A highland wetland carbon sink intelligent prediction and optimization system based on a multi-modal large model

Abstract

The application discloses a kind of highland wetland carbon sink intelligent prediction and optimization system based on multimodal big model, and relates to the technical field of ecological environment monitoring.The system includes: multi-modal data sensing module, for obtaining multi-source heterogeneous data;Highland environment adaptive coding module, encode the unique environmental factors of highland into model input features;Multimodal fusion big model module, realize the deep fusion of multi-modal features through cross-modal alignment and Transformer fusion layer;Freezing and thawing perception attention mechanism, model the freezing and thawing cycle of highland wetland, predict the influence of freezing and thawing on carbon flux;Carbon sink prediction module, based on the fusion feature output CO2 flux, CH4 flux and the prediction result of net ecosystem exchange;Optimization decision module, generate optimization management suggestions based on the prediction result.The application realizes high-precision, multi-time-scale carbon sink intelligent prediction and optimization decision.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and ecological environment monitoring technology, specifically to an intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model. Background Technology

[0002] As a core component of the Qinghai-Tibet Plateau ecosystem, plateau wetlands are not only an important ecological security barrier for my country but also a key link in the global carbon cycle. Their vast soil organic carbon pool plays an irreplaceable role in achieving dual carbon targets and addressing global climate change. Compared to plain wetlands, plateau wetlands possess unique characteristics such as high altitude, low oxygen partial pressure, large diurnal temperature range, significant freeze-thaw cycles, and fragile and sensitive ecosystems. Their carbon sequestration processes are coupled and regulated by multiple factors including hydrology, meteorology, vegetation, and soil freeze-thaw cycles, exhibiting strong spatiotemporal heterogeneity and nonlinearity, posing significant challenges to the accurate prediction and scientific management of carbon sequestration.

[0003] Currently, technologies related to carbon sequestration monitoring and prediction in plateau wetlands can be mainly divided into three categories: The first is in-situ monitoring technology based on ground stations, using eddy covariance flux towers as the core. This can achieve high-frequency continuous observation of carbon flux, but it suffers from inherent drawbacks such as high deployment costs, difficult operation and maintenance, and sparse site coverage. In the harsh environment and inconvenient transportation of the Qinghai-Tibet Plateau, it is difficult to achieve large-scale, high-density spatial coverage, thus failing to meet the needs of regional-scale carbon sequestration assessment. The second is monitoring technology based on remote sensing inversion. Using satellite and UAV remote sensing data, large-scale spatial inversion of wetland carbon sequestration can be achieved. However, it can only acquire surface information such as surface spectra and structure, and it is difficult to effectively integrate soil physicochemical properties, microbial activity, etc. The accuracy of inversion based on underground and dynamic information such as meteorological time series changes and freeze-thaw processes is greatly affected by atmospheric correction and topography, and it cannot achieve forward-looking predictions across multiple time scales. Thirdly, there are carbon cycle process models and machine learning prediction models. Among them, process models such as DNDC and Biome-BGC are built based on ecological mechanisms and can simulate the biogeochemical processes of the carbon cycle, but they require numerous input parameters and are complex to calibrate. They also have poor adaptability to the unique environment of plateau wetlands, such as low oxygen and freeze-thaw cycles, and exhibit large simulation errors in highly heterogeneous plateau scenarios. Furthermore, existing machine learning models are mostly built based on single-modal or limited-modal data, and their application in the prediction and management of carbon sinks in plateau wetlands faces the following core technical bottlenecks: First, the ability to deeply integrate multi-source heterogeneous data is insufficient. Data related to carbon sinks in plateau wetlands encompasses multimodal data such as remote sensing images, time-series flux data, geospatial data, textual expert knowledge, and scientific literature. Most existing technologies only achieve simple splicing and shallow fusion of different modal data, lacking effective cross-modal alignment mechanisms. This makes it difficult to fully explore the complementary information and intrinsic relationships between different modal data, and thus fails to comprehensively characterize the complex driving mechanisms of carbon sinks in plateau wetlands, resulting in limited prediction accuracy.

[0004] Second, there is a lack of adaptive modeling capabilities for plateau-specific environmental factors. Existing models do not have specific coding and modeling mechanisms designed for core unique environmental factors of plateau wetlands, such as altitude, hypoxia, and freeze-thaw cycles. In particular, for the freeze-thaw cycle, a key process regulating carbon flux in plateau wetlands, the models cannot accurately capture the nonlinear effects of freeze-thaw cycles on soil microbial metabolism, vegetation photosynthesis, and respiration. This makes it difficult to accurately predict the drastic fluctuations in carbon flux during freeze-thaw sensitive periods, and the models have extremely poor adaptability to plateau scenarios.

[0005] Third, the generalization ability of models is severely insufficient in small sample scenarios. Due to the harsh natural environment of the plateau, long-term continuous monitoring data of plateau wetlands is scarce and labeled samples are lacking. Existing pure data-driven machine learning models are prone to overfitting in small sample scenarios and cannot quickly adapt to different regions and types of plateau wetlands. The transferability and application capability of the models are limited. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide an intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model, comprising: The multimodal data sensing module is used to acquire remote sensing image data, ground monitoring flux data, meteorological data, UAV inspection data, and laboratory analysis data of plateau wetlands. The plateau environment adaptive coding module is used to encode plateau-specific environmental factors such as altitude, temperature, oxygen content, and freeze-thaw cycle into model input features; The multimodal fusion large model module includes an image encoder, a sequence encoder, a geoencoder, and a text encoder. The image encoder is used to extract visual features from remote sensing images, the sequence encoder is used to extract time-series features from flux data and meteorological data, the geoencoder is used to extract spatial topological features of wetlands, and the text encoder is used to extract semantic features from scientific literature and expert knowledge. The multimodal fusion large model module achieves deep fusion of multimodal features through cross-modal alignment and a Transformer fusion layer. The freeze-thaw sensing attention mechanism module is used to model the freeze-thaw cycle of plateau wetlands and predict the impact of freeze-thaw on carbon flux. The carbon sink prediction module is used to output prediction results of CO2 flux, CH4 flux and net ecosystem exchange based on the fused multimodal features; The optimization decision-making module is used to generate optimized management suggestions for improving carbon sequestration in plateau wetlands based on the prediction results.

