Forest ecosystem monitoring system and method based on space-time big data

Abstract

The invention discloses a forest ecosystem monitoring system and method based on space-time big data, and relates to the technical field of data processing systems. The system comprises a multi-source heterogeneous data acquisition and preprocessing module, an ecological process mechanism embedded spatio-temporal feature extraction module, a self-adaptive dynamic state characterization and evolution module, an abnormal event early perception and causal traceability module and a multi-dimensional ecological function comprehensive evaluation and early warning module. The method comprises the following steps: collecting and processing multi-source heterogeneous spatio-temporal data to generate a multi-source spatio-temporal data set; extracting a multi-dimensional spatio-temporal feature vector; constructing a dynamic forest state characterization model, and generating a forest state index map; executing abnormal event early perception and causal traceability analysis; and evaluating the health state of the forest ecosystem and generating early warning information. According to the method, the characterization capability of the complex ecological process is improved, more accurate and coherent dynamic description of the ecological system state is realized, and a reliable basis is provided for ecological management and decision making.

Description

Technical Field

[0001] This invention relates to the field of data processing system technology, and in particular to a forest ecosystem monitoring system and method based on spatiotemporal big data. Background Technology

[0002] As the main body of Earth's terrestrial ecosystems, forest ecosystems undertake key ecological functions such as water conservation, biodiversity maintenance, carbon sequestration regulation, and climate stability. Their health directly affects global ecological security and sustainable development. With the development of remote sensing technology, IoT sensing networks, and geographic information systems, the dynamic monitoring of forest ecosystems has gradually evolved from static sampling to continuous and refined monitoring. Current forest monitoring systems generally rely on the fusion analysis of multi-source spatiotemporal data, including satellite remote sensing imagery, ground sensor networks, meteorological observation stations, and UAV aerial photography data. These data exhibit periodic or near-real-time characteristics in the temporal dimension and cover multi-scale features ranging from point-based deployments to area coverage in the spatial dimension. However, existing monitoring methods still have significant limitations in data integration, status identification, and trend early warning, making it difficult to meet the needs for high-precision, high-timeliness, and high-robust comprehensive assessment of complex forest ecosystems.

[0003] Among these, forest ecosystem monitoring based on spatiotemporal big data focuses on constructing dynamic characterization models using multi-source heterogeneous spatiotemporal data to achieve a quantitative description of forest structure, function, and disturbance response. The core of this technical direction lies in transforming discrete observation data into continuous, interpretable sequences of ecological indicators through spatiotemporal alignment, feature extraction, and state mapping, thereby supporting key application scenarios such as degradation identification, fire risk assessment, and pest and disease transmission simulation. These methods typically rely on spatiotemporal interpolation, change detection algorithms, or shallow machine learning models, attempting to extrapolate the state of large-scale forest areas with limited computational resources. However, their modeling capabilities are constrained by inherent challenges such as data spatiotemporal resolution mismatch, observational noise interference, and nonlinear coupling of ecological processes, making it difficult for monitoring results to meet operational requirements in terms of accuracy, consistency, and generalization ability.

[0004] Existing technologies for forest ecosystem monitoring generally suffer from the following integrated deficiencies: First, multi-source spatiotemporal data lack a unified spatiotemporal benchmark framework, making it difficult to effectively coordinate the time lag of satellite remote sensing data with the local representativeness of ground sensors, resulting in discontinuities in the spatiotemporal continuity of monitoring results. Second, existing analytical models mostly employ static or linear assumptions, failing to accurately characterize non-stationary and nonlinear dynamic mechanisms such as vegetation growth, disturbance response, and recovery succession in forest ecological processes, leading to significant biases in state identification. Third, monitoring systems generally lack the ability to detect abnormal events early and trace their causal origins, making it difficult to provide effective early warnings within critical windows when facing sudden disturbances such as wildfires, pests, or illegal logging. Finally, existing methods rely too heavily on empirical thresholds or single ecological parameters in their indicator system design, failing to construct a comprehensive evaluation paradigm oriented towards multidimensional ecological functions, making it difficult for monitoring conclusions to support scientific decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a forest ecosystem monitoring system and method based on spatiotemporal big data to solve the problems of spatiotemporal continuity discontinuity of data, state identification bias and early warning lag in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a forest ecosystem monitoring system based on spatiotemporal big data includes: The multi-source heterogeneous data acquisition and preprocessing module collects raw spatiotemporal data from satellite remote sensing platforms, ground IoT sensor networks, meteorological observation stations, and UAV aerial survey platforms, and performs spatiotemporal benchmark unification, missing value imputation, and noise filtering operations on the raw spatiotemporal data to generate a standardized multi-source spatiotemporal dataset. The spatiotemporal feature extraction module embedded in the ecological process mechanism receives the standardized multi-source spatiotemporal dataset and extracts multi-dimensional spatiotemporal feature vectors representing forest structure, function and disturbance response from the dataset by coupling the physical ecological process model with a deep spatiotemporal neural network. The adaptive dynamic state representation and evolution module receives the multidimensional spatiotemporal feature vector and constructs a dynamic forest state representation model based on attention mechanism and graph neural network. The dynamic forest state representation model can adaptively learn the non-stationary evolution law of forest ecosystem in the time dimension and the heterogeneous correlation in the spatial dimension, and generate a forest state index map on a continuous spatiotemporal grid. The early perception and causal tracing module for abnormal events monitors the dynamic changes of the forest state index map in real time. By constructing an anomaly detection framework based on variational autoencoders and Granger causality tests, it identifies abnormal perturbation events that exceed the normal evolution threshold and performs causal probability inference on the potential driving factors of the abnormal events based on Bayesian networks. The multidimensional ecological function comprehensive assessment and early warning module integrates forest state index map and causal tracing results of abnormal events. Based on a preset comprehensive assessment index system oriented towards productivity, stability and resilience, it calculates the comprehensive health score of the forest ecosystem and generates graded early warning information when the comprehensive health score is lower than a preset threshold or when a high-risk abnormal event is detected.

