the features in the multi-source feature data cube to generate a context-enhanced feature vector; an inference module, which inputs the context-enhanced feature vector into a pre-trained basic soil carbon sink measurement model to infer a preliminary distribution map of soil carbon sink; a calibration module, which acquires new field measurement data, performs assimilation calibration on the preliminary distribution map of soil carbon sink based on the Kalman filter algorithm, and generates an assimilation-calibrated soil carbon sink distribution map; and an iteration module, which combines the event log list and the dynamic fusion weight set to perform post-processing on the assimilation-calibrated soil carbon sink distribution map to generate the final distribution map. The soil carbon sink measurement results map is generated by iteratively optimizing the weight generation rules based on historical task data. Optical imagery, soil temperature and humidity, precipitation, and pedestrian flow data are collected and preprocessed to generate a multi-source feature data cube. Based on preset rules, human and natural events such as heavy precipitation and vegetation disturbance are identified, forming an event log list. Based on the event log, a fusion weight set targeting different data source features is dynamically generated for the affected grid cells using the weight generation rules. This weight set is then used to weight and fuse features, generating a context-enhanced feature vector, which is then input into a pre-trained measurement model. A dynamic inference switch triggered by events switches between standard and dedicated paths to infer a preliminary carbon sink distribution map. New field measurement data and Kalman filtering are used for assimilation and calibration. The calibrated distribution map is then post-processed using the event log and dynamic weight set, overlaying event-related predictive annotations and calculating the uncertainty level to generate the final result map. The weight generation rules are iteratively optimized based on historical data, thus achieving high-precision and adaptive measurement of soil carbon sinks in a dynamic urban environment.

[0105] This embodiment also provides a computer device applicable to the urban soil carbon sequestration calculation method based on multi-source data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the urban soil carbon sequestration calculation method based on multi-source data as proposed in the above embodiment.

[0106] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0107] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for calculating urban soil carbon sequestration based on multi-source data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0108] In summary, this invention generates a multi-source feature data cube by collecting and preprocessing optical images, soil temperature and humidity, precipitation, and pedestrian flow data. Based on preset rules, it identifies human and natural events such as heavy precipitation and vegetation disturbance, forming an event log list. Using this event log, it dynamically generates fusion weight sets for affected grid cells based on different data source features using weight generation rules. These weight sets are then used to weight and fuse features, generating context-enhanced feature vectors, which are input into a pre-trained measurement model. A dynamic inference switch triggered by events switches between standard and dedicated paths to infer a preliminary carbon sink distribution map. New field measurement data and Kalman filtering are used for assimilation and calibration. The calibrated distribution map is then post-processed using the event log and dynamic weight set, overlaying event-related predictive annotations and calculating uncertainty levels to generate the final result map. Finally, the weight generation rules are iteratively optimized based on historical data, thereby achieving high-precision, adaptive measurement of soil carbon sinks in dynamic urban environments.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for measuring urban soil carbon sequestration based on multi-source data, characterized in that: This includes: collecting and preprocessing raw data from multiple sources to generate a multi-source feature data cube; analyzing the multi-source feature data cube according to preset rules to identify potential impact events and form an event log list; generating fusion weights for each affected spatial grid cell based on weight generation rules, forming a dynamic fusion weight set; and using the dynamic fusion weight set to weight and fuse the features in the multi-source feature data cube to generate a context-enhanced feature vector. The context-enhanced feature vector is input into a pre-trained soil carbon sink basic measurement model to infer a preliminary soil carbon sink distribution map. New field measurement data is acquired, and the preliminary soil carbon sink distribution map is assimilated and calibrated based on the Kalman filter algorithm to generate an assimilated and calibrated soil carbon sink distribution map. The assimilated and calibrated soil carbon sink distribution map is post-processed by combining an event log list and a dynamically fused weight set to generate a final soil carbon sink measurement result map. The weight generation rule is iteratively optimized based on historical task data.

2. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 1, characterized in that: The process of collecting and preprocessing multi-source raw data to generate a multi-source feature data cube includes the following steps: acquiring optical images, soil temperature and humidity, precipitation data, and green space pedestrian records, and aggregating them into an initial aggregation package; performing atmospheric correction on the optical images to generate a first corrected image; and performing outlier processing on the soil temperature and humidity to generate standardized sensor data; and then spatially resampling and temporally aligning the first corrected image, standardized sensor data, precipitation data, and green space pedestrian records to generate vegetation index features, soil moisture features, and environmental meteorological features, which are then integrated into a multi-source feature data cube.

3. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 2, characterized in that: The multi-source feature data cube is analyzed according to preset rules to identify potential impact events and form an event log list. This includes the following steps: extracting daily cumulative precipitation from environmental meteorological features, extracting the normalized vegetation index (NVI) trend from vegetation index features, and extracting the increase in pedestrian flow per unit time from green space pedestrian records; comparing the daily cumulative precipitation with the first threshold in the preset rules for heavy precipitation events, comparing the NVI trend with the second threshold in the preset rules for vegetation disturbance events, and comparing the increase in pedestrian flow per unit time with the third threshold in the preset rules for clustered events; identifying spatial units affected by heavy precipitation exceeding the first threshold, spatial units affected by vegetation disturbance exceeding the second threshold, and spatial units affected by clustered events exceeding the third threshold based on the comparison results; generating event record entries for each identified spatial unit, including event type, occurrence timestamp, coordinates of the affected spatial range, and event intensity level; integrating the event record entries of all spatial units and sorting them chronologically to form a structured event log list.

4. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 3, characterized in that: Based on the weight generation rules, a fusion weight is generated for each affected spatial grid cell, targeting different data source characteristics, to form a dynamic fusion weight set. This includes the following steps: Extracting vegetation index feature values ​​from the corresponding locations in the multi-source feature data cube for spatial grid cells recorded as heavy precipitation events in the event log list; inputting the event type from the event log list, the vegetation index feature values ​​from the multi-source feature data cube, and the combination of the event type (heavy precipitation event) and the highest vegetation index feature value into the weight generation rules; outputting a fusion weight allocation scheme for vegetation index feature weights, soil moisture feature weights, and environmental meteorological feature weights through the weight generation rules; applying the fusion weight allocation scheme to generate vegetation index feature fusion weights, soil moisture feature fusion weights, and environmental meteorological feature fusion weights for each spatial grid cell of the heavy precipitation event; and aggregating the vegetation index feature fusion weights, soil moisture feature fusion weights, and environmental meteorological feature fusion weights of the spatial grid cells to form a dynamic fusion weight set.

5. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 4, characterized in that: The method utilizes a dynamic fusion weight set to weight and fuse features in a multi-source feature data cube to generate a context-enhanced feature vector. This involves the following steps: extracting vegetation index feature values, soil moisture feature values, and environmental meteorological feature values ​​from the corresponding spatial grid cells of the multi-source feature data cube; performing a weighted operation on the vegetation index feature values ​​using vegetation index feature weights to generate weighted vegetation index features; performing a weighted operation on the soil moisture feature values ​​using soil moisture feature weights to generate weighted soil moisture features; and performing a weighted operation on the environmental meteorological feature values ​​using environmental meteorological feature weights to generate weighted environmental meteorological features; and combining the weighted vegetation index features, weighted soil moisture features, and weighted environmental meteorological features to generate a context-enhanced feature vector.

6. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 5, characterized in that: The process involves inputting context-enhanced feature vectors into a pre-trained soil carbon sink baseline measurement model to infer a preliminary distribution map of soil carbon sinks. This includes the following steps: loading a historical multi-source feature data cube and a training dataset containing corresponding field-measured soil carbon sink values; training the model on the training dataset using a gradient boosting decision tree algorithm to obtain the pre-trained soil carbon sink baseline measurement model; triggering a dynamic inference switch based on the type and intensity of active events in the event log list to selectively switch the pre-trained model between a standard forward propagation path and a feature attention transformation path specifically fine-tuned for different event types; performing differentiated inference on the context-enhanced feature vectors; and outputting an estimated soil organic carbon density value for each spatial grid cell. Finally, interpolating the estimated soil organic carbon density values ​​on the spatial grid cell to generate a preliminary distribution map of soil carbon sinks.

7. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 6, characterized in that: To acquire new field measurement data and assimilate and calibrate a preliminary soil carbon sink distribution map based on the Kalman filter algorithm, generating an assimilated and calibrated soil carbon sink distribution map, the following steps are included: Locating spatial grid cells corresponding to the geographical locations of the new field measurement data in the preliminary soil carbon sink distribution map; comparing the estimated soil organic carbon density located in the preliminary soil carbon sink distribution map with the measured soil organic carbon values ​​from the new field measurement data to obtain residuals; updating the state and error covariance of all spatial grid cells in the preliminary soil carbon sink distribution map using the residuals based on the Kalman filter algorithm; and generating a spatially continuous assimilated and calibrated soil carbon sink distribution map based on the updated state of the Kalman filter algorithm.

8. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 7, characterized in that: By combining the event log list and the dynamic fusion weight set, the assimilation calibration soil carbon sink distribution map is post-processed to generate the final soil carbon sink measurement result map. The steps include: identifying the corresponding spatial region in the assimilation calibration soil carbon sink distribution map based on the impact range of heavy precipitation events recorded in the event log list; assessing the data reliability level based on the vegetation index characteristic weight change pattern of the dynamic fusion weight set in the corresponding spatial region; overlaying predictive annotations of short-term carbon sink decay on the corresponding spatial region of the assimilation calibration soil carbon sink distribution map according to the data reliability level and the characteristics of the heavy precipitation event; calculating the confidence interval values ​​of the corresponding spatial region based on the data reliability level to obtain the uncertainty level; and generating the final soil carbon sink measurement result map by overlaying the predictive annotations and the uncertainty level annotations on the assimilation calibration soil carbon sink distribution map.

9. The method for calculating urban soil carbon sequestration based on multi-source data as described in claim 8, characterized in that: The weight generation rule is iteratively optimized based on historical task data, including the following steps: collecting event log lists, dynamic fusion weight sets, and corresponding verification error records of the final soil carbon sequestration measurement results from historical tasks to form historical task data; analyzing the historical task data, statistically analyzing the verification errors corresponding to the dynamic fusion weight sets used under different combinations of event types and environmental features, identifying event types and environmental feature combinations with verification errors exceeding the threshold, and determining them as low-reliability combinations that need optimization in the weight generation rule; adjusting the mapping logic corresponding to the low-reliability combinations in the weight generation rule, and updating the weight generation rule using the adjusted mapping logic.

10. A system for measuring urban soil carbon sequestration based on multi-source data, based on the method for measuring urban soil carbon sequestration based on multi-source data as described in any one of claims 1 to 9, characterized in that: This includes an identification module that collects and preprocesses raw data from multiple sources to generate a multi-source feature data cube, analyzes the multi-source feature data cube according to preset rules, identifies potential impact events, and forms an event log list. The weight generation module generates fusion weights for each affected spatial grid cell based on weight generation rules, targeting different data source features, forming a dynamic fusion weight set; the fusion module uses the dynamic fusion weight set to perform weighted fusion of features in the multi-source feature data cube, generating a context-enhanced feature vector. The inference module inputs the context-enhanced feature vector into the pre-trained soil carbon sink basic measurement model to infer a preliminary distribution map of soil carbon sink. The calibration module acquires new field measurement data and performs assimilation calibration on the preliminary distribution map of soil carbon sink based on the Kalman filter algorithm to generate an assimilated and calibrated soil carbon sink distribution map. The iteration module combines the event log list and the dynamically fused weight set to perform post-processing on the assimilated and calibrated soil carbon sink distribution map to generate the final soil carbon sink measurement result map, and iteratively optimizes the weight generation rules based on historical task data.

Images