An alpine cold region ecological quality observation positioning method and system

Abstract

The application relates to the technical field of ecological environment monitoring and intelligent data analysis, and discloses a method and system for observing and positioning ecological quality in an alpine endorheic region, which comprises the following steps: arranging a three-dimensional observation network in the target alpine endorheic region, continuously collecting data in a hierarchical classification mode, forming a multi-ecosystem collaborative monitoring data set, standardizing the data set, and constructing a structured standard database; inputting time-series monitoring data in the standard database into a pre-trained correlation analysis model, respectively extracting dynamic spatial correlation features between observation field nodes, multi-scale periodic features of time-series data, and nonlinear trend features, and adaptively fusing the features to generate comprehensive feature vectors of the observation field nodes; simultaneously outputting quantitative results representing the collaborative relationship between different observation field nodes; and using a prediction model to perform short-term, medium-term and long-term trend prediction on core ecological indexes. The method realizes multi-period and confidence interval precise prediction of core ecological indexes of a multi-ecosystem in an alpine endorheic region.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring and intelligent data analysis technology, specifically to a method and system for ecological quality observation and positioning in high-altitude inland drainage areas. Background Technology

[0002] With increasing demands for global climate change and ecological protection, the scientific rigor and forward-looking nature of ecological quality monitoring in high-altitude inland drainage basins—as ecologically fragile areas and core areas of the hydrological cycle—is becoming increasingly crucial. Currently, significant bottlenecks exist in ecological monitoring and analysis technologies for high-altitude inland drainage basins:

[0003] 1. The observation system lacks systematicness and specificity: existing observations are mostly scattered and have not formed an observation network that covers the entire ecosystem and coordinates all elements. It is difficult to capture the correlation effects between different ecosystems (such as wetlands and surrounding meadows), and the data lacks specificity and completeness.

[0004] 2. Weak cross-ecosystem correlation analysis capability: Traditional methods often analyze single ecological indicators in isolation, failing to effectively explore the spatiotemporal synergistic relationships between observation fields and ecosystems, and failing to identify the core driving factors of ecological quality changes (such as the coupling effect of precipitation and permafrost thawing).

[0005] 3. Poor adaptability of long-term time series forecasts: Ecological data in high-altitude inland drainage areas are characterized by strong seasonality, nonlinearity, and long-term dependence. Existing forecast models mostly adopt general time series modeling methods, which are not adapted to their unique ecological evolution patterns, resulting in low accuracy and high uncertainty in multi-period trend forecasts.

[0006] 4. Insufficient fusion of multi-dimensional features: Ecological data contains multi-dimensional features such as spatial correlation, time cycle, and nonlinear fluctuation. Existing technologies lack effective feature integration mechanisms, making it difficult to fully utilize the value of data to support accurate analysis and prediction.

[0007] Therefore, existing technologies cannot provide comprehensive, accurate, and forward-looking scientific support for ecological protection decisions in high-altitude inland drainage areas, and there is an urgent need to build an integrated technical solution of "observation-analysis-prediction". Summary of the Invention

[0008] To address the shortcomings of existing technologies, the core objective of this invention is to provide a method and system for ecological quality observation and positioning in high-altitude inland drainage basins. This invention focuses on the need for coordinated monitoring of all elements of "hydrology-soil-meteorology-ecology" in high-altitude inland drainage basins, integrating observation network layout, multi-source data standardization processing, adaptive spatiotemporal map attention analysis and long-term time series prediction technology. It is applicable to the synergistic correlation mining of multiple ecosystems (wetlands, meadows, deserts, lakes, etc.) in typical high-altitude inland drainage basins such as the Shaliu River Basin, accurate inversion of core ecological indicators, and short-term (six months), medium-term (3 years), and long-term (5 years) trend prediction. It can be widely applied to ecological protection and restoration, water resource regulation, and climate change response in high-altitude inland drainage basins.

[0009] To achieve the above objectives, the following technical solution is adopted:

[0010] In a first aspect, embodiments of the present invention provide a method for observing and locating the ecological quality of a high-altitude inland drainage basin, comprising the following steps:

[0011] A three-dimensional observation network was established in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field, and observation fields of various ecological types. Sampling points and monitoring instruments covering hydrological, soil, meteorological, and ecological elements were set up in each observation field. Continuous data collection was carried out in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset.

[0012] The multi-ecosystem collaborative monitoring dataset was formatted and normalized to construct a structured ecological quality observation standard database for high-altitude inland drainage areas.

[0013] The time-series monitoring data from the standard database is input into a pre-trained correlation analysis model based on an adaptive spatiotemporal graph attention network (ASTNet). The correlation analysis model extracts dynamic spatial correlation features between observation field nodes, multi-scale periodic features, and nonlinear trend features from the time-series data. These features are then adaptively fused to generate a comprehensive feature vector for each observation field node. Simultaneously, based on the model's learning results, a quantitative result representing the collaborative relationship between different observation field nodes is output.

