Dynamic carbon sink accounting system based on multi-modal ai remote sensing monitoring and blockchain-based evidence storage

Abstract

The present invention relates to the technical field of dynamic carbon sink accounting, and specifically relates to a dynamic carbon sink accounting system based on multi-modal AI remote sensing monitoring and blockchain-based evidence storage. The system collects optical remote sensing, radar, photosynthetically active radiation and meteorological data, performs unified spatiotemporal calibration on the data, and then fuses the calibrated data by means of a cross-modal attention mechanism, so as to generate multi-modal feature vectors, and inputs same into a TCN model for carbon stock and trend prediction. A prediction result and metadata are uploaded to a blockchain by means of smart contracts, so as to generate a carbon sink NFT including a geographic fence and a confidence level, thereby realizing trusted evidence storage. A residual mapping function is established in view of on-chain historical data, so as to dynamically optimize the model, and improve the accounting accuracy. The system improves the fusion capability and prediction accuracy, and enhances the credibility and transparency of carbon asset management and transactions.

Description

Carbon sink dynamic accounting system based on multi-modal AI remote sensing monitoring and blockchain storage TECHNICAL FIELD

[0001] The present application relates to the field of carbon sink dynamic accounting, in particular to a carbon sink dynamic accounting system based on multi-modal AI remote sensing monitoring and blockchain storage. BACKGROUND

[0002] With the intensification of global climate change, carbon sink, as an important natural mechanism to regulate atmospheric carbon dioxide concentration, its dynamic monitoring and accurate accounting are of great significance to achieve the carbon neutralization goal. Under the driving of the "double carbon" strategic goal, achieving high-precision and verifiable carbon sink dynamic accounting is the key to supporting the global carbon trading market.

[0003] However, existing carbon sink monitoring technologies mostly rely on single remote sensing data, which is difficult to fully integrate optical, radar, meteorological and ground sensing information, resulting in limited prediction accuracy. At the same time, the data preprocessing and feature extraction lack dynamic adaptive mechanism, which cannot effectively cope with complex environments such as vegetation shielding and climate disturbance. In terms of blockchain application, most systems are limited to simple storage, lack of feedback optimization and cross-chain verification mechanism closely integrated with AI models, and are difficult to realize dynamic updating and credible sharing of prediction results, affecting the authority and application of carbon sink assets.

[0004] Therefore, the present application provides a carbon sink dynamic accounting system based on multi-modal AI remote sensing monitoring and blockchain storage. SUMMARY

[0005] The present application provides a carbon sink dynamic accounting system based on multi-modal AI remote sensing monitoring and blockchain storage.

[0006] To achieve the above purpose, the present application provides the following technical solution:

[0007] The present application provides a carbon sink dynamic accounting system based on multi-modal AI remote sensing monitoring and blockchain storage.

[0008] The multi-modal data acquisition module is used to acquire target area data and perform unified space-time calibration;

[0009] The feature fusion and preprocessing module introduces a trainable modal channel weight factor based on a cross-modal attention mechanism driven by vegetation response, and fuses to form a multi-modal spatio-temporal feature vector;

[0010] The adaptive AI carbon sink prediction module inputs the multi-modal spatio-temporal feature vector into the TCN network to obtain carbon storage prediction value and future trend, and supports zero-knowledge proof cross-chain parameter adjustment;​

[0011] The blockchain credible record module encrypts the prediction result and the metadata, uploads the alliance chain through the smart contract, and generates a carbon sink NFT asset containing a three-dimensional attribute of a geographic fence;

[0012] The carbon sink accounting and verification report module generates a carbon asset report according to international standards, and constructs a residual mapping function based on historical record data on the chain. When the residual of the predicted value of the carbon storage and the historical certified value exceeds a set threshold, the model iteration optimization is triggered.

[0013] As a preferred technical solution of the application, 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 application, the unified space-time calibration includes:

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

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

[0017] When the consecutive missing data exceeds a set threshold, linear interpolation is used for completion.

