Carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage

Abstract

The present invention relates to the technical field of dynamic carbon sink accounting, and specifically to a dynamic carbon sink accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage. The system collects optical remote sensing, radar, photosynthetically active radiation, and meteorological data, unifies them in time and space, and fuses them through a cross-modal attention mechanism to generate multimodal feature vectors, which are input into a TCN model to predict carbon reserves and trends. The prediction results and metadata are uploaded to the chain through smart contracts to generate carbon sink NFTs with geofences and confidence levels, realizing credible evidence storage. The residual mapping function is established in combination with historical data on the chain, and the model is dynamically optimized to improve accounting accuracy. The system improves fusion capabilities and prediction accuracy, and enhances the credibility and transparency of carbon asset management and trading.

Description

Technical Field

[0001] The present invention relates to the field of dynamic carbon sink accounting, and specifically to a dynamic carbon sink accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage. Background Art

[0002] As global climate change intensifies, carbon sinks, as a key natural mechanism for regulating atmospheric carbon dioxide concentrations, require dynamic monitoring and accurate accounting, which are crucial for addressing climate change. Achieving highly accurate and verifiable dynamic carbon sink accounting is key to supporting the global carbon trading market.

[0003] However, existing carbon sink monitoring technologies often rely on single remote sensing data, struggling to fully integrate optical, radar, meteorological, and ground-based sensor information, resulting in limited prediction accuracy. Furthermore, data preprocessing and feature extraction lack dynamic adaptive mechanisms, making them ineffective in addressing complex environments such as vegetation obstruction and climate disturbances. Regarding blockchain applications, most systems are limited to simple evidence storage and lack feedback optimization and cross-chain verification mechanisms closely integrated with AI models. This makes it difficult to dynamically update and trustlessly share prediction results, impacting the authority and widespread application of carbon sink assets.

[0004] To this end, the present invention proposes a carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage. Summary of the Invention

[0005] The present invention provides a carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention is based on a multimodal AI remote sensing monitoring and blockchain-based carbon sink dynamic accounting system, including:

[0008] Multimodal data acquisition module, used to obtain target area data and perform unified spatiotemporal calibration;

[0009] The feature fusion and preprocessing module introduces a trainable modal channel weight factor based on the cross-modal attention mechanism driven by vegetation response, and fuses them into a multimodal spatiotemporal feature vector;

[0010] The adaptive AI carbon sink prediction module inputs the multimodal spatiotemporal feature vector into the TCN network to obtain carbon storage prediction values ​​and future trends, and supports zero-knowledge proof cross-chain parameter adjustment;

[0011] The blockchain trusted evidence module encrypts and hashes the prediction results and metadata and uploads them to the consortium chain through a smart contract to generate carbon sink NFT assets with three-dimensional geofence attributes;

[0012] The carbon sink accounting and verification reporting module generates carbon asset reports in accordance with international standards and constructs a residual mapping function based on the historical evidence data on the chain. When the residual between the carbon storage forecast value and the historical certification value exceeds the set threshold, the model iterative optimization is triggered.

[0013] As a preferred technical solution of the invention, the target area data includes optical remote sensing images, radar images, photosynthetically active radiation PAR sensor data, meteorological data and historical carbon sink records.

[0014] As a preferred technical solution of the invention, the unified spatiotemporal calibration includes:

[0015] The remote sensing images are resampled through the image processing library and uniformly converted to the CGCS2000 coordinate system. When distortion exists, projection correction is performed based on the feature points of the ground objects.

[0016] The ground sensor data is generated into a spatial grid through nearest neighbor interpolation and inserted into the corresponding layer. All data timestamps are converted to the UTC standard format. The remote sensing images are aggregated into daily scale layers. The ground data are time-aligned according to the set sliding window and weighted average.

[0017] When the number of consecutive missing values ​​exceeds the set threshold, linear interpolation is used to complete the missing values.

[0018] After data processing is completed, structured data vectors and metadata documents are constructed.