[0014] The integrated feature vector and / or the time-series monitoring data from the standard database are input into a prediction model constructed based on a time-series attention gating unit (TAGU) to perform short-term, medium-term, and long-term trend predictions on at least one core ecological indicator among permafrost active layer thickness, vegetation cover change rate, and hydrological runoff, and output prediction results with confidence intervals.

[0015] Furthermore, the three-dimensional observation network specifically includes: one comprehensive observation field, four evapotranspiration observation fields respectively located in alpine wetlands, meadows, deserts and around lakes, three meteorological observation fields, five hydrological observation fields, and seven ecological observation fields respectively covering alpine wetlands, shrublands, meadows, deserts, alpine vegetation, vegetated areas and lake shorelines.

[0016] Furthermore, the adaptive spatiotemporal graph attention network includes at least:

[0017] An adaptive spatial graph encoder is used to extract dynamic spatial correlation features between observation field nodes and output spatial feature vectors.

[0018] The seasonal capture module is used to extract multi-scale periodic features from the input time-series data and output seasonal feature vectors.

[0019] The frequency domain decomposition module is used to extract nonlinear frequency domain features from the input time-series data and output a frequency domain feature vector.

[0020] The feature fusion layer is used to adaptively fuse the spatial feature vector, seasonal feature vector and frequency domain feature vector to generate a comprehensive feature vector that characterizes the state of the observation field nodes and the potential correlation between nodes.

[0021] Furthermore, the adaptive spatial graph encoder calculates the feature similarity between each observation field node through a learnable node embedding matrix, and dynamically generates an adaptive adjacency matrix through nonlinear activation and normalization processing to quantify the time-varying spatial correlation strength between different observation field nodes.

[0022] Using the adaptive adjacency matrix and the input features of the observation field nodes, at least one layer of graph convolution operation is performed to aggregate the neighborhood features of each observation field node in order to learn and output the spatial feature vector.

[0023] Furthermore, the seasonal capture module performs one-hot encoding on the intraday time period, intraweek date and season contained in the time information corresponding to the input data to obtain three independent sparse encoding vectors.

[0024] The three sparse coding vectors are concatenated to form a high-dimensional feature vector that integrates multi-scale time period information.

[0025] Through a learnable linear transformation layer, the spliced ​​high-dimensional feature vector is mapped into a low-dimensional dense seasonal embedding vector, which is then output as the seasonal feature vector to explicitly characterize the multi-scale periodic evolution of alpine ecosystems.

[0026] Furthermore, the frequency domain decomposition module applies a discrete Fourier transform to the input standardized time-series monitoring data, converting it from a time-domain representation to a frequency-domain representation;

[0027] By analyzing the frequency domain representation, high-frequency fluctuation components and low-frequency trend components in the time series data are separated and extracted.

[0028] Based on the high-frequency fluctuation component and the low-frequency trend component, the frequency domain feature vector is constructed and output to enhance the capture of nonlinear features in the time series data.

[0029] Furthermore, the feature fusion layer generates the comprehensive feature vector in the following manner:

[0030] The system receives the spatial feature vector output by the adaptive spatial map encoder, the seasonal feature vector output by the seasonal capture module, and the frequency domain feature vector output by the frequency domain decomposition module.

[0031] The spatial feature vector, seasonal feature vector, and frequency domain feature vector are weighted linearly combined using a set of learnable fusion weight parameters, and a learnable bias term is added.

[0032] The result of the linear combination is input into the sigmoid activation function to achieve dynamic adaptive adjustment and fusion of the contribution weights of spatial correlation features, time periodic features and nonlinear trend features, and output the comprehensive feature vector.

[0033] Furthermore, the prediction model built based on the temporal attention gating unit adopts an encoder-decoder architecture and includes at least:

[0034] The location coding layer is used to embed location information into the input time-series data;

[0035] The encoder, consisting of four layers of stacked temporal attention gating units, is used to encode the position-encoded input sequence and extract deep features.

[0036] The decoder, consisting of two stacked layers of temporal attention gating units, is used to decode based on encoded features and generate the hidden state representation of the predicted sequence.

[0037] The output layer is used to map the hidden state representation output by the decoder to the final ecological index prediction value.

[0038] The temporal attention gating unit integrates a gating loop unit and a multi-head self-attention mechanism to collaboratively capture local short-term dependencies and global long-term dependencies in temporal data.

[0039] Furthermore, during the training process, the prediction model built based on the Temporal Attention Gating Unit (TAGU) uses a hybrid loss function to optimize the model parameters. The hybrid loss function is a weighted sum of the mean absolute error loss and the root mean square error loss.

[0040] When making trend predictions, the Monte Carlo simulation method is used to perform multiple random sampling predictions on the prediction model. The standard deviation of the predicted values ​​is calculated based on the statistical distribution of the multiple prediction results, and a prediction confidence interval with a specific confidence level is generated accordingly.

[0041] Secondly, embodiments of the present invention also provide an ecological quality observation and positioning system for high-altitude inland drainage basins, comprising:

[0042] The data acquisition module is used to deploy a three-dimensional observation network in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field, and observation fields of various ecological types. Within each observation field, sampling points and monitoring instruments covering hydrological, soil, meteorological, and ecological elements are deployed. Continuous data acquisition is carried out in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset.