[0018] After data processing, a structured data vector and a metadata document are constructed.

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

[0020] The modal channel includes an optical image channel, a radar image channel, a PAR sensor channel, and a meteorological data channel. The system adaptively adjusts the weight according to the dynamic characteristic index of each modal channel. The adjustment is based on the following rules:

[0021] When the time series slope of the normalized difference vegetation index (NDVI) is greater than a set threshold, the attention weight of the optical image channel and the PAR channel is automatically enhanced;

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

[0023] When the precipitation or temperature fluctuation amplitude in the next three days exceeds a set upper limit, the weight of the meteorological data channel is enhanced 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 at this time is greater than the frost threshold, the system automatically freezes the input of the meteorological channel, giving it a weight value of 0, and if the temperature at this time is less than the frost threshold, the input of the meteorological channel is enhanced;

[0025] The weight of each modal channel is calculated by a differentiable normalization function.

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

[0027] The carbon sink prediction result and its metadata are hashed to generate a unique identifier, and are encapsulated into an on-chain transaction combined with the model version, prediction time and regional code;

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

[0029] The system generates a carbon sink asset NFT according to the ERC-3525 standard, binds the ecological type, prediction value confidence and geographic fence attributes, and embeds the on-chain evidence hash to realize the right and traceability of the carbon sink asset;

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

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

[0032] The carbon storage prediction value and its trend sequence generated based on the prediction result are used to construct a summary hash;

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

[0034] The off-chain model prediction summary provided by the verification requester is matched with the on-chain record to prove the authenticity of the source and the consistency of the prediction content;

[0035] If the verification is passed, the prediction result and its trend output are allowed to be called and the NFT right operation is performed, otherwise the result is refused to be written and the asset is generated, ensuring the credibility and safety of the source of the prediction result.

[0036] As a preferred technical solution of the application, the prediction summary is structured by splicing rules of sequential fields to generate a structured data vector, including: prediction value, trend sequence, confidence, model version, ecological label, timestamp and block number, and after splicing, the hash summary is processed.

[0037] As a preferred technical solution of the application, the carbon asset report comprises:

[0038] Carbon sink increment estimation value and uncertainty range;

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

[0040] The increment storage certificate verification interface based on zero-knowledge proof.

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

[0042] As a preferred technical solution of the application, the model iteration optimization specifically comprises:

[0043] Based on the carbon sink accounting and verification report module, the residual data between the carbon storage prediction result and the historical certified carbon sink value is counted;

[0044] The error mapping function is constructed to establish a residual mapping relationship between the predicted value and the corresponding certified report carbon sink increment value;

[0045] The mapping function is introduced as a dynamic weighting item of the loss function into the training process, and the model is periodically retrained to realize weight adaptive convergence and recognition accuracy optimization.

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

[0047] 1. The application realizes efficient integration of multi-source heterogeneous data by multi-modal data acquisition and unified space-time calibration, and significantly improves the space-time resolution and data integrity of carbon sink information, laying a solid foundation for subsequent data analysis.

[0048] 2. The application adopts a vegetation response driven cross-modal attention mechanism to adaptively adjust the weight of each modal channel, enhance the system's ability to respond to environmental changes and data loss, significantly improve the effectiveness of feature extraction, and the carbon storage prediction error of the system under complex weather is significantly improved compared with traditional single-modal methods, providing a high-reliability data basis for carbon sink accounting.

[0049] 3. The application constructs a "prediction-storage-verification-optimization" whole-chain technology closed loop. Spatial verifiability is realized by binding ERC-3525 NFT through three-dimensional geographic fence; the zero-knowledge verification is automatically started with model residual anomaly as the trigger condition; the residual function is constructed based on the historical data on the chain to optimize the model in real time, giving the system the ability to continuously evolve, solving the problem that traditional static models cannot dynamically respond to ecological changes. BRIEF DESCRIPTION OF DRAWINGS

[0050] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are intended to serve as an exemplification of the application, and do not limit the application. In the drawings:

[0051] Fig. 1 is a system block diagram of a multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system provided by the present application;

[0052] Fig. 2 is an implementation flowchart of a multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system provided by the present application. DETAILED DESCRIPTION

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

[0054] The preferred embodiments of the present application will be described below with reference to the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0055] Embodiment one:

[0056] The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system, as shown in Fig. 1, comprises:

[0057] A multi-modal data acquisition module is used to acquire target area data and perform unified space-time calibration.