[0008] Furthermore, the multimodal data sensing module specifically includes: The remote sensing image acquisition submodule is used to receive satellite remote sensing data, including Sentinel-2 multispectral images, Landsat-8 images, and Gaofen series satellite images, and extract spectral indices. The ground monitoring flux data acquisition submodule is used to connect to the eddy covariance flux tower to acquire high-frequency observation data of CO2 flux, CH4 flux, latent heat flux, and sensible heat flux, with a data frequency of 10Hz or 20Hz. The meteorological data acquisition submodule is used to connect to meteorological stations and acquire meteorological satellite data, as well as meteorological element data. The UAV inspection data acquisition submodule is used to receive hyperspectral images, thermal infrared images, and LiDAR point cloud data collected by the UAV, and to acquire high spatial resolution vegetation structure and surface deformation information. The laboratory analysis data acquisition submodule is used to receive laboratory analysis data on soil organic carbon content, vegetation biomass, dissolved organic carbon in water, and soil microbial community structure.

[0009] Furthermore, the plateau environment adaptive coding module includes an altitude gradient location coding unit, a temperature adaptive coding unit, an oxygen content coding unit, and a freeze-thaw cycle embedding unit, which are used to convert altitude, temperature, oxygen content, and freeze-thaw state into corresponding feature codes to adapt to the special ecological environment of the plateau.

[0010] Furthermore, the freeze-thaw sensing attention mechanism module includes a seasonal location encoding submodule, a freeze-thaw state encoding submodule, a temperature response function submodule, and a multi-scale time attention submodule, which are used to capture the dynamic impact of the freeze-thaw cycle on carbon flux at multiple time scales such as day, month, and season.

[0011] Furthermore, the cross-modal alignment of the multimodal fusion large model module employs a contrastive learning method, including: Construct multimodal positive sample pairs by combining data from different modalities at the same time point into positive sample pairs; Construct multimodal negative sample pairs by using data from different time points or different wetlands to create negative sample pairs; The InfoNCE loss function is used to train a cross-modal alignment network, which brings the same semantic features from different modalities closer together in the feature space. By employing a modality-specific projection network, features from different modalities are projected onto the same dimensional space, achieving cross-modal feature alignment.

[0012] Furthermore, the system also includes a few-shot learning mechanism module and a physical constraint layer module; The few-shot learning mechanism module includes: The meta-learning framework is used to train a model to learn how to learn through a set of tasks, enabling the model to adapt quickly when faced with new plateau wetland scenarios. The prototype network submodule is used to extract prototype features for each category in the support set and perform classification or regression prediction in the query set by nearest neighbor matching. The transfer learning submodule is used to transfer models pre-trained on other wetland data to plateau wetland scenarios, and to fine-tune them using a small amount of plateau wetland data. The data augmentation submodule is used to augment the plateau wetland training data through time series interpolation, image transformation, and feature noise injection. The physical constraint layer module includes: The temperature-controlled breathing constraint submodule is used to determine the temperature based on Q. 10 Temperature coefficient constraints on the impact of temperature on ecosystem respiration ensure the physical justification for increased CO2 emissions when temperatures rise; The hydrological methane constraint submodule is used to constrain CH4 emissions based on the flooding status of soil or water bodies, ensuring the physical laws governing the increase in CH4 emissions in flooded environments. The net ecosystem exchange constraint submodule is used to constrain the net ecosystem exchange amount according to the phenological period of vegetation to ensure the ecological law that the growing season is a carbon sink and the non-growing season may be a carbon source. The mass conservation constraint submodule is used to ensure mass conservation during the carbon cycle process, and the predicted total carbon change is equal to the algebraic sum of the changes in each carbon pool.

[0013] Furthermore, the carbon sink prediction module supports multiple prediction time scales, including: Real-time monitoring and prediction; estimating carbon flux at the current moment based on current multimodal data. Short-term forecasts predict changes in CO2 and CH4 fluxes over the next 1-7 days, based on meteorological forecast data. Medium-term forecast: Predicts the trend of carbon sink changes over the next 1-3 months, based on seasonal patterns and historical data for the same period. Long-term scenario analysis predicts carbon sink response over the next 10-50 years under different climate change scenarios, based on global climate model output data. The optimization decision module includes: The water level regulation strategy generation submodule is used to generate the optimal water level regulation scheme based on the carbon sink optimization target, including water level elevation, regulation timing and duration; The vegetation configuration optimization submodule is used to predict the carbon sequestration effect of different vegetation combinations and recommend the optimal vegetation configuration scheme, including plant species selection, planting density and spatial distribution. The artificial intervention assessment submodule is used to evaluate the carbon sequestration benefits of water replenishment, vegetation restoration, and habitat creation measures, and to quantify the input-output ratio of these measures. The multi-objective optimization submodule is used to balance multiple objectives such as carbon sink enhancement, biodiversity conservation, water resource conservation, and economic benefits, and adopts the Pareto frontier to find the optimal decision-making solution.