[0019] As a preferred technical solution of the invention, the vegetation response-driven cross-modal attention mechanism introduces a trainable modal channel weight factor, specifically including:

[0020] The modal channels include optical image channel, radar image channel, PAR sensor channel and meteorological data channel. The system performs adaptive weight adjustment based on the dynamic characteristic index of each modal channel. The adjustment is based on the following rules:

[0021] When the slope of the time series of the Normalized Difference Vegetation Index (NDVI) is greater than the set threshold, the attention weights of the optical image channel and the PAR channel are automatically enhanced;

[0022] When the optical remote sensing image is blocked by clouds and the NDVI pixel loss rate exceeds the set ratio, the weight of the radar image channel is enhanced to use its penetration ability to supplement the vegetation structure and texture information;

[0023] When the precipitation or temperature fluctuations in the past three days exceed the set upper limit, the weight of the meteorological data channel is increased to improve the perception of short-term disturbances;

[0024] When the PAR value is lower than the set plant light compensation point, if the temperature is greater than the frost threshold, the system automatically freezes the input of the meteorological channel and assigns it a weight value of 0. If the temperature is less than the frost threshold, the input of the meteorological channel is enhanced.

[0025] The weights of each modal channel are calculated using a differentiable normalization function.

[0026] As a preferred technical solution of the invention, the blockchain trusted evidence module is built on the FISCO-BCOS alliance chain platform, specifically including:

[0027] The carbon sequestration prediction results and their metadata are hashed to generate a unique identifier, which is then combined with the model version, prediction time, and regional code to be packaged into an on-chain transaction.

[0028] The summary and its related information are stored in the alliance chain through a preset smart contract to form an unalterable record;

[0029] The system generates carbon sink asset NFTs based on the ERC-3525 standard, binds the ecological type, predicted value confidence, and geographic fence attributes, and embeds the on-chain evidence hash to achieve the confirmation and traceability of carbon sink assets;

[0030] The module supports on-chain token query and summary verification functions, and is linked to the zero-knowledge verification mechanism.

[0031] As a preferred technical solution of the invention, the zero-knowledge verification mechanism linkage includes:

[0032] Constructing summary hash based on the carbon storage prediction value and its trend sequence generated by the prediction results;

[0033] The system calls the on-chain evidence hash for consistency comparison and generates a non-interactive zero-knowledge proof through the zk-SNARK algorithm;

[0034] Proof-match the off-chain model prediction summary provided by the verification requester with the on-chain record to verify the authenticity of its source and the consistency of the prediction content;

[0035] If the verification is passed, the prediction results and their trend outputs will be allowed to be called and NFT rights confirmation operations will be performed. Otherwise, result writing and asset generation will be rejected to ensure the credibility and source security of the prediction results.

[0036] As a preferred technical solution of the invention, the prediction summary is generated by constructing a structured data vector using sequential field splicing rules, including: predicted value, trend sequence, confidence, model version, ecological label, timestamp and block number, and hash summary processing is performed after splicing.

[0037] As a preferred technical solution of the invention, the carbon asset report includes:

[0038] Estimated value and uncertainty range of carbon sink increment;

[0039] The ecological layer rendering hash generated by the verifiable random function VRF;

[0040] Incremental evidence verification interface based on zero-knowledge proof.

[0041] As a preferred technical solution of the invention, the residual threshold set in the carbon sink accounting and verification reporting module can be dynamically adjusted according to the type of carbon source area, and different eco-labels correspond to different error tolerance standards.

[0042] As a preferred technical solution of the invention, the model iterative optimization specifically includes:

[0043] Based on the carbon sink accounting and verification reporting module, the residual data between the carbon storage prediction results and the historical certified carbon sink values ​​are counted;

[0044] Construct an error mapping function to establish a residual mapping relationship between the predicted value and the corresponding carbon sink increment value in the certification report;

[0045] This mapping function is introduced into the training process as a dynamic weighted term of the loss function, and the model is periodically retrained to achieve adaptive weight convergence and optimization of recognition accuracy.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention achieves efficient integration of multi-source heterogeneous data through multimodal data acquisition and unified spatiotemporal calibration, integrating optical remote sensing, radar imagery, PAR sensors, and meteorological data. This significantly improves the spatiotemporal resolution and data integrity of carbon sink information, laying a solid foundation for subsequent data analysis.