[0007] On the other hand, a forest ecosystem monitoring method based on spatiotemporal big data includes: Step S110: Collect and preprocess multi-source heterogeneous spatiotemporal data to generate a standardized multi-source spatiotemporal dataset; Step S120: Based on the ecological process mechanism and deep spatiotemporal neural network, extract multidimensional spatiotemporal feature vectors from the standardized multi-source spatiotemporal dataset; Step S130: Construct a dynamic forest state representation model and use the multidimensional spatiotemporal feature vector to generate a forest state index map on a continuous spatiotemporal grid. Step S140: Based on the forest state index map, perform early perception and causal source analysis of abnormal events; Step S150: Based on the multidimensional ecological function comprehensive assessment index system, assess the health status of the forest ecosystem and generate early warning information.

[0008] Preferably, in the multi-source heterogeneous data acquisition and preprocessing module, the spatiotemporal reference unification operation specifically involves: uniformly reprojecting satellite remote sensing images and UAV aerial photography data to a preset universal geographic coordinate system, and aligning the timestamps of all data to a unified UTC time reference; the missing value imputation adopts a spatiotemporal kriging interpolation algorithm, which simultaneously considers spatial autocorrelation and the periodicity of the time series, and its interpolation weight function is defined as: in, Represents a known point For missing points Contribution weight, and Let Euclidean distance and time interval represent the spatial Euclidean distance and time interval between the two points, respectively. and These are preset spatial and temporal scale parameters, respectively; the noise filtering adopts a joint denoising method based on wavelet transform and 3σ criterion to filter out high-frequency noise and outlier observations.

[0009] Furthermore, the spatiotemporal feature extraction module embedded with the ecological process mechanism includes a dual-channel feature extraction network: the first channel is a physical ecological process simulation channel, integrating a photosynthesis model, a water stress model, and a nutrient cycling model, taking meteorological and soil data as input, and simulating and calculating the vegetation photosynthetically active radiation absorption ratio, water use efficiency, and nitrogen and phosphorus availability index at the grid scale; the second channel is a deep spatiotemporal neural network channel, adopting an architecture of cascaded 3D convolutional layers and long short-term memory network layers, directly learning spectral, texture, and phenological temporal variation features from multi-temporal remote sensing image sequences. The outputs of the dual channels are integrated through a feature fusion layer, which uses a gated attention mechanism to dynamically adjust the feature contribution weights from the physical model channel and the data-driven channel, ultimately outputting a multi-dimensional spatiotemporal feature vector that integrates prior ecological mechanisms and the inherent laws of the data.

[0010] Furthermore, in the adaptive dynamic state representation and evolution module, the dynamic forest state representation model represents forest spatial units and their relationships using a graph structure. Each spatial unit is a graph node, and its initial node features are the multidimensional spatiotemporal feature vectors. The edge weights between nodes are jointly determined by spatial adjacency and ecological similarity. Ecological similarity is obtained by calculating the cosine similarity of historical feature vectors between nodes. The core of the dynamic forest state representation model is a spatiotemporal graph attention network layer. When updating the node state at each time step, this layer not only aggregates the information of its spatial neighbor nodes but also aggregates its own state information at historical time steps. The node state update formula is: in, Represents a node In time The hidden state, Represents a node Spatial neighbor set, Spatial attention weights, For time-based self-attention weights, and For learnable parameter matrix, For bias terms, The activation function is denoted as . By stacking multiple spatiotemporal graph attention network layers, the dynamic forest state representation model can capture the complex nonlinear evolution pattern of forest state in the spatiotemporal dimension, and finally output the forest state index of each spatial unit at each time step, forming the forest state index map.