[0014] Furthermore, the system also includes an anomaly detection and early warning module, comprising: The flux anomaly detection submodule is used to identify abnormally high or low CO2 / CH4 emission events, and detects anomalies based on statistical methods and machine learning methods. The ecosystem health early warning submodule is used to assess the health status of wetland ecosystems through vegetation cover, water level, and biomass, and to issue early warnings when degradation trends appear. The freeze-thaw disaster early warning submodule is used to predict the risk of freeze-thaw erosion based on temperature prediction and soil condition, and issue early warnings during the high-incidence period of freeze-thaw disasters. The extreme weather early warning submodule is used to identify extreme weather events based on meteorological forecasts and to provide early warnings of the impact of extreme weather on carbon sinks.

[0015] Furthermore, the system also includes a knowledge base and decision support module, comprising: The expert knowledge fusion submodule is used to structure the experiential knowledge of plateau wetland ecology experts into rules and cases, and to encode expert knowledge into reusable knowledge representations. The scientific literature knowledge extraction submodule is used to extract knowledge related to carbon sequestration in plateau wetlands from massive scientific literature using natural language processing technology. The case retrieval submodule is used to retrieve treatment cases of similar wetlands from the case library based on the characteristics of the current wetland, providing a reference for decision-making; The interactive question-and-answer submodule supports users in querying questions related to carbon sequestration in plateau wetlands using natural language, and the system automatically generates interpretable answers based on data and models; The decision explanation submodule is used to explain the results of optimization decisions, show the basis for the decision, influencing factors and expected effects, and improve the credibility and acceptability of the decision.

[0016] Beneficial effects Compared with the prior art, the present invention has the following advantages: (1) This invention realizes deep cross-modal fusion of multi-source heterogeneous data, which comprehensively improves the expressive ability of carbon sink-related features. This invention acquires multi-source heterogeneous data such as remote sensing images, ground flux time series data, meteorological data, UAV inspection data, and laboratory analysis data through a multi-modal data perception module. It is equipped with image encoder, sequence encoder, geo-encoder and text encoder to extract the core features of different modalities. Then, it achieves unified spatial alignment of multi-modal features through cross-modal comparative learning. Finally, it completes deep nonlinear fusion of features with the help of the Transformer fusion layer, which fully explores the complementary information and intrinsic correlation between different modal data, comprehensively describes the multi-dimensional driving mechanism of carbon sink in plateau wetlands, and lays a solid data and feature foundation for high-precision carbon sink prediction.

[0017] (2) This invention constructs an adaptive coding and freeze-thaw sensing attention mechanism for plateau environment, realizing accurate adaptation to the plateau-specific environment and refined modeling of the freeze-thaw process. This invention designs a dedicated adaptive coding unit for plateau wetland-specific environmental factors such as altitude, temperature, oxygen content, and freeze-thaw cycle, encoding the plateau-specific environmental factors into key features that the model can recognize, enabling the model to autonomously learn the constraints of low oxygen and high altitude environment on plant photosynthesis and soil microbial metabolism; at the same time, it innovatively designs a freeze-thaw sensing attention mechanism integrated into the Transformer architecture, accurately capturing the dynamic regulation of carbon flux by daily, monthly, and seasonal multi-scale freeze-thaw cycles through freeze-thaw state coding, temperature response function, and multi-scale time attention branch, effectively solving the problems of poor adaptability of existing models to the complex plateau environment and low prediction accuracy of carbon flux fluctuations during freeze-thaw transition, and significantly improving the prediction performance and environmental adaptability of the model in plateau wetland scenarios.

[0018] (3) This invention designs a full-process few-shot learning mechanism, breaking through the industry bottleneck of scarce plateau wetland data. This invention constructs a full-process few-shot learning system including a meta-learning framework, a prototype network, transfer learning, and data augmentation. Through MAML meta-learning, the model gains the ability to quickly adapt to new scenarios. With the help of the prototype network, accurate classification and regression under few samples are achieved. Cross-scenario transfer learning is used to reduce the need for labeled data of the target wetland. Combined with multiple types of data augmentation methods to expand the training samples, the overfitting problem of the model under few-shot scenarios is effectively alleviated. The model can quickly adapt to different regions and different types of plateau wetlands with a small amount of labeled data, which greatly improves the generalization ability and engineering feasibility of the model. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a system architecture diagram provided in an embodiment of the present invention; Figure 2 This is a structural diagram of the multimodal data sensing module provided in an embodiment of the present invention; Figure 3 This is a structural diagram of the plateau environment adaptive coding module provided in an embodiment of the present invention; Figure 4 This is a structural diagram of the multimodal fusion large model module provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the freeze-thaw sensing attention mechanism provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of multi-timescale prediction of the carbon sink prediction module provided in an embodiment of the present invention; Figure 7 This is a flowchart of the method for constructing a multimodal large model of carbon sinks in plateau wetlands provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of cross-modal alignment training provided in an embodiment of the present invention; Figure 9 This is a diagram of the physical constraint layer structure provided in an embodiment of the present invention; Figure 10 This is a flowchart of the optimization decision-making module provided in an embodiment of the present invention; Figure 11 This is a cloud-edge collaborative deployment architecture diagram provided in an embodiment of the present invention; Figure 12 This is a system full-process data flow diagram provided in the embodiments of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0022] The objective of this invention is achieved through the following technical solution: This embodiment provides an intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model. The system adopts a cloud-edge collaborative deployment architecture, wherein: The system is deployed in the cloud on a server cluster equipped with eight NVIDIA A100 80GB GPUs, running Ubuntu 22.04, and using PyTorch 2.0+ as the deep learning framework for model pre-training, fine-tuning, updating, and long-term scenario analysis. The edge computing device is equipped with NVIDIA Jetson Orin NX and deployed at the on-site monitoring station in the plateau wetland. It is used for real-time data acquisition, lightweight model inference, local early warning and data transmission. The device can operate at temperatures from -40℃ to 70℃, which is suitable for the extreme low temperature environment of the plateau. The on-site data acquisition terminals include an eddy covariance flux tower, an automatic weather station, a drone inspection platform, a satellite remote sensing data receiving terminal, and soil / water sampling laboratory analysis equipment. All terminals interact with the edge and cloud via 4G / 5G / BeiDou satellite communication modules.