[0043] The data processing module is used to format and normalize the multi-ecosystem collaborative monitoring dataset to construct a structured ecological quality observation standard database for high-altitude inland drainage areas.

[0044] The intelligent analysis module is used to input the time-series monitoring data from the standard database into a pre-trained correlation analysis model based on the Adaptive Spatiotemporal Graph Attention Network (ASTNet). The correlation analysis model extracts the dynamic spatial correlation features between observation field nodes, the multi-scale periodic features and the nonlinear trend features in the time-series data, and adaptively fuses the dynamic spatial correlation features, multi-scale periodic features and nonlinear trend features to generate a comprehensive feature vector for each observation field node, and outputs a quantitative result representing the collaborative relationship between different observation field nodes.

[0045] The trend prediction module is used to input the comprehensive feature vector and / or the time-series monitoring data in the standard database into a prediction model constructed based on the temporal attention gating unit (TAGU), and to make short-term, medium-term and long-term trend predictions for at least one core ecological indicator among the frozen soil active layer thickness, vegetation cover change rate, and hydrological runoff, and output prediction results with confidence intervals.

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

[0047] 1. The observation system is highly systematic and the data support is reliable: The "1+4+3+5+7" observation field system achieves full coverage of multiple ecosystems, and the full-element sampling network ensures that no key indicators are missed. The annual data volume is ≥50,000, providing comprehensive and continuous high-quality data for subsequent analysis and prediction, which improves the data integrity by more than 80% compared with traditional scattered observations.

[0048] 2. High accuracy in correlation analysis and accurate identification of driving factors: The adaptive spatiotemporal graph attention network model mines multi-dimensional features through multi-component collaboration, achieving a spatial correlation capture accuracy of ≥85%. The accuracy of key driving factor identification is 30% higher than that of traditional methods. It can accurately quantify the contribution weight of factors such as precipitation and permafrost thawing, providing a scientific basis for ecological protection decision-making.

[0049] 3. Excellent predictive adaptability and reliable multi-period results: The TAGU model is specifically adapted to the ecological data characteristics of high-altitude inland drainage areas, with a long-term prediction error of ≤8%, which is 40% more accurate than general time series models. It supports multi-period prediction and outputs 90% confidence intervals, providing forward-looking support for ecological protection planning at different time scales.

[0050] 4. Highly practical technology with controllable deployment costs: The observation network is compatible with existing monitoring equipment, the data processing flow is standardized and modularized, the model training and inference efficiency is high, no complex hardware upgrades are required, and it can be directly connected to the existing ecological monitoring system in high-altitude inland drainage areas, making it easy to promote and apply in typical areas such as the Shaliu River Basin.

[0051] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0052] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0053] Figure 1 This is a schematic diagram of a method for monitoring and locating ecological quality in a high-altitude inland drainage basin provided by an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the main process architecture of the integrated observation-analysis-prediction constructed according to an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the intelligent analysis (Adaptive Spatiotemporal Graph Attention Network ASTNet) process according to an embodiment of the present invention;

[0056] Figure 4 This is a flowchart illustrating the prediction model based on a temporal attention gating unit (TAGU) according to an embodiment of the present invention.

[0057] Figure 5 This is a schematic diagram of a module of an ecological quality observation and positioning system for high-altitude inland drainage basins provided in an embodiment of the present invention;

[0058] Figure 6 This is a comparison chart of the accuracy of the algorithm of this invention and the baseline algorithm. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0060] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0061] like Figure 1 and Figure 2 As shown, a method for ecological quality observation and positioning in high-altitude inland drainage basins includes the following steps:

[0062] S1: Deploy a three-dimensional observation network in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field, and observation fields of various ecological types. In each observation field, deploy sampling points and monitoring instruments covering hydrological, soil, meteorological, and ecological elements. Collect data continuously in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset.

[0063] Step S1 is used to construct the observation and positioning system and acquire data, including the following steps:

[0064] S1.1: Precise positioning and layout of the observation field

[0065] In response to the characteristics of the coordinated evolution of multiple ecosystems in high-altitude inland drainage basins, a dedicated observation field system of "1+4+3+5+7" is constructed in core monitoring areas (such as the Shaliu River Basin) to achieve full coverage of the entire ecosystem and spatial range:

[0066] One comprehensive observation field: located in the core hub area of ​​the basin, integrating the monitoring functions of all elements of "hydrology-soil-meteorology-ecology", serving as a data integration and calibration center;

[0067] Four evapotranspiration observation sites are set up in alpine wetlands, meadows, deserts, and around lakes, respectively, to monitor the evapotranspiration process of different ecosystems and support hydrological cycle analysis.

[0068] Three meteorological observation fields are evenly distributed in the upper, middle and lower reaches of the basin to monitor meteorological elements such as air pressure, temperature, humidity and precipitation, and to capture regional climate spatial differences.