[0048] 2. This invention adopts a vegetation response-driven cross-modal attention mechanism to adaptively adjust the weights of each modal channel, enhance the system's ability to cope with environmental changes and data loss, and significantly improve the effectiveness of feature extraction. In addition, the carbon storage prediction error of this system under complex weather conditions is significantly improved compared to traditional single-modal methods, providing a highly reliable data foundation for carbon sink accounting.

[0049] 3. This invention establishes a closed-loop technology chain encompassing prediction, evidence storage, verification, and optimization. Spatial verifiability is achieved by binding ERC-3525 NFTs to three-dimensional geofencing. Zero-knowledge verification is automatically initiated using model residual anomalies as a trigger. A real-time optimization model for the residual function is constructed based on on-chain historical data, enabling the system to continuously evolve and addressing the inability of traditional static models to dynamically respond to ecological changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 This is a system block diagram of the carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage provided by the present invention;

[0052] Figure 2 This is a flowchart of the implementation of the carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage provided by the present invention. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0054] Example 1:

[0055] Based on multimodal AI remote sensing monitoring and blockchain carbon sink dynamic accounting system, such as Figure 1 Shown, including:

[0056] Multimodal data acquisition module, used to obtain target area data and perform unified spatiotemporal calibration;

[0057] Furthermore, the target area data includes optical remote sensing images, radar images, photosynthetically active radiation PAR sensor data, meteorological data and historical carbon sink records.

[0058] The unified spatiotemporal calibration includes:

[0059] The remote sensing images are resampled through the image processing library and uniformly converted to the CGCS2000 coordinate system. When distortion exists, projection correction is performed based on the feature points of the ground objects.

[0060] The ground sensor data is generated into a spatial grid through nearest neighbor interpolation and inserted into the corresponding layer. All data timestamps are converted to the UTC standard format. The remote sensing images are aggregated into daily scale layers. The ground data are time-aligned according to the set sliding window and weighted average.

[0061] When the number of consecutive missing values ​​exceeds the set threshold, linear interpolation is used to complete the missing values.

[0062] After data processing is completed, structured data vectors and metadata documents are constructed.

[0063] Specifically, the system user first defines the geographic scope of the target monitoring area, specifically inputting geofences, including latitude and longitude boundaries, terrain type, and administrative division or forest management unit number. The system then accesses an existing database of forest and grassland resources and assigns initial eco-labels to the target area, including coniferous forest, broad-leaved forest, grassland, shrubland, and tropical rainforest, for subsequent use by the ecological type identification module.

[0064] The system schedules the following multimodal data sources and collects them regularly according to the set period:

[0065] Optical remote sensing images: Priority access to commercial satellite platforms to obtain 2-meter resolution visible light and near-infrared band image data;

[0066] Radar remote sensing imagery: Accesses SAR data to penetrate clouds and detect changes in surface texture and structure;

[0067] PAR sensor data: photosynthetically active radiation sensors deployed on the ground record the light intensity available to plants per unit time, with a sampling frequency of generally 5 minutes;

[0068] Meteorological data: collected through the National Meteorological Observatory API, including temperature, precipitation, wind speed, humidity, etc.

[0069] Historical carbon sink records: Import carbon storage or carbon sink monitoring reports from past certification cycles as comparison and training labels.

[0070] Due to the heterogeneous data sources and inconsistent spatial coordinate systems of the datasets, the system first performs spatial registration on the optical and radar data, which includes the following sub-steps:

[0071] Call the OpenCV library to implement remote sensing image resampling;

[0072] All data are uniformly converted to the CGCS2000 coordinate system and aligned to a uniform pixel grid;

[0073] If the image has parallax or tilt distortion, perform projection transformation based on the feature points of the ground object;

[0074] The ground sensor point data is converted into spatial grid values ​​using the nearest neighbor method and inserted into the registration layer.

[0075] If the time dimensions of different data sources vary greatly, the system performs the following time alignment processing:

[0076] Take UTC standard time as the unified benchmark;

[0077] Aggregate remote sensing images into daily scale layers according to observation date;

[0078] PAR and meteorological data are averaged by sliding window on a daily basis. The time and step size of the sliding window are set. The data in the window are weighted averaged. When the boundary is less than the window size, the most recent historical value is used to supplement it.