[0023] The overall system architecture of this embodiment is as follows: Figure 1 As shown, it includes a multimodal data perception module, a plateau environment adaptive coding module, a multimodal fusion large model module, a freeze-thaw perception attention mechanism, a carbon sink prediction module, and an optimization decision-making module. It is also equipped with a few-shot learning mechanism, a physical constraint layer, an anomaly detection and early warning module, and a knowledge base and decision support module. The specific implementation methods of each module are as follows: 1. Multimodal data sensing module The specific structure of this module is shown in the attached instruction manual. Figure 2 As shown, this module is used to acquire multi-source heterogeneous data of plateau wetlands. The multi-source heterogeneous data includes remote sensing image data, ground monitoring flux data, meteorological data, UAV inspection data, and laboratory analysis data. This module includes a multi-source data acquisition unit and a data preprocessing unit.

[0024] 2. Plateau Environment Adaptive Coding Module The specific structure of this module is shown in the attached instruction manual. Figure 3 As shown, the system encodes plateau-specific environmental factors such as altitude, temperature, oxygen content, and freeze-thaw cycle into model input features. It includes an altitude gradient location encoding unit, a temperature adaptive encoding unit, an oxygen content encoding unit, and a freeze-thaw cycle embedding unit. The specific implementation of each unit is as follows: In practice, the elevation gradient location coding unit first acquires the digital elevation model (DEM) data of the target wetland, matching the spatial resolution with the remote sensing image. The elevation is divided into several gradient intervals at 60m intervals, and an independent learnable embedding vector is assigned to each gradient interval. The dimension of the embedding vector is consistent with the feature dimension of the model backbone network (default 256 dimensions). The elevation gradient topology is constructed through discretization coding, enabling the model to capture the nonlinear relationship between elevation and vegetation distribution and microbial activity. At the same time, the elevation embedding vector is added element-wise with the spatial location code of the remote sensing image and the temporal code of the time series, realizing the fusion of elevation features and basic features.

[0025] The temperature adaptive coding unit first constructs a historical temperature distribution statistical database based on more than 10 years of historical daily temperature data of the target wetland, and calculates the multi-year monthly average temperature, temperature standard deviation, and extreme temperature threshold. The real-time collected temperature data is Z-score standardized relative to the historical temperature distribution of the same period to generate temperature standardized features. At the same time, the diurnal variation, seasonal variation, and interannual variation features of temperature are extracted, and the dynamic trend features of temperature are extracted through a 1D convolutional layer. The vegetation phenological features (greening period, peak growth period, withering period, and dormancy period) extracted from NDVI time series data are fused to generate the final temperature perception feature vector.

[0026] The oxygen content coding unit calculates the relative oxygen content based on the atmospheric pressure-oxygen content relationship model. The calculation formula is as follows: O2 relative content = exp(-h / H) In the formula, h is the altitude of the target point (unit: m), and H is the elevation (value is 8500m). The calculated relative oxygen content is calibrated with the measured atmospheric oxygen content data. The oxygen content value is encoded into a feature vector that matches the dimension of the backbone network through a fully connected layer. At the same time, the oxygen content feature is concatenated with the temperature feature and altitude feature. Key features are selected through an attention gating mechanism, so that the model can learn the constraint effect of low oxygen environment on plant photosynthesis and soil microbial metabolism.

[0027] The freeze-thaw cycle embedding unit, based on soil temperature and moisture content data at a depth of 5 cm, classifies soil freeze-thaw states into three categories: frozen state (soil temperature consistently below 0℃, soil liquid water content below 5%), thawed state (soil temperature consistently above 0℃, soil liquid water content above 20%), and transitional state (soil temperature fluctuates around 0℃, with alternating freeze-thaw cycles). An independent, learnable state embedding vector is assigned to each freeze-thaw state. Simultaneously, the state embedding vector is fused with seasonal location coding and diurnal time coding to generate freeze-thaw cycle features, enabling the model to perceive the dynamic regulatory effect of the freeze-thaw cycle on carbon flux.

[0028] 3. Multimodal fusion large model module The specific structure of this module is shown in the attached instruction manual. Figure 4 As shown, it includes an image encoder, sequence encoder, geoencoder, and text encoder. Deep fusion of multimodal features is achieved through cross-modal alignment and a Transformer fusion layer. The specific implementation of each part is as follows: (1) Image encoder This encoder is used to extract visual features from remote sensing images and UAV hyperspectral / thermal infrared images. It adopts the VisionTransformer (ViT-B / 16) architecture. The input is preprocessed remote sensing images and UAV multispectral images. First, it is pre-trained on the ImageNet large-scale image dataset, and then fine-tuned using a wetland remote sensing image dataset of the target wetland and surrounding area. The fine-tuning process adopts the masked autoencoder (MAE) self-supervised learning method, randomly masking 75% of the image patches, training the model to reconstruct the masked areas, and finally outputting a 768-dimensional image visual feature vector.

[0029] (2) Sequence encoder This encoder is used to extract time-series features from flux data and meteorological data. It adopts a TemporalTransformer architecture with 6 encoder layers and 16 attention heads. The input consists of preprocessed flux time-series data, meteorological time-series data, and soil environmental time-series data, and the input sequence length is configurable. The encoder contains 6 Transformer encoder layers, each with a 16-head self-attention mechanism. In the pre-training stage, a masked time-series modeling method is used, randomly masking 30% of the time steps. The model is trained to predict the masked values. The pre-training dataset consists of continuous observation datasets from more than 20 wetland ecological stations across the country for more than 10 years. After pre-training, the model is fine-tuned using target wetland time-series data, and a 512-dimensional time-series feature vector is output.