[0058] Further, the target area data comprises optical remote sensing images, radar images, photosynthetic active radiation (PAR) sensor data, meteorological data and historical carbon sink records.

[0059] The unified space-time calibration comprises:

[0060] The remote sensing images are resampled through an image processing library and are uniformly converted to the CGCS2000 coordinate system. When there is distortion, projection correction is performed based on feature points of ground objects.

[0061] The ground sensor data is generated into a spatial grid through nearest neighbor interpolation, is inserted into a corresponding layer, and all data timestamps are converted into UTC standard format. Remote sensing images are aggregated into daily scale layers, and ground data is time-aligned through a set sliding window and weighted average.

[0062] When continuous missing data exceeds a set threshold, linear interpolation is used to complete the missing data.

[0063] After data processing is completed, a structured data vector and a metadata document are constructed.

[0064] Specifically, first, the system user sets the geographical range of the target monitoring area, specifically in the form of a geofence, including latitude and longitude boundaries, terrain types, administrative regions, or forest management unit numbers. The system calls the existing forest and grass resources database to add initial ecological labels to the target area, including coniferous forest, broadleaf forest, grassland, shrub, and tropical rainforest, for subsequent ecological type identification module calls.

[0065] The system schedules the following multi-modal data sources and collects them according to the set period:

[0066] Optical remote sensing images: preferentially access commercial satellite platforms to obtain 2-meter resolution visible and near-infrared band image data;

[0067] Radar remote sensing images: access SAR data for detecting surface texture and structural changes through clouds;

[0068] PAR sensor data: deploy photosynthetically active radiation sensors on the ground to record the available light intensity for plants per unit time, with a sampling frequency of 5 minutes;

[0069] Weather data: collected through the National Meteorological Observatory API, including temperature, precipitation, wind speed, humidity, etc.;

[0070] Historical carbon sink records: import carbon storage or carbon sink monitoring reports from previous certification periods as comparison and training labels.

[0071] Due to the heterogeneity of data sources and the inconsistency of spatial coordinate systems of each dataset, the system first performs spatial registration on optical and radar data, including the following sub-steps:

[0072] Call the OpenCV library to realize remote sensing image resampling;

[0073] Convert all data to the CGCS2000 coordinate system and align to a unified pixel grid;

[0074] If the image has parallax or tilt distortion, perform projection transformation based on feature points;

[0075] Generate spatial grid values for ground sensor point data using the nearest neighbor method and insert them into the registration layer.

[0076] For different data sources with large differences in time dimension, the system performs the following time alignment processing:

[0077] Use UTC standard time as the unified reference;

[0078] Aggregate remote sensing images by observation date into daily scale layers;

[0079] The PAR and meteorological data are windowed and averaged by day, and the time and step size of the sliding window are set. The data in the window are averaged by weighting, and the recent historical values are used to supplement the insufficient window size.

[0080] When there is a time gap in the data, the system first determines whether the missing duration exceeds the set threshold. For all missing time periods, linear interpolation method is used to complete, which is: taking the observation data of the adjacent two effective time points before and after the missing value as the end point, and estimating the numerical value of the missing time point in the middle through linear function.

[0081] After interpolation, the system constructs a structured data vector and a metadata document, including location encoding, time, modal type and data integrity, which provides the basis for subsequent data fusion and model prediction stage.