[0011] Preferably, the workflow of the early detection and causal tracing module for abnormal events includes two stages. The first stage is anomaly detection, which uses a variational autoencoder to reconstruct and train the forest state index sequence under historical normal conditions to learn its latent distribution; in the monitoring stage, the reconstruction error of the current state index sequence is calculated, and when the reconstruction error continuously exceeds a preset threshold... When three time steps have elapsed, it is identified as a potential anomaly. The second stage is causal attribution. For the potential anomaly, multi-dimensional driving factor data related to the forest state are extracted within a specific time window before and after the event, including meteorological factors, soil moisture, and human activity index. Granger causality tests are used to initially screen the set of driving factors that have a statistical causal relationship with the abnormal state changes. Then, a Bayesian network is constructed, with the anomaly as the child node and the screened driving factors as the parent node. The conditional probability table of the network is learned based on historical data, and finally, the probability value of each driving factor causing the anomaly is output, realizing the quantitative inference of causal relationship.

[0012] Furthermore, in the multidimensional ecological function comprehensive assessment and early warning module, the comprehensive assessment index system includes three primary indicators: productivity indicators, stability indicators, and resilience indicators. The productivity indicator is a weighted combination of the vegetation greenness index and the simulated net primary productivity value in the forest state index map; the stability indicator is measured by calculating the coefficient of variation of the forest state index in the time dimension and the semivariance function in the spatial dimension; the resilience indicator is quantified by analyzing the rate and extent to which the forest state index recovers to its pre-disturbance level after historical disturbance events. The multidimensional ecological function comprehensive assessment and early warning module uses the analytic hierarchy process (AHP) to determine the weights of each primary indicator and secondary sub-indicator, and finally calculates the comprehensive health score for each assessment unit. The calculation formula is as follows: in, For the first The weight of each primary indicator, This is the normalized value of the indicator. The warning level is based on the comprehensive health score. The predetermined range into which the event falls is determined by the highest risk probability value output by the abnormal event causal tracing module.

[0013] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention constructs a multi-source heterogeneous data acquisition and preprocessing module and adopts spatiotemporal kriging interpolation and joint denoising methods to achieve deep fusion and high-quality preprocessing of multi-scale and multi-temporal observation data under a unified spatiotemporal reference. This effectively makes up for the defects of spatiotemporal continuity discontinuity in existing technologies and provides a reliable data foundation for subsequent high-precision analysis.

[0014] This invention combines a physical ecological model with a deep spatiotemporal neural network by designing a spatiotemporal feature extraction module that embeds ecological process mechanisms, and uses a gated attention mechanism for feature fusion. This allows the extracted feature vectors to simultaneously contain prior ecological mechanisms and inherent data patterns, significantly improving the feature's ability to represent complex nonlinear ecological processes in forests and overcoming the state identification bias problem caused by static or linear assumptions in existing models.

[0015] This invention constructs an adaptive dynamic state representation and evolution module based on a spatiotemporal graph attention network. This model can explicitly model the heterogeneous associations and temporal evolution dependencies between forest spatial units, adaptively learn non-stationary and nonlinear dynamic mechanisms, and thus generate a continuous and high-precision forest state index map, achieving a more accurate and coherent dynamic characterization of the forest ecosystem state.

[0016] This invention integrates a variational autoencoder and a Bayesian network into an early detection and causal tracing module for abnormal events, enabling early detection of sudden disturbances such as wildfires and pests, and quantitatively tracing their key driving factors. It provides a closed-loop analysis capability from "phenomenon perception" to "cause diagnosis," effectively solving the pain points of existing systems' delayed early warning and lack of causal tracing capabilities.