[0030] (3) Geographic encoder This encoder is used to extract the spatial topological features of wetlands. It adopts the Graph Attention Network (GAT) architecture in Graph Neural Networks (GNNs). First, a graph structure is constructed based on the hydrological connectivity, spatial adjacency, and vegetation type distribution of the target wetland. The wetland is divided into several 100m×100m grid cells, with each grid cell serving as a node in the graph. The edges between nodes are constructed based on spatial adjacency and hydrological flow direction. In the pre-training stage, a graph contrastive learning method is used. A contrastive view is constructed through node feature masks and edge perturbations to train the model to learn the spatial topological features of the nodes. Finally, a 256-dimensional spatial topological feature vector is output.

[0031] (4) Text encoder This encoder is used to extract semantic features from scientific research literature and expert knowledge. It adopts the BERT-base-Chinese pre-trained language model as its basic architecture. The input includes scientific research literature, industry standards, expert experience knowledge, and monitoring report text data in the field of wetland ecology. First, it is pre-trained on a general Chinese corpus, and then it is domain-adaptive pre-trained through a constructed plateau wetland ecology professional corpus. The pre-training tasks include masked language model (MLM) and sentence-level contrastive learning tasks. Finally, it outputs a 256-dimensional semantic feature vector.

[0032] (5) Cross-modal alignment and Transformer fusion layer This section achieves cross-modal alignment through contrastive learning, and then achieves deep feature fusion through a Transformer fusion layer. The specific implementation is as follows: construct multimodal positive sample pairs, construct positive sample pairs from different modal data at the same time point and the same spatial location, and construct negative sample pairs from data at different time points, different spatial locations, or different wetlands. Each positive sample pair is matched with 128 negative sample pairs; design a modality-specific projection head network to project the features output by each encoder to a unified 128-dimensional embedding space through the modality-specific projection head; and train the cross-modal alignment network using the InfoNCE loss function.

[0033] The Transformer fusion layer employs a 6-layer Transformer encoder architecture, with each layer featuring a 16-head multi-attention mechanism and a feedforward network dimension of 1024. Aligned image features, sequence features, geographic features, and text features are concatenated along the channel dimension to form a fused feature matrix, which is then input into the Transformer fusion layer. The multi-head self-attention mechanism captures the correlation between features from different modalities, uncovers cross-modal feature interaction information, and achieves nonlinear transformation and deep fusion of features through the feedforward network, ultimately outputting a global fused feature vector.

[0034] 4. Freeze-thaw perception attention mechanism module The specific structure of this module is shown in the attached instruction manual. Figure 5 As shown, this method is used to model the freeze-thaw cycle of plateau wetlands and predict the impact of freeze-thaw on carbon flux. It is integrated into the self-attention module of the Transformer fusion layer and specifically includes a seasonal location encoding submodule, a freeze-thaw state encoding submodule, a temperature response function submodule, and a multi-scale time attention submodule.

[0035] In practice, the seasonal location coding submodule generates a fixed-dimensional location coding vector for each of the 365 days of the year.

[0036] The freeze-thaw state coding submodule, based on the three types of state embedding vectors output by the freeze-thaw cycle embedding unit, uses a gated fusion network to fuse the state embedding vectors with soil temperature and humidity features and water level features to generate freeze-thaw state coding features. At the same time, state attention weights are set to dynamically adjust the weight ratio of different state codes according to real-time soil temperature data, so that the model can automatically increase the weight of transition state codes during the freeze-thaw transition period and enhance its ability to capture drastic fluctuations in carbon flux.

[0037] The temperature response function submodule is based on Q. 10 The temperature coefficient model calculates the nonlinear response of temperature to ecosystem respiration. The calculation formula is as follows: In the formula, R is the ecosystem respiration rate, R0 is the baseline respiration rate at the reference temperature (calibrated using measured data from the target wetland), and Q... 10 Here, T represents the temperature coefficient (optimized to 2.2 for the plateau wetland scenario), where T is the soil temperature at a depth of 5 cm. ref The reference temperature is 15℃. The calculated physical constraint features of the breathing rate are input into the attention mechanism and fused with the breathing features learned by the model. At the same time, the output of the temperature response function is used as the bias term of the attention weight to adjust the model's attention to temperature-sensitive features.

[0038] The multi-scale temporal attention submodule constructs temporal attention branches at three time scales: daily, monthly, and seasonal, to capture the multi-scale dynamic impact of freeze-thaw cycles on carbon flux. Diurnal branch: Input hourly flux and meteorological data for 24 hours, set the attention window to 24 steps, and capture the impact of intraday freeze-thaw cycles on carbon flux; Monthly scale branch: Input 30 days of daily data, set the attention window to 30 steps, and capture the impact of monthly freeze-thaw state changes on carbon flux; Seasonal Scale Branch: Input 12 months of monthly data, set the attention window to 12 steps, and capture the impact of the annual freeze-thaw cycle on total carbon. The features output by each branch are weighted and fused through a multi-scale fusion gating network. The weights are dynamically learned by the model based on the input data. The final output is an attention feature that incorporates the effects of multi-scale freeze-thaw cycles, which is then input into the carbon sink prediction module.

[0039] 5. Carbon Sequestration Forecasting Module The specific structure of this module is as follows: Figure 6 As shown, the carbon sink prediction module is used to output prediction results of CO2 flux, CH4 flux and net ecosystem exchange based on fusion characteristics, and supports multiple prediction time scales such as real-time monitoring prediction, short-term prediction, medium-term prediction and long-term scenario analysis.