[0079] When data has time gaps, the system first determines whether the missing time period exceeds a set threshold. Linear interpolation is used to fill in all missing time periods. Specifically, the system uses the observation data of two valid time points before and after the missing value as endpoints and uses a linear function to estimate the value of the missing time point in the middle.

[0080] After the interpolation is completed, the system constructs a structured data vector and metadata document, including location coding, time, modality type and data integrity, providing a basis for subsequent data fusion and model prediction stages.

[0081] The feature fusion and preprocessing module introduces a trainable modal channel weight factor based on the cross-modal attention mechanism driven by vegetation response, and fuses them into a multimodal spatiotemporal feature vector;

[0082] Specifically, the standard normalized vegetation index is calculated based on the collected optical remote sensing data. , the formula is as follows:

[0083]

[0084] in represents the reflectivity in the near-infrared band, Indicates the reflectivity of the red light band.

[0085] System Construct a slope change graph based on time series and calculate the growth slope using a local sliding window:

[0086]

[0087] Among them, t is the index of the current time point, indicating a specific time, and k is the length of the time window, indicating how many time units to go back.

[0088] The slope is the dominant factor in the dynamic response of vegetation and is used for weight adjustment in the subsequent attention mechanism. A trainable weight factor is introduced for each modal data source, and the system initializes the following weight channels:

[0089] : optical image channel weight;

[0090] : radar image channel weight;

[0091] : PAR sensor channel weight;

[0092] : Weather channel weight.

[0093] The weight adjustment rules are as follows:

[0094] when When the slope rises, the growth is vigorous, and the and ;

[0095] When the precipitation or temperature fluctuation in the past three days exceeds the set threshold, the ;

[0096] When PAR is lower than the plant light compensation point, freeze the meteorological channel to eliminate environmental interference. .

[0097] The weight calculation is implemented by the following differentiable function:

[0098]

[0099] in is the Sigmoid activation function, which is used to normalize the output range. is the weight sensitization factor of each channel, It is the dynamic index value of the channel, including NDVI slope and temperature change rate.

[0100] After weighting, the eigenvalues ​​of all modal channels are aligned according to the unified timestamp and concatenated into a length of , dimension is The feature vector sequence of is used as the input of the subsequent model. That is:

[0101]

[0102] in Indicates the The multimodal spatiotemporal feature vector after fusion of time points.

[0103] The adaptive AI carbon sink prediction module inputs the multimodal spatiotemporal feature vector into the TCN network to obtain carbon storage prediction values ​​and future trends, and supports zero-knowledge proof cross-chain parameter adjustment;

[0104] Specifically, the system combines the feature fusion with the unified spatiotemporal aligned multimodal feature vector output by the preprocessing module, and constructs a sliding time window according to the set time step. The length of each window is The time series tensor of . The input dimension of this tensor is ,in represents the length of the time series, The multimodal feature dimensions after fusion include NDVI, radar texture, PAR lighting factor, and meteorological indicators.

[0105] In order to improve the convergence speed of the model and avoid the influence of scale differences of different modal data, the data at each time step are standardized so that the mean of all modal features is 0 and the standard deviation is 1.

[0106] The above input tensor is then fed into the TCN structure to obtain the prediction result output in vector form.

[0107] To ensure the uniqueness of the prediction source and the traceability of the prediction process, the system packages the predicted value, trend series, confidence, model version, eco-label, timestamp and region number, performs a hash summary extraction operation, generates a summary fingerprint, and binds it to the predicted geographic location as the input of the trusted evidence module.

[0108] During this process, the system supports cross-chain verification of prediction outputs by verifying the consistency of the prediction summary and its metadata with the existing model version and summary data on the chain through a zero-knowledge verification mechanism. This module does not disclose the specific structure and parameters of the model, and only uses hash summaries and non-interactive zero-knowledge proofs to achieve privacy protection and trusted verification of the prediction path, source, and confidence level.