[0017] This invention establishes a multi-dimensional ecological function comprehensive evaluation index system oriented towards productivity, stability, and resilience, and uses the analytic hierarchy process (AHP) for integrated evaluation. This system can comprehensively and quantitatively reflect the overall health status of forest ecosystems. The generated comprehensive health score and graded early warning information provide a direct and reliable basis for ecological governance and scientific decision-making, breaking through the limitations of existing methods that rely on single parameters or empirical thresholds. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical architecture of the forest ecosystem monitoring system based on spatiotemporal big data proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal feature extraction module embedded in the ecological process mechanism in this invention; Figure 3 This is a schematic diagram illustrating the specific steps of the forest ecosystem monitoring method based on spatiotemporal big data proposed in this invention. Detailed Implementation

[0019] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0020] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0021] Example 1 In the long-term monitoring and protection scenarios of national forest parks or key ecological function zones, forest ecosystems face multiple pressures, including outbreaks of pests and diseases, forest fire risks, extreme weather events, and human interference. Traditional monitoring methods rely on manual patrols and analysis of single data sources, which have prominent problems such as limited coverage, delayed response, and difficulty in quantifying and assessing comprehensive ecological health. This embodiment applies the system and method provided by the present invention to achieve high-precision, dynamic, and intelligent monitoring and early warning of target forest areas.

[0022] See Figure 1 This system includes a multi-source heterogeneous data acquisition and preprocessing module, a spatiotemporal feature extraction module embedding ecological process mechanisms, an adaptive dynamic state representation and evolution module, an early perception and causal tracing module for abnormal events, and a multi-dimensional ecological function comprehensive assessment and early warning module. These modules are connected sequentially to form a complete technology chain from data input to decision output.

[0023] The multi-source heterogeneous data acquisition and preprocessing module is responsible for integrating various observational data from space-based, air-based, and ground-based sources. Specifically, this module acquires multispectral, hyperspectral, and radar imagery data covering the target area from multiple optical and synthetic aperture radar satellite remote sensing platforms at preset revisit intervals, with spatial resolution ranging from 30 meters to sub-meter levels. Simultaneously, a ground-based IoT sensor network deployed within the forest area continuously collects near-surface ecological parameters such as soil temperature and humidity, sap flow, forest carbon dioxide concentration, and light intensity, with a sampling frequency of once per hour. Regional meteorological observation stations provide hourly meteorological data such as temperature, precipitation, wind speed, and relative humidity. Furthermore, regularly or urgently deployed UAV aerial survey platforms acquire centimeter-resolution orthophotos and laser point cloud data for fine-scale topographic and forest stand structure inversion. All of the above raw spatiotemporal data are transmitted to the central data processing server in real-time or near real-time.

[0024] The core preprocessing operations of this module include spatiotemporal reference unification, missing value interpolation, and noise filtering. The spatiotemporal reference unification operation specifically involves unifying all satellite remote sensing imagery and UAV aerial data to a preset universal geographic coordinate system using a reprojection algorithm, such as the CGCS2000 coordinate system and UTM projection. The timestamps of all data are converted and aligned to a unified Coordinated Universal Time (UTC) reference. For data missing due to cloud cover, sensor malfunction, or communication interruption, a spatiotemporal kriging interpolation algorithm is used for filling in the missing data. This algorithm considers both spatial autocorrelation and the periodicity of the time series, and its interpolation weight function is defined as: in, Represents known data points For missing data points Contribution weight, This represents the spatial Euclidean distance between two points. Indicates the time interval between two points. and The spatial and temporal scale parameters are preset based on data characteristics to control the decay rate of spatial and temporal correlation. The algorithm iterates through each missing point, calculates its weight with all known points within a certain spatiotemporal window, and sums the weighted values ​​to obtain the interpolated value. For noise filtering, a joint denoising method based on wavelet transform and the 3σ criterion is adopted. First, wavelet decomposition is performed on the time-series sensor data and image pixel value sequence to separate components of different frequencies. Then, thresholding based on the 3σ criterion is applied to the high-frequency detail components, and coefficients with amplitudes exceeding three standard deviations of the historical sequence mean are regarded as noise and removed. Finally, wavelet reconstruction is performed to obtain a smooth data sequence after filtering out high-frequency noise and outliers. After the above preprocessing, a standardized, spatiotemporally aligned, and quality-controllable multi-source spatiotemporal dataset is generated and stored in a distributed spatiotemporal database for subsequent modules to use.