[0040] In practice, the carbon sink prediction module adopts a multi-task learning framework, taking the global fusion features output by the multimodal fusion large model module and the freeze-thaw attention features output by the freeze-thaw sensing attention mechanism as inputs. Three parallel prediction heads are set up, corresponding to CO2 flux prediction, CH4 flux prediction and net ecosystem exchange (NEE) prediction respectively. Each prediction head adopts a 2-layer fully connected network, with ReLU as the activation function and a linear activation function as the output layer.

[0041] This module supports four prediction time scales: Real-time monitoring and prediction; estimating carbon flux at the current moment based on current multimodal data. Short-term forecasts predict changes in CO2 and CH4 fluxes over the next 1-7 days, based on meteorological forecast data. Medium-term forecast: Predicts the trend of carbon sink changes over the next 1-3 months, based on seasonal patterns and historical data for the same period. Long-term scenario analysis predicts carbon sink responses over the next 10-50 years under different climate change scenarios (RCP 2.6, RCP 4.5, RCP 8.5), based on global climate model output data.

[0042] 6. Optimize the decision-making module This module generates optimized management recommendations for plateau wetlands based on carbon sink prediction results. The workflow is as follows: Figure 10 As shown, it includes a water level regulation strategy generation submodule, a vegetation configuration optimization submodule, a human intervention evaluation submodule, and a multi-objective optimization submodule. The water level regulation strategy generation submodule is used to generate the optimal water level regulation scheme based on the carbon sink optimization target, including water level elevation, regulation timing and duration; The vegetation configuration optimization submodule is used to predict the carbon sequestration effect of different vegetation combinations and recommend the optimal vegetation configuration scheme, including plant species selection, planting density and spatial distribution. The artificial intervention assessment submodule is used to assess the carbon sequestration benefits of measures such as water replenishment, vegetation restoration, and habitat creation, and to quantify the input-output ratio of these measures. The multi-objective optimization submodule is used to balance multiple objectives such as carbon sink enhancement, biodiversity conservation, water resource conservation, and economic benefits, and adopts the Pareto frontier to find the optimal decision-making solution.

[0043] 7. Specific Implementation Methods of Supporting Modules and Mechanisms (1) Few-shot learning mechanism This mechanism addresses the scarcity of data for plateau wetlands, enabling models to quickly adapt to target wetland scenarios with limited labeled data. It specifically includes a meta-learning framework, a prototype network submodule, a transfer learning submodule, and a data augmentation submodule, implemented as follows: ① Meta-learning framework: The Model-independent Meta-learning (MAML) algorithm is adopted to construct a meta-task set. Each meta-task contains a support set (5-10 labeled samples) and a query set (20-30 samples to be predicted). The meta-task set covers different regions and types of plateau wetland data of the Qinghai-Tibet Plateau. Through multiple rounds of training on the meta-task set, the model learns general feature extraction and fast adaptation capabilities. The training objective is to minimize the prediction loss of the model after 1-5 steps of gradient update on a new task. ② Prototype Network Submodule: For tasks such as carbon sink status classification and wetland degradation level classification, the prototype features (mean of all sample features within the category) of each category are calculated in the support set. In the query set, the Euclidean distance from the query sample to the prototype of each category is calculated to complete the classification or regression prediction. Adaptation can be completed without a large amount of labeled data. ③ Transfer learning submodule: The model weights pre-trained on multiple plain wetlands and other plateau wetlands across the country are used as the initial weights of the target wetland model. The encoder weights of the model are frozen. Only a small amount of labeled data from the target wetland is used to fine-tune the top-level parameters of the model's prediction head and Transformer fusion layer, so as to quickly adapt to the unique attributes of the target wetland. ④ Data Augmentation Submodule: For time series data, linear interpolation, time warp, and Gaussian noise injection are used for data augmentation; for remote sensing / UAV imagery, random rotation, flipping, color jitter, and masking are used for data augmentation; for feature data, feature mixing (MixUp) is used to generate augmented samples, expand the training dataset, and alleviate the model overfitting problem in small sample scenarios.

[0044] (2) Physical constraint layer This constraint layer is integrated into the model's output to ensure that the model's predictions conform to the physical laws and ecological principles of the carbon cycle. The specific structure is shown in the attached instruction manual. Figure 9 As shown, it includes submodules for temperature and respiration constraints, hydrological methane constraints, net ecosystem exchange constraints, and mass conservation constraints. The specific implementation is as follows: ① Temperature-based breathing constraint submodule: based on Q 10 A temperature coefficient model is used to calculate the theoretical ecosystem respiration rate and CO2 emission flux, and a temperature-respiration constraint loss function L is designed. temp The mean square error between the CO2 flux predicted by the calculation model and the theoretical value is added as a loss term to the total loss function; the model output is constrained to conform to the physical law that when the temperature rises, the respiration of the ecosystem increases and CO2 emissions increase. ② Hydrological Methane Constraint Submodule: Based on the response relationship between flooding and CH4 emissions, a hydrological-methane theoretical model is constructed. When the soil / water body is in a flooded anaerobic state, the theoretical CH4 emission flux increases significantly. A hydrological methane constraint loss function L is designed. hydro The deviation between the CH4 flux predicted by the calculation model and the theoretical value is used to constrain the model output to conform to the biogeochemical laws of increased CH4 emissions in flooded environments. ③ Net Ecosystem Exchange Constraint Submodule: Based on vegetation phenology, constraints are imposed on NEE. During the growing season (June-September), vegetation photosynthesis is vigorous, and NEE should be negative (carbon sink). During the non-growing season, NEE can be positive (carbon source). Design the phenological constraint loss function L. pheno When the growing season model predicts a positive NEE value, a penalty term is applied to ensure that the prediction results conform to the phenological laws of the ecosystem. ④ Mass Conservation Constraint Submodule: Based on the law of conservation of carbon cycle mass, the change in total carbon in an ecosystem is equal to the algebraic sum of the changes in the vegetation carbon pool, soil organic carbon pool, and water carbon pool; design the mass conservation constraint loss function L. mass The deviation between the total carbon change predicted by the calculation model and the sum of the changes in each carbon pool is calculated to ensure the quality conservation of the carbon cycle process. ⑤ Total Loss Function: The total loss function is obtained by weighted summing of the physical constraint loss term and the master prediction loss term. L total =α×L pred +β×L temp +γ×L hydro +δ×L pheno +ε×L mass Here, α, β, γ, δ, and ε are the weight coefficients of each constraint loss. The weights are adjusted according to the actual scenario to balance the relative importance of data-driven prediction and physical law constraints. Model training is completed through backpropagation.