[0069] Five hydrological observation sites: covering the main streams and tributaries of rivers, lakes, permafrost areas and groundwater recharge areas, to monitor hydrological indicators such as water level, flow rate and water quality;

[0070] Seven ecological observation sites: specifically covering alpine wetlands, alpine shrublands, alpine meadows, alpine deserts, alpine vegetation, areas without vegetation, and lake shore ecosystems, achieving full coverage of multiple ecological types.

[0071] S1.2: Deployment of a full-element sampling network

[0072] A comprehensive sampling network encompassing hydrology, soil, meteorology, and ecology was constructed within the observation field to ensure that no key indicators were missed.

[0073] (1) Long-term artificial sampling points: a total of 11 points were set up, including 5 river water sampling points (covering key sections of the main stream and tributaries), 2 permafrost water sampling points (core area of ​​active permafrost layer), 3 precipitation sampling points (different altitude gradients), and 2 soil sampling points (typical areas of different ecosystems).

[0074] (2) Monitoring instrument configuration: Equipped with more than 100 sets of special instruments, including hydro-meteorological observation system, frozen soil active layer hydrothermal process monitoring system, flow meter, rain gauge, vegetation growth monitoring instrument (EM50 series), soil physicochemical property analyzer, etc., to realize continuous automated monitoring of indicators such as water level, flow rate, temperature, humidity, soil fertility, vegetation coverage, and dissolved oxygen in water.

[0075] S1.3: Hierarchical and Categorical Data Acquisition Strategy

[0076] Data collection schemes are designed in tiers according to ecosystem type, highlighting the monitoring priorities of different ecosystems and improving the relevance of the data:

[0077] Wetland ecosystem: The key indicators collected include hydrological connectivity, dissolved oxygen in water, wetland vegetation coverage, and soil moisture content. The sampling interval is 30 minutes, with a focus on capturing dynamic changes in hydrology.

[0078] Meadow ecosystem: Focusing on indicators such as soil moisture, vegetation biomass, vegetation cover change rate, and soil nutrient content, sampling was conducted at 1-hour intervals, and vegetation phenological changes were recorded simultaneously;

[0079] Desert ecosystems: Focus on monitoring indicators such as wind and sand activity intensity, soil fertility, surface temperature, and water evaporation, with a sampling interval of 2 hours, and strengthen data capture under extreme weather conditions;

[0080] Lake hydrological system: Core water quality indicators (pH value, nitrogen and phosphorus concentration, transparency), water level changes, lake shore vegetation coverage and other indicators are collected at 1-hour intervals and synchronously calibrated in combination with satellite remote sensing data;

[0081] General data collection requirements: All observation data must be recorded with accurate geographic coordinates and timestamps to ensure data traceability and correlation, with a cumulative annual data volume of ≥50,000 records, forming a multi-ecosystem collaborative monitoring dataset.

[0082] S2: The multi-ecosystem collaborative monitoring dataset is formatted and normalized to construct a structured ecological quality observation standard database for high-altitude inland drainage areas;

[0083] Step S2 is used to standardize the observation data, standardize the monitoring data from different observation sites and instruments, and establish a unified format ecological quality observation standard database for high-altitude inland drainage areas.

[0084] (1) Format standardization: The unified data format is CSV format, and the fields include observation time (accurate to the second), geographic coordinates (latitude and longitude accurate to 0.001°), indicator name, indicator value, instrument number, observation field type, etc.

[0085] (2) Dimensional unification: Normalize indicators with different units, such as temperature (°C), humidity (%), soil fertility (g / kg), etc. Use Min-Max normalization to map the data to the [0,1] interval, which facilitates cross-indicator correlation analysis;

[0086] (3) Database construction: MySQL database is used for data storage. A four-level index structure of "observation field-ecosystem type-indicator category-time series" is constructed to support fast query and data association. The database supports concurrent access and regular backup to ensure data security.

[0087] S3: Input the time-series monitoring data from the standard database into a pre-trained correlation analysis model based on an adaptive spatiotemporal graph attention network (ASTNet). The correlation analysis model extracts dynamic spatial correlation features between observation field nodes, multi-scale periodic features, and nonlinear trend features from the time-series data. The dynamic spatial correlation features, multi-scale periodic features, and nonlinear trend features are adaptively fused to generate a comprehensive feature vector for each observation field node. Simultaneously, based on the learning results of the model, a quantitative result representing the collaborative relationship between different observation field nodes is output.

[0088] Step S3 is used to perform ecosystem synergy analysis.

[0089] A correlation model based on an adaptive spatiotemporal graph attention network is constructed, comprising at least: an adaptive spatial graph encoder (ASGE) for extracting dynamic spatial correlation features between nodes in the observation field and outputting a spatial feature vector; a seasonal capture module (SCM) for extracting multi-scale periodic features from the input time-series data and outputting a seasonal feature vector; a frequency domain decomposition module (FDD) for extracting nonlinear frequency domain features from the input time-series data and outputting a frequency domain feature vector; and a feature fusion layer for adaptively fusing the spatial feature vector, seasonal feature vector, and frequency domain feature vector to generate a comprehensive feature vector characterizing the state of nodes in the observation field and potential correlations between nodes. Through core components such as the adaptive spatial graph encoder (ASGE), seasonal capture module (SCM), and frequency domain decomposition module (FDD), the synergistic relationships between different observation fields and different ecosystems are explored, and key driving factors of ecological quality changes are identified.