[0025] The spatiotemporal feature extraction module embedded with ecological process mechanisms receives the aforementioned standardized multi-source spatiotemporal dataset. See also... Figure 2This module contains a dual-channel feature extraction network designed to integrate prior knowledge of physical and ecological processes with data-driven deep features. The first channel is a physical and ecological process simulation channel. This channel integrates a locally calibrated photosynthesis model, a water stress model, and a nutrient cycling model. Input data includes preprocessed hourly meteorological data and monthly soil property data. The photosynthesis model, based on light energy utilization theory, simulates and calculates the proportion of photosynthetically active radiation absorbed by vegetation at each spatial grid scale. The water stress model combines soil moisture, evapotranspiration, and vegetation root distribution to calculate the vegetation water stress index. The nutrient cycling model simulates the migration and transformation processes of key elements such as nitrogen and phosphorus in the soil-vegetation system, outputting a nitrogen and phosphorus availability index. These model outputs constitute a physical feature vector reflecting the ecological mechanisms. The second channel is a deep spatiotemporal neural network channel. This channel employs an architecture consisting of cascaded 3D convolutional layers and long short-term memory (LSTM) network layers. Input is a preprocessed, chronologically ordered sequence of multi-temporal remote sensing images. 3D convolutional layers perform convolution operations simultaneously in both spatial and temporal dimensions to extract local spatiotemporal features from image sequences, such as greenness fluctuations caused by phenological changes and spectral abrupt changes in texture due to forest fires or logging. Subsequently, long short-term memory (LSTM) network layers capture the dependencies and evolution trends of these local features over longer time scales, learning a data-driven deep spatiotemporal feature vector. The dual-channel output is fed into a feature fusion layer. This fusion layer employs a gated attention mechanism, containing a learnable attention weight generation network. This network dynamically calculates the contribution weights of features from the physical model channel and features from the data-driven channel to the final output based on the context information of the current input. Specifically, for the physical feature vector... and deep feature vectors The fusion layer generates attention weights. and ,satisfy The final output is a multidimensional spatiotemporal feature vector. for: in, This represents a linear transformation layer used to map features from different channels to the same dimension. Through this mechanism, the module can adaptively balance the interpretability of the mechanistic model with the flexibility of the data model in different regions and at different times, outputting a highly representative multidimensional spatiotemporal feature vector that integrates prior ecological mechanisms and the inherent laws of the data.

[0026] The adaptive dynamic state representation and evolution module receives the aforementioned multidimensional spatiotemporal feature vectors. This module constructs and runs a dynamic forest state representation model based on an attention mechanism and graph neural network. First, the monitoring area is discretized into regular spatial grid cells, each cell serving as a graph node. The initial features of a node are its corresponding multidimensional spatiotemporal feature vectors. Edges between nodes are defined by two relationships: spatial adjacency, where adjacent grid cells sharing a boundary are automatically connected; and ecological similarity, calculated by the cosine similarity of historical feature vector sequences between nodes. If the similarity exceeds a preset threshold, a connection is established between the two nodes. The edge weights combine the reciprocal of the spatial distance and the ecological similarity value. The core of the model is the spatiotemporal graph attention network layer. This layer performs two aggregation operations at each time step when updating the node state. One is spatial neighbor aggregation: for nodes... Calculate its relationship with each spatial neighbor node. attention weights This weight is determined by the node. and In time The characteristics of each node, along with the edge weights between them, jointly determine the varying importance of different neighbors to the current node's state. Secondly, there is temporal self-aggregation: computing nodes... The attention weights between the features at the current time step and the hidden state at the previous time step This is used to capture the evolutionary dependencies of its own state. Node In time Hidden state The updated formula is: in, Represents a node In time The hidden state, Represents a node Spatial neighbor set, and For a learnable parameter matrix, For bias terms, The activation function is nonlinear. By stacking multiple such spatiotemporal graph attention network layers, the model can abstract layer by layer, capturing the complex nonlinear evolution patterns and heterogeneous relationships of forest states in the spatiotemporal dimension. Finally, the hidden states of the nodes in the last network layer are passed through a fully connected output layer and mapped to the forest state index of each spatial unit at each time step. This index is a normalized scalar value that comprehensively reflects the state information of the unit, such as vegetation vitality and structural integrity. The indices of all spatial units constitute a forest state index map on a continuous spatiotemporal grid in time series. This map is stored in raster data form with a temporal resolution of up to days or weeks.

[0027] The early detection and causal attribution module for abnormal events monitors the dynamic changes in the forest state index map in real time. The module's workflow is divided into two stages: abnormal event detection and causal attribution. The first stage of anomaly detection employs a variational autoencoder (VAE) architecture. First, a forest state index sequence that has not experienced any known major perturbations over a long historical period is used as training data to reconstruct the VAE. The encoder compresses the input state index sequence into a low-dimensional latent vector, and the decoder attempts to reconstruct the original sequence from this latent vector. The training objective is to minimize the reconstruction error while making the distribution of the latent vector approximate a standard normal distribution. After training, the VAE learns the latent distribution and reconstruction capability of the forest state index sequence under normal conditions. In the real-time monitoring phase, the state index sequences of the current time step and several previous time steps are input into the trained VAE, and its reconstruction error is calculated. The system sets a dynamically adjusted reconstruction error threshold. This threshold is determined based on the statistical distribution of reconstruction errors in historical normal sequences. When the reconstruction error of a spatial cell exceeds this threshold for three consecutive time steps, a potential abnormal event is identified in that cell, triggering an alarm and recording the time and spatial location of the event.