[0045] (3) Anomaly detection and early warning module This module is used to identify and issue early warnings for abnormal events related to carbon sequestration in plateau wetlands, specifically including: ① Flux Anomaly Detection Submodule: Based on the 3σ statistical criterion and the isolated forest machine learning algorithm, it identifies abnormally high emissions or abnormally low absorption events in CO2 / CH4 flux. When the flux value exceeds three times the standard deviation of the historical period, it is judged as an anomaly, triggers an anomaly alarm, and simultaneously pushes the anomaly location, anomaly value and possible cause analysis. ② Ecosystem Health Early Warning Submodule: Based on indicators such as vegetation coverage, water level, soil organic carbon content, and biomass, a wetland ecosystem health evaluation system is constructed, which is divided into five levels: healthy, sub-healthy, slightly degraded, moderately degraded, and severely degraded. When a wetland shows a degradation trend for three consecutive months, a degradation early warning is issued, and the analysis of the degradation area and degradation driving factors is pushed out. ③ Freeze-thaw disaster early warning submodule: Based on temperature prediction and soil freeze-thaw status prediction, predict the risk of freeze-thaw erosion and permafrost thawing. During the high-incidence period of freeze-thaw disasters in spring freeze-thaw transition period and autumn freeze transition period, issue early warnings 7 days in advance and push high-risk areas and protection suggestions. ④ Extreme Weather Early Warning Submodule: Connects with meteorological forecast data, identifies extreme weather events such as rainstorms, droughts, high temperatures, and cold waves, simulates the impact of extreme weather on wetland carbon sinks based on carbon sink prediction models, issues early warnings, and pushes out suggestions for countermeasures.

[0046] (4) Knowledge base and decision support module This module is designed to integrate expert knowledge with research findings, enhancing the interpretability of decisions. Specifically, it includes: ①Expert Knowledge Integration Submodule: The expert experience in the fields of plateau wetland ecology, hydrology and water resources, and ecological restoration is structured into production rules in the form of "condition-action-result" to build an expert rule base; at the same time, typical cases of plateau wetland ecological restoration are collected to build a case base, and expert knowledge and cases are encoded into reusable knowledge representations and integrated into a text encoder. ② Scientific Literature Knowledge Extraction Submodule: Based on named entity recognition and relation extraction technologies, it extracts knowledge such as influencing factors, carbon sink mechanisms, management measures, and key parameters of plateau wetland carbon sink from massive scientific literature, and constructs a structured domain knowledge graph to provide knowledge support for model prediction and decision-making; ③ Interactive Question and Answer Submodule: Based on a domain-adaptive pre-trained large language model, it supports users to query questions related to carbon sequestration in plateau wetlands using natural language. The system combines real-time monitoring data, model prediction results, and knowledge base content to automatically generate interpretable answers, realizing intelligent question and answer interaction for wetland management. ④ Decision Interpretation Submodule: For the generated optimized decision scheme, the SHAP interpretability analysis method is used to calculate the contribution of each influencing factor to the decision result, and output the core basis of the decision, expected carbon sink benefits, influencing factors and risk warnings to improve the credibility and acceptability of the decision.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model, characterized in that, include: The multimodal data sensing module is used to acquire remote sensing image data, ground monitoring flux data, meteorological data, UAV inspection data, and laboratory analysis data of plateau wetlands. The plateau environment adaptive coding module is used to encode plateau-specific environmental factors such as altitude, temperature, oxygen content, and freeze-thaw cycle into model input features; The multimodal fusion large model module includes an image encoder, a sequence encoder, a geoencoder, and a text encoder. The image encoder is used to extract visual features from remote sensing images, the sequence encoder is used to extract time-series features from flux data and meteorological data, the geoencoder is used to extract spatial topological features of wetlands, and the text encoder is used to extract semantic features from scientific literature and expert knowledge. The multimodal fusion large model module achieves deep fusion of multimodal features through cross-modal alignment and a Transformer fusion layer. The freeze-thaw sensing attention mechanism module is used to model the freeze-thaw cycle of plateau wetlands and predict the impact of freeze-thaw on carbon flux. The carbon sink prediction module is used to output prediction results of CO2 flux, CH4 flux and net ecosystem exchange based on the fused multimodal features; The optimization decision-making module is used to generate optimized management suggestions for improving carbon sequestration in plateau wetlands based on the prediction results.

2. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The multimodal data sensing module specifically includes: The remote sensing image acquisition submodule is used to receive satellite remote sensing data, including Sentinel-2 multispectral images, Landsat-8 images, and Gaofen series satellite images, and extract spectral indices. The ground monitoring flux data acquisition submodule is used to connect to the eddy covariance flux tower to acquire high-frequency observation data of CO2 flux, CH4 flux, latent heat flux, and sensible heat flux, with a data frequency of 10Hz or 20Hz. The meteorological data acquisition submodule is used to connect to meteorological stations and acquire meteorological satellite data, as well as meteorological element data. The UAV inspection data acquisition submodule is used to receive hyperspectral images, thermal infrared images, and LiDAR point cloud data collected by the UAV, and to acquire high spatial resolution vegetation structure and surface deformation information. The laboratory analysis data acquisition submodule is used to receive laboratory analysis data on soil organic carbon content, vegetation biomass, dissolved organic carbon in water, and soil microbial community structure.

3. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The plateau environment adaptive coding module includes an altitude gradient location coding unit, a temperature adaptive coding unit, an oxygen content coding unit, and a freeze-thaw cycle embedding unit, which are used to convert altitude, temperature, oxygen content, and freeze-thaw state into corresponding feature codes to adapt to the special ecological environment of the plateau.

4. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The freeze-thaw sensing attention mechanism module includes a seasonal location encoding submodule, a freeze-thaw state encoding submodule, a temperature response function submodule, and a multi-scale time attention submodule, which are used to capture the dynamic impact of the freeze-thaw cycle on carbon flux at multiple time scales such as daily, monthly, and seasonal.

5. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The cross-modal alignment of the multimodal fusion large model module adopts a contrastive learning method, including: Construct multimodal positive sample pairs by combining data from different modalities at the same time point into positive sample pairs; Construct multimodal negative sample pairs by using data from different time points or different wetlands to create negative sample pairs; The InfoNCE loss function is used to train a cross-modal alignment network, which brings the same semantic features from different modalities closer together in the feature space. By employing a modality-specific projection network, features from different modalities are projected onto the same dimensional space, achieving cross-modal feature alignment.

6. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The system also includes a few-shot learning mechanism module and a physical constraint layer module; The few-shot learning mechanism module includes: The meta-learning framework is used to train a model to learn how to learn through a set of tasks, enabling the model to adapt quickly when faced with new plateau wetland scenarios. The prototype network submodule is used to extract prototype features for each category in the support set and perform classification or regression prediction in the query set by nearest neighbor matching. The transfer learning submodule is used to transfer models pre-trained on other wetland data to plateau wetland scenarios, and to fine-tune them using a small amount of plateau wetland data. The data augmentation submodule is used to augment the plateau wetland training data through time series interpolation, image transformation, and feature noise injection. The physical constraint layer module includes: The temperature-controlled breathing constraint submodule is used to determine the temperature based on Q. 10 Temperature coefficient constraints on the impact of temperature on ecosystem respiration ensure the physical justification for increased CO2 emissions when temperatures rise; The hydrological methane constraint submodule is used to constrain CH4 emissions based on the flooding status of soil or water bodies, ensuring the physical laws governing the increase in CH4 emissions in flooded environments. The net ecosystem exchange constraint submodule is used to constrain the net ecosystem exchange amount according to the phenological period of vegetation to ensure the ecological law that the growing season is a carbon sink and the non-growing season may be a carbon source. The mass conservation constraint submodule is used to ensure mass conservation during the carbon cycle process, and the predicted total carbon change is equal to the algebraic sum of the changes in each carbon pool.

7. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model as described in claim 1, characterized in that, The carbon sink prediction module supports multiple prediction time scales, including: Real-time monitoring and prediction; estimating carbon flux at the current moment based on current multimodal data. Short-term forecasts predict changes in CO2 and CH4 fluxes over the next 1-7 days, based on meteorological forecast data. Medium-term forecast: Predicts the trend of carbon sink changes over the next 1-3 months, based on seasonal patterns and historical data for the same period. Long-term scenario analysis predicts carbon sink response over the next 10-50 years under different climate change scenarios, based on global climate model output data. The optimization decision module includes: The water level regulation strategy generation submodule is used to generate the optimal water level regulation scheme based on the carbon sink optimization target, including water level elevation, regulation timing and duration; The vegetation configuration optimization submodule is used to predict the carbon sequestration effect of different vegetation combinations and recommend the optimal vegetation configuration scheme, including plant species selection, planting density and spatial distribution. The artificial intervention assessment submodule is used to evaluate the carbon sequestration benefits of water replenishment, vegetation restoration, and habitat creation measures, and to quantify the input-output ratio of these measures. The multi-objective optimization submodule is used to balance multiple objectives such as carbon sink enhancement, biodiversity conservation, water resource conservation, and economic benefits, and adopts the Pareto frontier to find the optimal decision-making solution.

8. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model according to claim 1, characterized in that, The system also includes an anomaly detection and early warning module, including: The flux anomaly detection submodule is used to identify abnormally high or low CO2 / CH4 emission events, and detects anomalies based on statistical methods and machine learning methods. The ecosystem health early warning submodule is used to assess the health status of wetland ecosystems through vegetation cover, water level, and biomass, and to issue early warnings when degradation trends appear. The freeze-thaw disaster early warning submodule is used to predict the risk of freeze-thaw erosion based on temperature prediction and soil condition, and issue early warnings during the high-incidence period of freeze-thaw disasters. The extreme weather early warning submodule is used to identify extreme weather events based on meteorological forecasts and to provide early warnings of the impact of extreme weather on carbon sinks.

9. The intelligent prediction and optimization system for carbon sequestration in plateau wetlands based on a multimodal large model according to claim 1, characterized in that, The system also includes a knowledge base and decision support module, including: The expert knowledge fusion submodule is used to structure the experiential knowledge of plateau wetland ecology experts into rules and cases, and to encode expert knowledge into reusable knowledge representations. The scientific literature knowledge extraction submodule is used to extract knowledge related to carbon sequestration in plateau wetlands from massive scientific literature using natural language processing technology. The case retrieval submodule is used to retrieve treatment cases of similar wetlands from the case library based on the characteristics of the current wetland, providing a reference for decision-making; The interactive question-and-answer submodule supports users in querying questions related to carbon sequestration in plateau wetlands using natural language, and the system automatically generates interpretable answers based on data and models; The decision explanation submodule is used to explain the results of optimization decisions, show the basis for the decision, influencing factors and expected effects, and improve the credibility and acceptability of the decision.

Images