[0090] S3.1: Core Structure Design of the Model

[0091] like Figure 3 The diagram shown is a schematic of the intelligent analysis (Adaptive Spatiotemporal Graph Attention Network ASTNet) process according to an embodiment of the present invention. The input is standardized time-series monitoring data, and a single input sample contains... Each observation field node The core metrics, with a time dimension length of [missing information], are [missing information]. (336 time steps, corresponding to 7 days of continuous monitoring data), input feature dimension is .

[0092] Using standardized three-dimensional monitoring data of "observation field-indicator-time" as a unified input, the system first mines the dynamic correlations in the spatial dimension through an adaptive spatial graph encoder (ASGE). Simultaneously, the seasonal capture module (SCM) and frequency domain decomposition module (FDD) extract periodic and frequency domain features in the temporal dimension, respectively. Finally, a feature fusion layer adaptively integrates the three types of features to form a comprehensive feature set that combines spatial correlation, temporal periodicity, and nonlinear trends, providing high-quality feature support for subsequent collaborative relationship mining and driving factor identification. The specific functions of each component are as follows:

[0093] 1. Adaptive Spatial Graph Encoder (ASGE): Dynamically learns time-varying spatial dependencies through a learnable node embedding matrix. The feature similarity between each observation field node is calculated, and an adaptive adjacency matrix is ​​dynamically generated through nonlinear activation and normalization. This is used to quantify the strength of time-varying spatial correlation between different observation field nodes, and the formula is: This means that the model can automatically learn and update the association patterns between different ecosystem observation fields based on input multi-factor time-series data (such as water level, vegetation, and weather). For example, during the snowmelt season, the association weight between the lake and the upstream meadow ( The adjacency matrix may increase, indicating enhanced hydrological connectivity. Then, an adaptive adjacency matrix is ​​used. Given the input features of the observation field nodes, perform at least one layer of graph convolution operation to aggregate the neighborhood features of each observation field node, learn and output a spatial feature vector. Specifically: A two-layer graph convolutional network is used to aggregate neighborhood features. The first layer has an output dimension of 64, and the second layer has an output dimension of 32. The activation function is ReLU, and the graph convolution formula is as follows: .

[0094] An adaptive adjacency matrix dynamically characterizes the temporal spatial correlation strength between nodes of different observation fields in a cold, inland drainage region (e.g., the seasonal variation of hydrological recharge correlation between a lake observation field and surrounding meadow observation fields). Each element of this matrix... The value represents the dynamic correlation strength between node i (e.g., a lake observation field) and node j (e.g., a surrounding meadow observation field) at the current moment; this is a mathematical representation of the cooperative relationship. Softmax The normalization function transforms the ReLU-activated node embedding similarity matrix into a probability distribution, making the sum of the elements in each row of the adjacency matrix equal to 1, which facilitates the quantification of the association weights between nodes. The rectified linear unit activation function introduces nonlinearity, suppresses negative values ​​in the node embedding matrix, highlights effective spatial correlation information, and avoids gradient vanishing. Node embedding matrix, dimension 1 ( The number of observation field nodes, (As an embedding dimension), it is learned through model training and contains the core features of each observation field (such as temperature and precipitation characteristics of meteorological observation fields, and water level and flow characteristics of hydrological observation fields). Node embedding matrix The transpose of the matrix is ​​used to calculate the feature similarity between nodes in different observation fields. : No. The input feature matrix of the layer graph convolution has dimensions of ( For the first Layer input feature dimension). The original standardized index characteristics of the observation field nodes. : No. The output feature matrix of the layer graph convolution has a dimension of ( For the first (Layer output feature dimension) Time output dimension The output dimension is 32, which is a spatial feature vector. . : No. The learnable weight matrix of the layer graph convolution has dimensions of . It is used to perform linear transformations on input features and mine spatial correlation features.

[0095] 2. Seasonal Capture Module (SCM): Performs one-hot encoding on the intraday time period, intraweek date, and season contained in the time information corresponding to the input data, respectively, to obtain three independent sparse encoded vectors. , , The three sparse coding vectors are concatenated to form a high-dimensional feature vector that integrates multi-scale time period information. A learnable linear transformation layer then maps the concatenated high-dimensional feature vector into a low-dimensional, dense seasonal embedding vector. As a seasonal feature vector output, it is used to explicitly characterize the multi-scale periodic evolution of alpine ecosystems.