[0109] The blockchain trusted evidence module encrypts and hashes the prediction results and metadata and uploads them to the consortium chain through a smart contract to generate carbon sink NFT assets with three-dimensional geofence attributes;

[0110] Furthermore, the blockchain trusted evidence module is built on the FISCO-BCOS alliance chain platform, specifically including:

[0111] The carbon sequestration prediction results and their metadata are hashed to generate a unique identifier, which is then combined with the model version, prediction time, and regional code to be packaged into an on-chain transaction.

[0112] The summary and its related information are stored in the alliance chain through a preset smart contract to form an unalterable record;

[0113] The system generates carbon sink asset NFTs based on the ERC-3525 standard, binds the ecological type, predicted value confidence, and geographic fence attributes, and embeds the on-chain evidence hash to achieve the confirmation and traceability of carbon sink assets;

[0114] The module supports on-chain token query and summary verification functions, and is linked to the zero-knowledge verification mechanism.

[0115] Specifically, the system structures and organizes the carbon sequestration results and auxiliary metadata generated within each time segment, including:

[0116] Carbon storage prediction value at the current time; carbon storage prediction value, trend sequence, prediction time, eco-label, model version number; feature input summary information: , radar texture, PAR lighting factor, meteorological indicators; geospatial fence: use polygon coordinate set to represent the prediction area range; timestamp and block number.

[0117] The above content is organized in a structured data vector format and forms a logical "prediction record unit";

[0118] The system extracts the structured vector content from the predicted record unit and concatenates it into a normalized data string according to the set field order. It then performs a digest extraction operation on this data string to generate a fixed-length data fingerprint.

[0119] This digest extraction process uses a well-known one-way encryption algorithm, the core features of which are:

[0120] The same input must produce the same summary;

[0121] Any small change in the input will result in a completely different output summary;

[0122] The original content cannot be inferred through the summary, and it has the ability to hide information and resist tampering.

[0123] This summary is defined as the "Carbon Sink Forecast Summary Identification Code" and is the basis for subsequent chain-up and asset identification.

[0124] The system encapsulates the prediction summary identifier, metadata summary, timestamp, and prediction task number into a "deposit transaction" and prepares to write it into the consortium chain. The consortium chain platform uses a permission-based consensus mechanism for multi-node consensus writing. The specific process includes:

[0125] The evidence transaction is submitted to the blockchain gateway node;

[0126] Multiple consensus nodes verify the data format, signature, and hash legitimacy;

[0127] After reaching a consensus, the system writes the transaction into the latest generated block and records the block number, transaction index number and timestamp.

[0128] After this step is completed, the prediction record can be considered to be permanently stored on the chain. Any attempt to tamper with the record content afterwards will cause the summary verification to fail, thereby ensuring credibility.

[0129] Furthermore, the system generates a carbon sink data asset NFT for each prediction based on the ERC-3525 semi-fungible token standard. This NFT records the following attributes: ecological type label; predicted carbon storage value and unit; confidence level; geofence coordinate point set; on-chain evidence hash (for reverse tracing hash summary); creation time and expiration date fields; NFT issuer ID and ownership relationship.

[0130] NFT is unique, transferable and interface-callable, supporting subsequent transactions, audits, carbon asset valuations and mortgages.

[0131] Furthermore, the system opens an API interface for on-chain asset query and verification, supporting the following operations:

[0132] According to the prediction number, timestamp or geo-fence, the corresponding NFT information is reversed, and any carbon sink data file is input. The system calculates its SHA-256 digest and compares the hash value on the chain to complete the consistency check. The interface returns the verification result, the current status of the NFT, the issuance time, the prediction version and the current owner.

[0133] To ensure the verifiability and privacy of prediction data in a cross-subject environment, the system integrates a zero-knowledge verification mechanism to achieve consistency verification between off-chain data copies and on-chain evidence data. The process is as follows:

[0134] Constructing summary hash based on the carbon storage prediction value and its trend sequence generated by the prediction results;

[0135] The system calls the on-chain evidence hash for consistency comparison and generates a non-interactive zero-knowledge proof through the zk-SNARK algorithm;

[0136] Proof-match the off-chain model prediction summary provided by the verification requester with the on-chain record to verify the authenticity of its source and the consistency of the prediction content;

[0137] If the verification is passed, the prediction results and their trend outputs will be allowed to be called and NFT rights confirmation operations will be performed. Otherwise, result writing and asset generation will be rejected to ensure the credibility and source security of the prediction results.