[0028] The second phase of causal attribution focuses on the identified potential anomalous events. First, it extracts multi-dimensional driving factor data sequences related to the spatial unit within a specific time window before and after the event's occurrence. These sequences include, but are not limited to: meteorological factor sequences such as temperature, precipitation, and drought index; soil moisture sensor data sequences; and regional human activity intensity index sequences based on nighttime light or traffic data. Next, using the Granger causality test, it analyzes the statistical causal relationship between each driving factor sequence and the anomalous state index sequence. The Granger causality test compares the performance of a prediction model containing historical information about the driving factor with that model when predicting future values ​​of the state index sequence, determining whether the driving factor statistically "caused" the change in state. This step filters out a subset of driving factors with significant Granger causal relationships to the anomalous event. Subsequently, a Bayesian network is constructed for deep causal probability inference. In this Bayesian network, the anomalous event is a child node, and the selected driving factors are parent nodes. Based on historical data from the same period, the network's conditional probability table is learned using maximum likelihood estimation or Bayesian estimation methods. This quantifies the probability of child nodes (abnormal events) occurring when each parent node (driving factor) is in different states. Finally, the early perception and causal attribution module outputs the probability value of each driving factor leading to the current abnormal event, such as "the probability that continuous high temperatures and drought have caused an abnormal forest state in this area is 85%." This achieves quantitative causal attribution from abnormal phenomena to key driving factors.

[0029] The multidimensional ecological function comprehensive assessment and early warning module integrates forest state index maps and causal tracing results of abnormal events to perform comprehensive assessments and generate early warnings. This module calculates based on a pre-set comprehensive assessment index system, which includes three primary indicators: productivity, stability, and resilience. The productivity indicator is composed of the vegetation greenness index extracted from the forest state index map and the net primary productivity value simulated by the physical model in the spatiotemporal feature extraction module embedded with ecological process mechanisms, weighted according to preset weights, and is used to characterize the material production capacity of the ecosystem. The stability indicator is measured through two dimensions: temporal stability, calculating the coefficient of variation of the forest state index for each spatial unit within the assessment time window; and spatial stability, analyzing the spatial autocorrelation and heterogeneity of the forest state index within the assessment area by calculating its semivariogram function. The resilience indicator is quantified by analyzing the rate and degree of recovery of the forest state index to the pre-disturbance baseline level after similar abnormal disturbance events in historical records. The rate is obtained by fitting the parameters of the index recovery curve, and the degree of recovery is obtained by comparing the ratio of the recovered stable value to the baseline value. Each primary indicator can be further subdivided into several secondary sub-indicators. The module employs the analytic hierarchy process (AHP) to determine the weights of each indicator: experts in ecology and forestry are invited to conduct pairwise comparisons of the importance of each indicator, constructing a judgment matrix. After passing a consistency test, the eigenvectors of each indicator are calculated as their weights. Finally, a comprehensive health score is assigned to each assessment unit. Calculated using the weighted summation formula: in, For the first The weight of each primary indicator, This is the normalized value of the indicator. Normalization maps each indicator value to a range of 0 to 1, with 1 representing the optimal state. Determining the warning level is a multi-rule decision-making process. First, based on the comprehensive health score... The preset intervals into which the object falls are initially classified, for example: For "health", "Sub-health" It is labeled "unhealthy". Then, it is corrected by combining the highest risk probability value output by the abnormal event causal tracing module. For example, even if a certain area... The score is in the "sub-healthy" range, but if a high-risk forest fire event is detected and its driving factor probability exceeds 90%, the warning level is raised to the highest level, "red alert." The final warning information includes the target area location, comprehensive health score, main degradation indicators, types of detected abnormal events, key driving factors and their probabilities, and recommended response measures, which are pushed to forest management departments and decision-makers through various channels such as visualization platforms, SMS, and email.

[0030] Example 2 In the forest carbon sequestration measurement and monitoring project carried out by provincial forestry departments, it is necessary to make high-precision estimates of the current status and dynamic changes of carbon storage of all forest resources within the jurisdiction, and to assess the impact of various natural and human activities on carbon sequestration function. This embodiment applies the system and method of the present invention, focusing on deepening the dimensions of carbon sequestration function assessment.