[0096] Specifically: By embedding "day-week-season" periodic information through one-hot encoding, the seasonal evolution pattern of alpine ecosystems is explicitly modeled, with an output dimension of 32. The formula is as follows:

[0097] ;

[0098] The seasonal embedding vector, with a dimension of 32, is a dense representation of the "day-week-season" periodic information, adapted to the seasonal evolution patterns of alpine ecosystems (such as the periodic characteristics of winter permafrost freezing and summer vegetation growth). Concat The feature concatenation operation concatenates the one-hot encoded vectors of the day, week, and season according to the feature dimension, integrates multi-scale periodic information, and forms a high-dimensional sparse feature vector. One-hot encoding function converts discrete-time indices into binary vectors, such as converting weekday date index 2 (Wednesday) into a binary vector. , which explicitly represents the characteristics of a single periodic dimension. Intraday time period index, value (Corresponding to a 24-hour day) to capture the periodic changes in intraday observation indicators (such as the pattern of daytime vegetation transpiration and nighttime soil temperature and humidity stability). : Weekday date index, value (Corresponding to Monday to Sunday), capturing the periodic fluctuations of ecological indicators within the week (such as the cycle of surrounding vegetation disturbance caused by the intensity of human activities). Seasonal index, with values ​​from 0 to 3 (corresponding to spring, summer, autumn, and winter respectively), adapted to the significant seasonal differences in high-altitude inland drainage areas (such as the ecological characteristics of concentrated summer precipitation and frozen soil in winter). Learnable weight matrix, dimension 1 ( Encoding dimension for intraday time periods, For weekday date encoding dimensions, (For seasonal encoding dimension), the high-dimensional sparse one-hot encoding is mapped to a 32-dimensional dense vector. : Bias vector, with a dimension of 32, used to fine-tune the baseline value of the seasonal embedding vector, improving the flexibility of model fitting. : Feature dimension of the unique hot encoding for intraday time periods, with a value of 24, corresponding to the 24 time periods within the day. : Feature dimension of the one-hot encoding of dates within the week, with a value of 7, corresponding to the 7 days of a week. : The feature dimension of seasonal unique heat encoding, with a value of 4, corresponding to the four seasons.

[0099] 3. Frequency Domain Decomposition Module (FDD): The input standardized time-series monitoring data is transformed from a time-domain representation to a frequency-domain representation using Discrete Fourier Transform (DFT). Through analysis of the frequency-domain representation, high-frequency fluctuation components and low-frequency trend components are separated and extracted from the time-series data. Based on these components, a frequency-domain feature vector is constructed and output to enhance the capture of nonlinear features in the time-series data. The decomposition formula is as follows:

[0100] ;

[0101] Frequency domain data is the result of time domain data after discrete Fourier transform. It includes high-frequency fluctuation components (such as sudden changes in soil moisture caused by short-term extreme weather and fluctuations in vegetation cover caused by wind and sand activities) and low-frequency trend components (such as long-term vegetation degradation and the annual change trend of permafrost active layer thickness). The Discrete Fourier Transform function converts time-series monitoring data (such as soil temperature and humidity and vegetation coverage data for 7 consecutive days) into frequency domain representation, decomposes feature components of different frequencies, and enhances the capture of nonlinear features. Time-domain data, i.e., standardized time-series monitoring data, has the following dimensions: The time step corresponds to 7 days. (Indicators), recording the dynamic changes of observed indicators over time.

[0102] 4. Feature Fusion Layer: Receives spatial feature vectors output by the Adaptive Spatial Graph Encoder (ASGE). Seasonal feature vectors output by the Seasonal Capture Module (SCM) and the frequency domain feature vector output by the frequency domain decomposition module (FDD). ; By using a set of learnable fusion weight parameters, the spatial feature vectors are respectively... Seasonal feature vectors and frequency domain eigenvectors Perform a weighted linear combination and add a learnable bias term. The result of the linear combination is input into the sigmoid activation function, and spatial features, seasonal features, and frequency domain features are adaptively fused using a gating mechanism. This achieves dynamic adaptive adjustment and fusion of the contribution weights of spatial correlation features, time periodic features, and nonlinear trend features, and outputs a comprehensive feature vector. The formula is:

[0103] ;

[0104] The comprehensive feature vector, with a dimension of 64, integrates three types of features: spatial correlation, time period, and nonlinear trend, providing high-quality input for subsequent collaborative relationship mining and driving factor identification. : S-type activation function, maps the linearly combined feature values ​​to The contribution weights of the three types of features are dynamically adjusted within the interval. For example, the weight of spatial features is strengthened during the ecological synergy period, and the weight of frequency domain features is strengthened during the extreme weather period. Spatial feature fusion weight matrix, with dimension 1 Used for the spatial feature vector output by ASGE Perform a linear transformation to quantify the contribution of spatial features. Seasonal feature fusion weight matrix, dimension 1 , used for the seasonal feature vector output by SCM Perform a linear transformation to quantify the contribution of periodic characteristics. Frequency domain feature fusion weight matrix, with dimension 1 , used for the frequency domain feature vector of FDD output Perform a linear transformation to quantify the contribution of nonlinear trend characteristics. : Spatial feature vector, with a dimension of 32, is output by a 2-layer graph convolution of ASGE and contains dynamic spatial correlation information between observation field nodes. Seasonal feature vector, with a dimension of 32, is output by SCM and contains multi-dimensional periodic information of "day-week-season". : Frequency domain feature vector, with a dimension of 32, output by FDD, containing high-frequency fluctuations and low-frequency trend features of time series data. : Bias term, with a dimension of 64, is used to fine-tune the baseline value of linear combination of features to improve the representational ability of fused features.