[0138] Even if the prediction model structure or input features are not disclosed, reliable verification of the prediction path, source, and confidence level can be achieved without leaking sensitive data, further improving the security and practicality of the system in multi-agent collaboration, cross-chain deployment, and financial trusted applications.

[0139] The carbon sink accounting and verification reporting module is used to convert the predicted value into a carbon asset report in accordance with international standards, and iteratively optimize the parameters of the ecological type identification model based on the historical blockchain evidence data.

[0140] The system calls the AI ​​module to generate the carbon storage prediction value , and the geocoding of the target area at the moment of prediction;

[0141] The predicted values ​​were normalized per unit area;

[0142] Convert the predicted value into carbon credit units based on international carbon sink measurement standards and corresponding carbon conversion factors;

[0143] The above data are combined to generate carbon sink accounting entries in a standard format, which serve as the basic input for carbon asset construction.

[0144] The system builds a carbon asset report template, which contains the following fields:

[0145] Incremental carbon storage prediction value; uncertainty range; regional ecological type identification result; prediction time and model version identification; layer rendering hash; zero-knowledge proof check code.

[0146] After all fields are automatically filled in, the system exports the report in PDF and JSON formats for verification by regulatory authorities.

[0147] The system uses a spatial grid map of the prediction area as a base map, overlaying carbon storage prediction values, eco-labels, and prediction confidence levels;

[0148] Render as pseudo-color image and output layer file;

[0149] Use hash algorithm to generate summary of layer file;

[0150] The hash value is used as an embedded field in the report to ensure that the spatial prediction results are non-repudiable.

[0151] The system reads historical prediction values ​​and their corresponding third-party authentication results from the blockchain;

[0152] Predicted values ​​for the same area With authentication value Calculate residuals ;

[0153] Clustering and regression are performed on the residual data under different ecological types to construct a residual mapping model under the ecological type dimension;

[0154] Whenever the amount of new data reaches the preset threshold, the system automatically starts the model backtracking optimization process:

[0155] The residual mapping curve is added to the loss function as a penalty factor;

[0156] Fine-tune the sub-model weights for each ecological type;

[0157] Generate a new model version and replace the old model, and store it on the chain.

[0158] The system constructs zero-knowledge proofs for key fields in the report, proving that the field value was indeed calculated by a certain prediction model for a certain area at a certain time;

[0159] Provide public parameters and validator APIs, allowing regulators or exchanges to verify without exposing the underlying data;

[0160] The zero-knowledge structure ensures that even if the report is published on the chain, no original remote sensing images, geographic locations or specific model structures will be leaked.

[0161] This module implements a closed-loop process of forecasting data → asset reporting → auditable verification → model self-evolution through data standardization, residual feedback optimization mechanism, trusted reporting structure and zero-knowledge encryption verification mechanism, greatly enhancing the practicality and verifiability of the system in real transactions and compliance supervision.

[0162] Example 2:

[0163] In a carbon sequestration project, in order to calculate, authenticate and confirm the dynamic changes of carbon reserves from October 2024 to April 2025, a carbon sequestration dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage was adopted. Figure 2 , the following whole process was implemented:

[0164] First, deploy optical remote sensing, radar remote sensing, ground PAR sensors, and meteorological monitoring stations in the target mountainous area to continuously record daily sunlight, precipitation, vegetation index, etc.

[0165] All remote sensing images were uniformly projected into the CGCS2000 coordinate system and resampled to a 2-meter resolution. For pixels with occlusion in the optical image, pixel values ​​were replaced using the texture channel of the radar image.

[0166] All ground-truth data are used to construct a spatial grid layer. For short-term data interruptions caused by weather or equipment problems, a time series-based linear interpolation algorithm is used to fill in missing values. The completed data are then mapped to the remote sensing image coverage using nearest neighbor interpolation.

[0167] For each type of data, a three-day sliding time window was constructed to extract dynamic indicators such as NDVI, radar backscatter, PAR change rate, precipitation gradient, and temperature change at a daily granularity;

[0168] The above multi-source data are fused into multimodal channels, and the system adjusts the fusion weight of each channel according to the following rules;

[0169] Each channel is compressed to the same numerical scale according to the normalization standard, and the NDVI slope, radar texture change rate, PAR and temperature change are taken as the dominant features and fused according to the weight factor to form dimensional feature vector as subsequent input.