[0031] The system architecture of this embodiment is the same as that of Embodiment 1, but the evaluation index system has been specifically expanded and customized in the multi-dimensional ecological function comprehensive assessment and early warning module. Under the productivity index, a secondary sub-index, "carbon absorption rate," has been added. This index is directly coupled to the net primary productivity simulated by the physical model in the spatiotemporal feature extraction module embedded with ecological process mechanisms, and is obtained through the carbon content conversion coefficient. In the stability index, the measurement of "spatiotemporal volatility of carbon storage" has been strengthened. It not only calculates the coefficient of variation of the forest state index as a proxy indicator for carbon storage, but also pays special attention to the magnitude and duration of carbon storage decline under stress conditions such as drought and frost. The resilience index specifically targets "carbon sink function resilience," analyzing the path and time for the regional carbon absorption rate to recover to its original level after historical disturbance events.

[0032] In terms of data acquisition, in addition to the data sources in Example 1, the access and integration of forest resource inventory data and measured forest growth data have been strengthened. The multi-source heterogeneous data acquisition and preprocessing module needs to integrate these non-spatiotemporally continuous point or sample plot survey data into a unified spatiotemporal grid dataset through spatiotemporal interpolation and scale push-up methods.

[0033] In the spatiotemporal feature extraction module embedded with ecological process mechanisms, the physical ecological process simulation channel enhances the embedding of carbon cycle process models. For example, it integrates process-based biogeochemical models to more precisely simulate processes such as vegetation photosynthesis, respiration, litter decomposition, and soil organic carbon transformation, and outputs grid-scale carbon flux and carbon storage change characteristics.

[0034] In this embodiment, the forest state index map generated by the adaptive dynamic state characterization and evolution module is given a more explicit indication of its carbon sink function. By establishing a regression relationship between the state index and ground-measured carbon storage data, the map can be calibrated, allowing changes in the state index to more directly reflect the dynamics of carbon storage.

[0035] The early detection and causal attribution module for abnormal events focuses on anomalous events that may lead to carbon loss, such as large-scale illegal logging, severe forest fires, and widespread tree death caused by pests and diseases. The factor analysis driven by causal attribution needs to incorporate management information such as timber harvesting permit data and forest fire prevention deployment data to distinguish between natural and human disturbances.

[0036] Ultimately, the comprehensive health score output by the multidimensional ecological function assessment and early warning module will be directly related to the health of forest carbon sink functions. The early warning information not only includes the ecological health status but also provides estimates of carbon storage changes, early warnings of carbon loss risks, and suggestions for carbon sequestration and enhancement measures based on causal tracing, providing precise data support and decision-making basis for the development, management, and trading of forestry carbon sink projects.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A forest ecosystem monitoring system based on spatiotemporal big data, characterized in that, It includes the following components: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire raw spatiotemporal data from satellite remote sensing platforms, ground IoT sensor networks, meteorological observation stations, and UAV aerial survey platforms, and to preprocess the raw spatiotemporal data to generate standardized multi-source spatiotemporal datasets. The spatiotemporal feature extraction module embedded in the ecological process mechanism is used to receive the standardized multi-source spatiotemporal dataset, and extract multi-dimensional spatiotemporal feature vectors representing forest structure, function and disturbance response from the dataset by coupling the physical ecological process model with the deep spatiotemporal neural network. An adaptive dynamic state representation and evolution module is used to receive the multidimensional spatiotemporal feature vector, construct a dynamic forest state representation model based on attention mechanism and graph neural network, and generate a forest state index map on a continuous spatiotemporal grid. The module for early detection and causal attribution of abnormal events is used to monitor the dynamic changes of the forest state index map in real time. By constructing an anomaly detection framework based on variational autoencoder and Granger causality test, it identifies abnormal perturbation events that exceed the normal evolution threshold and performs causal probability inference on the potential driving factors of abnormal events. The multidimensional ecological function comprehensive assessment and early warning module is used to integrate the forest state index map and the causal tracing results of abnormal events, calculate the comprehensive health score of the forest ecosystem based on the preset comprehensive assessment index system, and generate graded early warning information when the comprehensive health score is lower than the preset threshold or when a high-risk abnormal event is detected.

2. The forest ecosystem monitoring system based on spatiotemporal big data according to claim 1, characterized in that, The spatiotemporal feature extraction module embedded with the ecological process mechanism includes a dual-channel feature extraction network: the first channel is a physical ecological process simulation channel, which integrates a photosynthesis model, a water stress model, and a nutrient cycling model. It takes meteorological data and soil data as input to simulate and calculate the vegetation photosynthetically active radiation absorption ratio, water use efficiency, and nitrogen and phosphorus availability index at the grid scale; the second channel is a deep spatiotemporal neural network channel, which adopts an architecture of cascaded 3D convolutional layers and long short-term memory network layers to directly learn the spectral, texture, and phenological temporal change features from multi-temporal remote sensing image sequences; the outputs of the dual channels are integrated through a feature fusion layer with a gated attention mechanism, which dynamically adjusts the feature contribution weights from the physical model channel and the data-driven channel, and outputs a multi-dimensional spatiotemporal feature vector that integrates prior ecological mechanisms and the inherent laws of the data.