[0105] S3.2: Analysis of Synergistic Relationships and Driving Factors

[0106] Based on the learning results of the aforementioned Adaptive Spatiotemporal Graph Attention Network (ASTNet), collaborative relationship quantification and key driving factor identification are performed:

[0107] (1) Quantitative output of collaborative relationships:

[0108] The adaptive adjacency matrix dynamically generated by the Adaptive Spatial Graph Encoder (ASGE) The output serves as the core quantitative basis for characterizing the collaborative relationships between nodes in different observation fields. Each element of this matrix... The dynamic correlation strength between observation field node i and node j under the current spatiotemporal context was directly quantified, thus explicitly quantifying the spatiotemporal collaborative relationship between different ecosystems in matrix form.

[0109] (2) Identification of key driving factors:

[0110] Based on the comprehensive feature vector output by the feature fusion layer An additional fully connected layer and a Softmax activation function are used to calculate the contribution weights of each input ecological indicator (such as precipitation, temperature, and soil moisture) to the current ecological state change. Specifically, the additional fully connected layer integrates the feature vector... The input is the number of candidate driving factors (e.g., precipitation, temperature, permafrost active layer thickness), and the output dimension is equal to the number of such factors. The model is trained to learn the contribution of each factor to the change in the target ecological indicator. These contributions are normalized to weights using a softmax function, and the factors with the highest weights are identified as key driving factors. Higher weights indicate a greater impact of the factor on changes in ecological quality, thus identifying the key driving factors.

[0111] S4: Input the integrated feature vector and / or the time-series monitoring data in the standard database into the prediction model constructed based on the time-series attention gating unit (TAGU) to make short-term, medium-term and long-term trend predictions for at least one core ecological indicator among the frozen soil active layer thickness, vegetation cover change rate and hydrological runoff, and output the prediction results with confidence intervals.

[0112] Step S4 is used to achieve long-term monitoring trend prediction.

[0113] A prediction model based on Temporal Attention Gated Units (TAGUs) is constructed to adapt to the long-term, nonlinear evolution characteristics of ecosystems in high-altitude inland drainage basins, enabling multi-period trend prediction of core indicators. For example... Figure 4 The diagram shown is a flowchart of a prediction model based on a temporal attention gating unit (TAGU) according to an embodiment of the present invention, which specifically includes the following steps:

[0114] S4.1: Model Structure and Input / Output Logic

[0115] The model structure adopts an "encoder-decoder" architecture. The encoder consists of four layers of temporal attention gating units stacked together, which are used to encode the position-encoded input sequence and extract deep features. The decoder consists of two layers of temporal attention gating units stacked together, which are used to decode based on the encoded features and generate the hidden state representation of the predicted sequence.

[0116] 1. Temporal Attention Gated Unit (TAGU): It integrates the gated recurrent unit (GRU) with the multi-head self-attention mechanism to collaboratively capture local short-term dependencies and global long-term dependencies in temporal data.

[0117] (1) First, local short-term dependencies are captured using GRU. The GRU update formula is:

[0118] ;

[0119] ;

[0120] Update the gate vector, with the same dimension as the hidden state, and its value range is... Used to control the history hidden state Retention ratio and current input The update ratio (e.g., in the alpine ecological indicators, the long-term permafrost change trend is retained, while the short-term precipitation impact is updated). : Reset the gate vector, with the same dimension as the hidden state, and its value range is... Used to control the history hidden state The degree of forgetting (e.g., weakening information about wind and sand activity interference during non-critical periods). The sigmoid activation function maps the result of a linear combination to... The interval is used to achieve gated on / off control. Update the learnable weight matrix of the gate. Dimensions ( For the hidden layer dimension, (Input feature dimension), used to analyze the current input. Perform a linear transformation; Dimensions Used for hiding historical states Perform a linear transformation. : Reset the learnable weight matrix of the gate, Dimensions Dimensions Its function is the same as the update gate weight matrix, and it adapts to the linear transformation of the current input and the historical hidden state respectively. : The input feature vector at time step , with dimension . Standardized time-series data corresponding to a certain observation indicator in a cold inland drainage region (e.g.) (Data on soil moisture and vegetation cover at any given time). The hidden state vector at time step , with dimension . Contains Temporal characteristics prior to the current moment (such as historical permafrost thickness changes and vegetation growth trends). Update the gate and reset its bias vector, with dimension 1. It is used to fine-tune the gated output and improve the model's fitting ability. : Candidate hidden state vector, dimension is By resetting the gate modulation of the historical hidden state and the linear combination of the current input, through... Activated and containing New feature information at any given moment. The hyperbolic tangent activation function maps the linear combination result to... The interval introduces nonlinearity to enhance the feature representation capability. : The learnable weight matrix of the candidate hidden states, with dimension . Used for the current input Perform a linear transformation. : The bias vector of the candidate hidden state, with dimension , used to fine-tune the output of candidate hidden states. The element-wise multiplication operator performs element-wise multiplication of vectors, enabling gating and modulation of features. : The final hidden state vector at time step , with dimension . By updating the gate, historical hidden states and candidate hidden states are fused, and key temporal features are preserved.