[0082] The feature fusion and preprocessing module introduces trainable modal channel weight factors based on vegetation response driven cross-modal attention mechanism, and fuses to form a multi-modal spatio-temporal feature vector;

[0083] Specifically, the standard normalized vegetation index NDVI is calculated for the collected optical remote sensing data , the formula is as follows:

[0084]

[0085] Wherein represents the near-infrared band reflectivity, represents the red band reflectivity.

[0086] The system constructs a slope change graph according to the time sequence , and uses a local sliding window to calculate the growth slope:

[0087]

[0088] Where, t is the index of the current time point, represents a specific time, and k is the length of the time window, which represents how many time units to backtrack.

[0089] The slope, as the dominant factor of vegetation dynamic response, is used for weight adjustment of the subsequent attention mechanism. A trainable weight factor is introduced for each modal data source, and the system initializes the following weight channels:

[0090] : optical image channel weight;

[0091] : radar image channel weight;

[0092] : PAR sensor channel weight;

[0093] : weather channel weight.

[0094] The weight adjustment rule is as follows:

[0095] When The slope rises, that is, the growth is vigorous, and the weight is increased And ;

[0096] When the precipitation or temperature fluctuation amplitude in the last 3 days exceeds the set threshold, the weight is increased ;

[0097] When the PAR is lower than the light compensation point of the plant, the weather channel is frozen to eliminate environmental interference, and the weight is set .

[0098] The weight calculation is realized by the following differentiable function:

[0099]

[0100] wherein is a Sigmoid activation function for normalizing the output range, is a weight sensitivity factor of each channel, is the dynamic index value of the channel, including the NDVI slope and the temperature change rate.

[0101] After the feature values of all modal channels are weighted, they are aligned according to a unified timestamp and spliced into a feature vector sequence with a length of and a dimension of , which is used as the input of the subsequent model. That is:

[0102]

[0103] wherein represents the multi-modal spatio-temporal feature vector after fusion at the th time point.

[0104] The adaptive AI carbon sink prediction module inputs the multi-modal spatio-temporal feature vector into a TCN network to obtain the carbon storage prediction value and future trend, and supports zero-knowledge proof cross-chain parameter adjustment;

[0105] Specifically, the system constructs a sliding time window according to the set time step based on the unified spatio-temporal aligned multi-modal feature vector output by the feature fusion and preprocessing module, and each window constructs a time series tensor with a length of . The input dimension of the tensor is , wherein represents the time series length, The post-fusion multi-modal feature dimension includes NDVI, radar texture, PAR light factor, and meteorological index.

[0106] To improve the convergence speed of the model and avoid the influence of the scale difference of different modal data, the data at each time step is standardized to make the mean of all modal features 0 and the standard deviation 1.

[0107] The above input tensor is then sent into the TCN structure to obtain the prediction result output in vector form.

[0108] To ensure the uniqueness of the prediction source and the traceability of the prediction process, the system packages the prediction value, trend sequence, confidence, model version, ecological label, timestamp, and region number, and then performs a hash digest extraction operation to generate a digest fingerprint, which is bound to the predicted geographic location as input to the trusted evidence module.

[0109] In this process, the system supports consistency checking of the prediction digest and its metadata with the existing model version and digest data on the chain through a zero-knowledge verification mechanism, realizing cross-chain verification of the prediction output. This module does not disclose the specific structure and parameters of the model, but only realizes privacy protection and trusted verification of the prediction path, source, and confidence content through hash digest and non-interactive zero-knowledge proof.

[0110] The blockchain trusted evidence module encrypts and hashes the prediction results and metadata, and uploads them to the alliance chain through a smart contract to generate carbon sink NFT assets containing three-dimensional attributes of the geographic fence;

[0111] Further, the blockchain trusted evidence module is constructed based on the FISCO-BCOS alliance chain platform, specifically including:

[0112] The carbon sink prediction results and their metadata are hashed to generate a unique identifier, which is packaged with the model version, prediction time, and region code as an on-chain transaction;

[0113] The digest and its associated information are stored in the alliance chain through a preset smart contract to form an unalterable record;

[0114] The system generates carbon sink asset NFTs according to the ERC-3525 standard, binds ecological types, prediction value confidence, and geographic fence attributes, and embeds on-chain evidence hashes to realize the right protection and traceability of carbon sink assets;

[0115] The module supports on-chain token query and digest verification functions, and is linked with the zero-knowledge verification mechanism.