[0170] A general TCN model with multi-layer dilated convolution is systematically constructed to capture the short-term disturbances and long-term trends of carbon sinks over time from multimodal feature sequences.

[0171] After completing the general modeling, the system routes the temporal features to the predefined ecologically specific branching model according to the ecologically labeled shrubs carried by the input samples to perform type-specific predictions.

[0172] The final model output is a carbon storage forecast of 193.6 tons per hectare in the Jishui County project area, and a trend change forecast sequence is given for the next five days. The system also outputs a confidence interval of [186.2, 201.0] based on the confidence parameters used during model training.

[0173] The above carbon storage prediction values, trend series, confidence intervals, and ecological types are packaged, and a hash tool is used to calculate the summary and generate a hash value;

[0174] The hash is uploaded to the consortium chain through the smart contract deployed on the FISCO-BCOS platform, binding the carbon reserves and location codes to form an unalterable on-chain record;

[0175] Based on this evidence, the system generates carbon sink asset NFTs in accordance with the ERC-3525 semi-fungible token standard, recording carbon reserves, ecological types, prediction confidence, spatial location and hash traceability information.

[0176] According to the establishment of asset value update mechanism: deploy oracle contracts to regularly obtain ground sensor data; when the carbon sink change is detected to be more than 5%, a new version of NFT is automatically created and metadata is added to the original NFT;

[0177] The latest valid version is used by default when trading assets, and a historical version traceability interface is provided;

[0178] Holders can query the prediction source, geo-fence and traceability data corresponding to the asset through the on-chain browser.

[0179] The results of the field verification in April 2025 showed that the actual carbon storage in the area was 195.4 tons / hectare, the system calculated the residual per unit area was -1.8 tons / hectare, and the relative error was -0.92%.

[0180] According to the carbon flux deviation record, the residual is collated together with the original prediction features and added to the model feedback training set;

[0181] Reload the shrub model parameters and perform fine-tuning training, retaining the core structure of the original model. During the fine-tuning process, use the sliding window sequence input to recalibrate the TCN network model, and generate a hash summary of the corrected model prediction results again and upload it to the chain for recording.

[0182] Adjacent regions need to deploy similar prediction modules and apply for the latest model weights;

[0183] The system first obtains the model hash summary through on-chain query, and performs consistency verification on the prediction result summary submitted by the caller through a zero-knowledge verification mechanism. After successful verification, the system loads the off-chain model copy into the regional task without the need for retraining, ensuring prediction consistency and deployment efficiency.

[0184] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. Based on multimodal AI remote sensing monitoring and blockchain evidence storage carbon sink dynamic accounting system, it is characterized by: include: Multimodal data acquisition module, used to obtain target area data and perform unified spatiotemporal calibration; The feature fusion and preprocessing module introduces a trainable modal channel weight factor based on the cross-modal attention mechanism driven by vegetation response. The modal channel feature values ​​obtained by the multimodal data acquisition module and their corresponding weight factors are weighted and concatenated to form a multimodal spatiotemporal feature vector. An adaptive AI carbon sink prediction module inputs the multimodal spatiotemporal feature vector into the TCN network to obtain a carbon storage prediction value; The blockchain trusted evidence module encrypts and hashes the predicted value and metadata and uploads them to the alliance chain through a smart contract to generate a carbon sink NFT asset with three-dimensional attributes of the geofence; The carbon sink accounting and verification reporting module generates carbon asset reports according to international standards and constructs a residual mapping function based on historical evidence data on the chain. When the residual between the carbon storage forecast value and the historical certification value exceeds the set threshold, the model iterative optimization is triggered; The vegetation response-driven cross-modal attention mechanism introduces trainable modal channel weight factors, specifically including: The modal channels include optical remote sensing image channel, radar image channel, photosynthetically active radiation channel and meteorological data channel. The system performs weight adaptive adjustment based on the dynamic characteristic indicators of each modal channel. The adaptive adjustment is based on the following rules: When the slope of the time series of the normalized vegetation index is greater than the set threshold, the attention weights of the optical image channel and the photosynthetically active radiation channel are automatically enhanced; When the optical remote sensing image is obscured by clouds and the normalized vegetation index pixel missing rate exceeds the set ratio, the weight of the radar image channel is enhanced; When the precipitation or temperature fluctuation in the past three days exceeds the set upper limit, the weight of the meteorological data channel is increased; When the photosynthetically active radiation value is lower than the set plant light compensation point, if the temperature is greater than the frost threshold, the system automatically freezes the input of the meteorological channel and assigns it a weight value of 0. If the temperature is less than the frost threshold, the input of the meteorological channel is enhanced. The model iterative optimization specifically includes: Based on the carbon sink accounting and verification reporting module, the residual data between the carbon storage forecast value and the historical certified carbon sink value are calculated; Construct an error mapping function to establish a residual mapping relationship between the predicted value and the corresponding carbon sink increment value in the certification report; This mapping function is introduced into the training process as a dynamic weighted term of the loss function, and the model is periodically retrained to achieve adaptive weight convergence and optimization of recognition accuracy.

2. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 1 is characterized in that: The target area data includes optical remote sensing images, radar images, photosynthetically active radiation data, and meteorological data.

3. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 1 is characterized in that: The unified spatiotemporal calibration includes: The remote sensing images are resampled through the image processing library and uniformly converted to the CGCS2000 coordinate system. If the remote sensing images are distorted after projection conversion, projection correction is performed based on the feature points of the ground objects. The ground sensor data is generated into a spatial grid using the nearest neighbor interpolation method and inserted into the corresponding layer. All data timestamps are converted to the UTC standard format. The optical remote sensing image is aggregated into a daily scale layer. The ground sensor data is segmented according to the preset sliding time window, and the weighted average of each segment data is performed to achieve time alignment. When there is a time gap in the data and the missing duration exceeds the set threshold, linear interpolation is used to fill it in; After data processing is completed, structured data vectors and metadata documents are constructed.

4. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 1 is characterized in that: The blockchain trusted evidence module is built on the FISCO-BCOS alliance chain platform and specifically includes: The carbon storage prediction value and its metadata are hashed to generate a unique identifier, which is then combined with the model version, prediction time, and regional code to be packaged into an on-chain transaction. The summary and its related information are stored in the alliance chain through a preset smart contract to form an unalterable record; The system generates carbon sink asset NFTs based on the ERC-3525 standard, binds the ecological type, predicted value confidence, and geographic fence attributes, and embeds the on-chain evidence hash to achieve the confirmation and traceability of carbon sink assets; The module supports on-chain token query and summary verification functions, and is linked to the zero-knowledge verification mechanism.

5. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 4 is characterized in that: The zero-knowledge verification mechanism linkage includes: Perform field concatenation on the carbon storage prediction value and calculate the hash value; The system calls the on-chain evidence hash for consistency comparison and generates a non-interactive zero-knowledge proof through the zk-SNARK algorithm; Proof-match the off-chain model prediction summary provided by the verification requester with the on-chain record to verify the authenticity of its source and the consistency of the prediction content; If the verification is passed, the prediction results and their trend outputs will be allowed to be called and NFT rights confirmation operations will be performed. Otherwise, result writing and asset generation will be rejected to ensure the credibility and source security of the prediction results.

6. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 5 is characterized in that: The prediction summary adopts a sequential field concatenation rule to sequentially combine the predicted value, trend sequence, confidence, model version, eco-label, timestamp and block number into a structured data vector, and performs hash calculation on the vector.

7. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 1 is characterized in that: The carbon asset report includes: Estimated value and uncertainty range of carbon sink increment; The ecological layer rendering hash generated by the verifiable random function VRF; Incremental evidence verification interface based on zero-knowledge proof.

8. The carbon sink dynamic accounting system based on multimodal AI remote sensing monitoring and blockchain evidence storage according to claim 1 is characterized in that: The residual threshold set in the carbon sink accounting and verification reporting module is dynamically adjusted according to the carbon source area type, and different eco-labels correspond to different error tolerance standards.

Images