3. The forest ecosystem monitoring system based on spatiotemporal big data according to claim 1, characterized in that: The dynamic forest state representation model can adaptively learn the non-stationary evolution patterns of forest ecosystems in the time dimension and the heterogeneous relationships in the spatial dimension. It represents forest spatial units and their relationships in a graph structure, with each spatial unit as a graph node. Its initial node features are the multi-dimensional spatiotemporal feature vectors, and the edge weights between nodes are jointly determined by spatial adjacency and ecological similarity. The core of the dynamic forest state representation model is a spatiotemporal graph attention network layer. When the network layer updates the node state at each time step, it aggregates the information of its spatial neighbor nodes and its own state information in historical time steps.

4. The forest ecosystem monitoring system based on spatiotemporal big data according to claim 1, characterized in that, The workflow of the early perception and causal tracing module for abnormal events includes two stages: the first stage is abnormal perception, which uses a variational autoencoder to reconstruct and train the forest state index sequence under historical normal conditions. During the monitoring stage, the reconstruction error of the current state index sequence is calculated. When the reconstruction error exceeds a preset threshold for three consecutive time steps, it is determined to be a potential abnormal event. The second stage is causal attribution. For the potential abnormal events, multi-dimensional driving factor data within a specific time window before and after the event occurs are extracted. Granger causality test is used to screen the set of driving factors that have a statistical causal relationship with the abnormal state change. Then, a Bayesian network is constructed with the abnormal event as the child node and the screened driving factors as the parent node. The conditional probability table of the network is learned based on historical data, and the probability value of each driving factor causing the abnormal event is output.

5. The forest ecosystem monitoring system based on spatiotemporal big data according to claim 1, characterized in that: In the multidimensional ecological function comprehensive assessment and early warning module, the comprehensive assessment index system includes three primary indicators: productivity index, stability index, and resilience index. The productivity index is composed of a weighted combination of the vegetation greenness index and the simulated net primary productivity value in the forest state index map. The stability index is measured by calculating the coefficient of variation of the forest state index in the time dimension and the semivariance function in the spatial dimension. The resilience index is quantified by analyzing the rate and extent to which the forest state index recovers to the pre-disturbance level after historical disturbance events.

6. A method for monitoring forest ecosystems based on spatiotemporal big data, characterized in that, Includes the following steps: Step S110: Collect and preprocess multi-source heterogeneous spatiotemporal data to generate a standardized multi-source spatiotemporal dataset; Step S120: Based on the ecological process mechanism and deep spatiotemporal neural network, extract multidimensional spatiotemporal feature vectors from the standardized multi-source spatiotemporal dataset; Step S130: Construct a dynamic forest state representation model and use the multidimensional spatiotemporal feature vector to generate a forest state index map on a continuous spatiotemporal grid. Step S140: Based on the forest state index map, perform early perception and causal source analysis of abnormal events; Step S150: Based on the multidimensional ecological function comprehensive assessment index system, assess the health status of the forest ecosystem and generate early warning information.

7. The forest ecosystem monitoring method based on spatiotemporal big data according to claim 6, characterized in that: In step S110, the preprocessing operations include spatiotemporal benchmark unification, missing value interpolation, and noise filtering; the missing value interpolation adopts a spatiotemporal kriging interpolation algorithm; the interpolation algorithm considers both spatial autocorrelation and the periodicity of the time series.

8. The forest ecosystem monitoring method based on spatiotemporal big data according to claim 6, characterized in that: In step S130, the construction of the dynamic forest state representation model is specifically as follows: forest spatial units are represented by a graph structure, each unit is a graph node, and its initial feature is the multidimensional spatiotemporal feature vector; the node state is updated through the spatiotemporal graph attention network layer, and finally the hidden state of the node is mapped to the forest state index to form a forest state index map.

9. The forest ecosystem monitoring method based on spatiotemporal big data according to claim 6, characterized in that: In step S150, calculations are performed based on a multi-dimensional ecological function comprehensive evaluation index system, which includes three primary indicators: productivity, stability, and resilience. A comprehensive health score for each evaluation unit is calculated using a weighted summation formula. The warning level is determined jointly based on the preset range into which the comprehensive health score falls and the highest risk probability value output by the causal source analysis of the abnormal event.

Images