[0116] Specifically, the system structures and organizes the carbon sink results and auxiliary metadata generated in each time slice, including:

[0117] Carbon stock prediction value at current time; carbon stock prediction value, trend sequence, prediction time, ecological label, model version number; feature input abstract information: Radar texture, PAR light factor, meteorological index; Geospatial fence: use polygon coordinate set to represent the prediction area range; Timestamp and block number.

[0118] The above content is arranged in a structured data vector format, and forms a logical "prediction record unit";

[0119] The system extracts the above structured vector content from the prediction record unit, and splices it into a set of standardized data strings according to the set field order. Then the data string is executed abstract extraction operation, generating a fixed length data fingerprint.

[0120] This abstract extraction process uses a well-known one-way encryption algorithm, which has the following core features:

[0121] The same input must produce the same abstract;

[0122] Any slight change in input will result in a completely different output abstract;

[0123] The original content cannot be inferred from the abstract, and has information hiding and tamper resistance.

[0124] This abstract is defined as "carbon sink prediction abstract identification code", which is the basis for subsequent on-chain and asset identification.

[0125] The system encapsulates the prediction abstract identification code together with the metadata abstract, timestamp and prediction task number into a "certificate storage transaction" and prepares to write it into the alliance chain. The alliance chain platform uses a permission-based consensus mechanism for multi-node consensus writing, and the specific process includes:

[0126] The certificate storage transaction is submitted to the blockchain gateway node;

[0127] The data format, signature and hash legality are verified by multiple consensus nodes;

[0128] After consensus is reached, the system writes the transaction into the latest generated block, and records the block number, transaction index number and timestamp.

[0129] After this step is completed, it is considered that the prediction record has been permanently stored on the chain, and any attempt to tamper with the record content afterwards will result in abstract verification failure, thereby ensuring credibility.

[0130] Further, the system generates a carbon sink data asset NFT for each prediction behavior according to the ERC-3525 semi-homogeneous token standard. The NFT records the following attributes: ecological type label; predicted carbon storage value and unit; confidence; geofencing coordinate point set; on-chain notarization hash: for reverse tracing hash digest; creation time and validity period field; NFT issuer ID and ownership relationship.

[0131] The NFT has uniqueness, transferability, and interface invocability, supporting subsequent transactions, audits, carbon asset valuation, and mortgage behaviors.

[0132] Further, the system opens up on-chain asset query and verification API interfaces, supporting the following operations:

[0133] According to the prediction number, timestamp, or geofencing, the corresponding NFT information is retrieved. By inputting any carbon sink data file, the system calculates its SHA-256 digest and compares it with the on-chain hash value to complete consistency verification. The interface returns the verification result, NFT current state, issuance time, prediction version, and current owner.

[0134] To ensure the verifiability and privacy security 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 notarized data. The process is as follows:

[0135] Based on the prediction result, the carbon storage prediction value and its trend sequence are constructed into a digest hash.

[0136] The system calls the on-chain notarization hash for consistency comparison and generates a non-interactive zero-knowledge proof through zk-SNARK algorithm.

[0137] The off-chain model prediction digest provided by the verification requester is matched with the on-chain record to verify its source authenticity and prediction content consistency.

[0138] If the verification is successful, the prediction result and its trend output are allowed to be called and the NFT ownership operation is performed. Otherwise, the result is rejected from being written and the asset is generated, ensuring the credibility and source security of the prediction result.

[0139] Even without disclosing the prediction model structure or input features, the prediction path, source, and confidence can be verified without revealing sensitive data, further improving the security and practicality of the system in multi-subject collaboration, cross-chain deployment, and financial trust applications.

[0140] The carbon sink accounting and verification report module is used to convert the prediction value into a carbon asset report according to international standards and iteratively optimize the ecological type identification model parameters based on the historical notarization data on the blockchain.

[0141] Carbon storage prediction value generated by system call AI module , and target area geocoding at prediction time;

[0142] Standardize the prediction value per unit area;

[0143] Convert the prediction value into carbon credit units according to international carbon sink measurement standards and corresponding carbon conversion factors;

[0144] Combine the above data to generate carbon sink accounting items in standard format as the basis for building carbon assets.

[0145] The system builds a carbon asset report template, including the following fields:

[0146] Incremental carbon storage prediction value; uncertainty range; regional ecological type identification result; prediction time and model version identification; layer rendering hash; zero-knowledge proof verification code.

[0147] After all fields are automatically filled in, the system exports the report in PDF and JSON formats for regulatory agencies to verify.

[0148] The system uses the spatial grid map of the predicted area as the base map, and superimposes the carbon storage prediction value, ecological label, and prediction confidence;

[0149] Rendered as a pseudo-color map and output layer file;

[0150] Use a hash algorithm to generate a digest of the layer file;

[0151] Take this hash value as an embedded field in the report to ensure the spatial prediction result is non-repudiable.

[0152] The system reads historical prediction values and their corresponding third-party authentication results from the blockchain;

[0153] Calculate the residual error of the prediction value and the authentication value for the same area ;

[0154] Cluster and regress residual data under different ecological types to build a residual mapping model under the ecological type dimension;

[0155] Whenever the amount of new data reaches the preset threshold, the system automatically starts the model backtracking optimization process:

[0156] Add the residual mapping curve as a penalty factor to the loss function;

[0157] Fine-tune the training of the sub-model weight for each ecological type;

[0158] Generate a new model version and replace the old model, while storing it on the chain for evidence.

[0159] The system builds zero-knowledge proof for key fields in the report, proving that the field value is indeed calculated by a certain prediction model at a certain time for a certain area;

[0160] Provide public parameters and verifier API to allow regulatory units or exchanges to verify without exposing underlying data;

[0161] Zero-knowledge structure ensures that even if the report is published on the chain, it does not leak any original remote sensing images, geographic locations, or specific model structures.

[0162] This module realizes the closed-loop process of prediction data → assetized report → auditable verification → model self-evolution through data standardization, residual feedback optimization mechanism, credible report structure and zero-knowledge encryption verification mechanism, greatly enhancing the practicality and verifiability of the system in real transactions and compliance supervision.

[0163] Embodiment two:

[0164] In the forestry carbon sink project of Jishui County, Jiangxi Province, in order to account for the dynamic changes of carbon storage from October 2024 to April 2025, adopt the multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system, as shown in Figure 2, and implement the following whole process:

[0165] First, optical remote sensing, radar remote sensing, ground PAR sensors and meteorological monitoring stations are deployed in the target mountainous area to continuously record daily light, precipitation, vegetation index, etc.

[0166] All remote sensing images are resampled to 2-meter resolution after being projected to the CGCS2000 coordinate system. For pixels with occlusion in optical images, the pixel value is replaced through the texture channel of radar images.

[0167] All ground measured data are constructed into spatial grid layers, and short-term missing data caused by weather or equipment reasons are filled using linear interpolation algorithm based on time series, and then the data after filling process are mapped to the coverage range of remote sensing images using nearest neighbor interpolation.

[0168] For each type of data, a three-day sliding time window is constructed, and daily granularity is extracted for NDVI, radar backscatter, PAR rate of change, precipitation gradient, and temperature change dynamic indicators.

[0169] The above multi-source data are fused into multi-modal channels, and the system adjusts the fusion weight of each channel according to the following rules;

[0170] Each channel is compressed to the same numerical scale according to the normalization standard, and NDVI slope, radar texture rate of change, PAR and temperature change are taken as the dominant features, and are fused into The eigenvector is used as the subsequent input.

[0171] The system constructs a general TCN model with multi-layer dilated convolution to capture short-term disturbances and long-term trends of carbon sinks over time from multi-modal feature sequences.

[0172] After completing general modeling, the system routes the time series features to the pre-defined ecologically specific branch model according to the ecological label carried by the input sample, to perform type-specific prediction.

[0173] The final model output of the Jishui County project area carbon storage prediction value is 193.6 tons / acre, and the trend change prediction sequence after 5 days is given, and the system outputs the confidence interval as [186.2, 201.0] according to the confidence parameter in the model training process;

[0174] Pack the above carbon storage prediction value, trend sequence, confidence interval, and ecological type, calculate the digest using the hash tool, and generate the hash value;

[0175] Upload this hash to the alliance chain through the smart contract deployed on the FISCO-BCOS platform, bind the carbon storage and location code, and form an unalterable on-chain record;

[0176] Based on this evidence, the system generates carbon sink asset NFT according to the ERC-3525 semi-homogeneous token standard, recording information including carbon storage, ecological type, prediction confidence, spatial location, and hash traceability.

[0177] According to the establishment of asset value updating mechanism: deploy the oracle contract to obtain ground sensor data at regular intervals; When the detected carbon sink changes more than 5%: automatically create a new version of NFT, and add metadata to the original NFT;

[0178] The default asset transaction uses the latest valid version, and provides a historical version tracing interface;

[0179] The holder can query the prediction source, geographic fence, and traceability data corresponding to the asset through the on-chain browser.

[0180] The field verification result in April 2025 shows that the actual measured carbon storage in this area is 195.4 tons / acre, and the system calculates the unit area residual as -1.8 tons / acre, with a relative error of -0.92%.

[0181] According to the carbon flux deviation record, the residual is sorted together with the original prediction features and added to the model feedback training set;

[0182] Reload the shrub model parameters and perform fine-tuning training, keep the original model core structure, use the sliding window sequence input to correct the TCN network model during the fine-tuning process, generate the hash digest of the prediction result of the corrected model again and upload it to the chain for recording.

[0183] The same type of prediction module needs to be deployed in adjacent areas, and the latest model weight is called by the application;

[0184] The system first obtains the model hash digest through on-chain query, and then performs consistency check on the prediction result digest submitted by the calling party through the zero-knowledge verification mechanism. After successful verification, the system loads the off-chain model copy to the regional task, without the need for retraining, ensuring the consistency of prediction and the efficiency of deployment.

[0185] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for the purpose of limiting the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it still can be modified, or part of the technical features of the equivalent replacement. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the scope of the present application.

Claims

1. A dynamic accounting system for carbon sinks based on multi-modal AI remote sensing monitoring and blockchain storage, characterized in that, Comprise: A multi-modal data acquisition module for acquiring target area data and performing unified space-time calibration; A feature fusion and preprocessing module that introduces trainable modal channel weight factors based on a vegetation response driven cross-modal attention mechanism to fuse and form a multi-modal space-time feature vector; An adaptive AI carbon sink prediction module that inputs the multi-modal space-time feature vector into a TCN network to obtain carbon storage prediction values and future trends, supporting zero-knowledge proof cross-chain parameter adjustment; A blockchain trusted storage module that encrypts and hashes the prediction results and metadata and uploads them to the alliance chain through a smart contract to generate carbon sink NFT assets with three-dimensional attributes of geographic fences; A carbon sink accounting and verification report module that generates carbon asset reports based on international standards and constructs a residual mapping function based on historical storage data on the chain. When the residual error between the carbon storage prediction value and the historical certified value exceeds the set threshold, the model is triggered for iterative optimization. The target area data includes optical remote sensing images, radar images, photosynthetically active radiation (PAR) sensor data, weather data, and historical carbon sink records.

2. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, The unified space-time calibration includes:

3. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, Resampling the remote sensing images through an image processing library and converting them to the CGCS2000 coordinate system. When there is distortion, perform projection correction based on feature points. Generate a spatial grid by nearest neighbor interpolation for ground sensor data, insert it into the corresponding layer, and convert all data timestamps to UTC standard format. Aggregate remote sensing images into daily scale layers and time align ground data by setting the sliding window and weighted average. When consecutive missing data exceeds the set threshold, use linear interpolation to complete it. After data processing, construct a structured data vector and metadata document. The vegetation response driven cross-modal attention mechanism introduces trainable modal channel weight factors, which specifically include:

4. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, The modal channels include optical image channels, radar image channels, PAR sensor channels, and weather data channels. The system adjusts the weights adaptively based on the dynamic characteristics of each modal channel. The adjustment is based on the following rules: When the time series slope of the normalized difference vegetation index (NDVI) is greater than the set threshold, automatically enhance the attention weights of the optical image channel and the PAR channel. When the optical remote sensing image is obscured by clouds and the NDVI pixel missing rate exceeds the set ratio, increase the weight of the radar image channel to supplement the vegetation structure texture information. When the precipitation or temperature fluctuation amplitude exceeds the set upper limit within the next three days, increase the weight of the weather data channel to improve the perception of short-term disturbances. When the PAR value is below the set plant light compensation point, if the temperature is greater than the frost threshold, the system automatically freezes the input of the weather channel and assigns it a weight value of 0. If the temperature is less than the frost threshold, increase the input of the weather channel. The weights of each modal channel are calculated through a differentiable normalization function. The blockchain trusted storage module is based on the FISCO-BCOS alliance chain platform and specifically includes:

5. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, Hash the carbon sink prediction results and their metadata to generate a unique identifier, and combine the model version, prediction time, and region code to package it as a chain transaction. ​ The summary and its associated information are stored in the alliance chain through a preset smart contract, forming an unalterable record; The system generates carbon sink asset NFTs according to the ERC-3525 standard, binds ecological types, predicted value confidence, and geographic fence attributes, and embeds on-chain storage hashes to realize the right and traceability of carbon sink assets; The module supports on-chain token query and summary verification functions, and is linked with a zero-knowledge verification mechanism.

6. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 5, characterized in that, The zero-knowledge verification mechanism linkage includes: The carbon storage prediction value and its trend sequence generated based on the prediction results are used to construct a summary hash; The system calls the on-chain storage hash for consistency comparison, and generates a non-interactive zero-knowledge proof through a zk-SNARK algorithm; The off-chain model prediction summary provided by the verification requester is matched with the on-chain record to prove its authenticity and consistency of the prediction content; If the verification is passed, the prediction result and its trend output are allowed to be called and the NFT right operation is performed, otherwise the result is rejected to be written and the asset is generated, ensuring the credibility and safety of the prediction result.

7. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 6, characterized in that, The prediction summary is generated by constructing a structured data vector using a sequential field splicing rule, and the field sequence includes: prediction value, trend sequence, confidence, model version, ecological label, timestamp, and block number. After splicing, the hash summary is processed.

8. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, The carbon asset report includes: Carbon sink increment estimation value and uncertainty range; Ecological layer rendering hash generated by a verifiable random function VRF; Increment storage verification interface based on zero-knowledge proof.

9. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, The residual threshold set in the carbon sink accounting and verification report module can be dynamically adjusted according to the carbon source area type, and different ecological labels correspond to different error tolerance standards.

10. The multi-modal AI remote sensing monitoring and blockchain storage carbon sink dynamic accounting system based on claim 1, characterized in that, The model iteration optimization specifically includes: Based on the carbon sink accounting and verification report module, the residual data between the carbon storage prediction result and the historical certified carbon sink value is counted; An error mapping function is constructed to establish a residual mapping relationship between the prediction value and the corresponding certified report carbon sink increment value; The mapping function is introduced as a dynamic weighting item of the loss function into the training process to periodically retrain the model, so as to realize weight adaptive convergence and recognition accuracy optimization.

Images