[0121] (2) Then through multi-head self-attention (number of heads) Focusing on key historical moments, the attention formula is: . These are query, key, and value matrices, respectively. For feature dimensions.

[0122] : Multi-head self-attention output vector, with dimension This study focuses on key historical moments (such as the permafrost thawing period and peak vegetation growth season) in high-altitude ecological time-series data. Softmax The normalization function transforms the attention score matrix into a probability distribution, making the sum of attention weights at each time step equal to 1, thus quantifying the contribution of key moments. Query matrix, dimension is ( For time step, (The feature dimension) is obtained from the hidden state transformation of TAGU and is used to query key historical features. : Key matrix, dimension The value is obtained from the hidden state transition of TAGU and is used to calculate the attention score with the query matrix. Value matrix, dimension 1 , obtained from the hidden state transition of TAGU, is the feature value corresponding to the attention score. : Key matrix The transpose of the matrix is ​​used to query the matrix. Perform matrix multiplication and calculate the similarity (attention score) between time steps. The square root of the feature dimension is used to scale the attention score, avoiding bias. Excessive head count can lead to score overflow, thus improving training stability. :Will The attention is split into 4 groups for parallel computation to capture key temporal features at different scales (such as short-term precipitation, medium-term seasonal changes, and long-term ecological evolution), and finally spliced ​​together for output.

[0123] 2. Position Encoding Layer: This layer uses trigonometric functions to encode temporal position information, embedding position information into the input temporal data. The formula is as follows:

[0124] ;

[0125] ;

[0126] The positional encoding vector is at the 1st position. The time step, the first The value of dimension (even-dimensional or odd-dimensional) is used to represent the location information of time series data and avoid the model from confusing the time order (such as distinguishing data of the same season in different years). pos: time step index, corresponding to the order in the time series data (e.g., 0 corresponds to the observation start time, 335 corresponds to the last time step of 7-day monitoring). Dimension index of position encoding, range of values For model dimension (and hidden layer dimension) Consistent). Model dimension, i.e., hidden layer dimension This is used to define the total dimension of the positional encoding, which is consistent with the hidden state dimension of TAGU. Sine and cosine functions are used to generate position codes through trigonometric functions of different frequencies, making the codes at different time steps unique. The larger the time step interval, the more significant the code differences, which is suitable for the position representation of long time series data.

[0127] 3. Output Layer: This layer maps the hidden state representation output by the decoder to the final predicted ecological metric values. Specifically, it outputs the prediction results through a two-layer fully connected network: the first fully connected layer has a dimension of... The hidden state vector output by TAGU is mapped to 128 dimensions; the second fully connected network has a dimension of... This method maps high-dimensional features to one-dimensional predicted values ​​(such as the future value of an ecological indicator). The first layer activation function is ReLU, and the second layer is a linear activation function. The rectified linear unit activation function is used in the first fully connected layer of the network. It introduces nonlinearity, strengthens key prediction features, suppresses invalid information, and avoids gradient vanishing. The linear activation function is used in the second fully connected layer of the network. It directly outputs predicted values ​​(such as future values ​​of permafrost active layer thickness and vegetation cover change rate) without introducing additional nonlinear transformations, ensuring the continuity and reasonableness of the predicted values.

[0128] S4.2: Model Training Configuration and Key Parameters

[0129] The input sequence length is 365 (1 year of time series data), and the output sequence length supports 90 (short-term), 1095 (medium-term), and 1825 (long-term), corresponding to prediction periods of 6 months, 3 years, and 5 years.

[0130] The training parameters use the AdamW optimizer, with an initial learning rate of [missing information]. The learning rate scheduling uses a cosine annealing strategy with a minimum learning rate. The training rounds are 200, and the batch size is 32.

[0131] Normalization is performed using a Layer Norm layer to normalize the input features, as shown in the formula. .

[0132] The output of the Layer Norm layer, i.e. the normalized input features, keeps the mean and variance of ecological index data (such as permafrost thickness and vegetation cover time series data) in the high-altitude inland drainage area stable, thereby improving the model training convergence speed and generalization ability. The input feature vector of the Layer Norm layer has a dimension of ( (Input feature dimension), corresponding to standardized alpine ecological time series index data. Input features The mean value is obtained by averaging all elements of the input feature, and reflects the overall level of the batch of data. Input features The variance measures the dispersion of input feature elements and reflects the fluctuation of the data. Minimum value (value) This is used to avoid division errors when the variance is 0, and to ensure the stability of the formula calculation. Learnable scaling parameters, dimensions and Consistency is used to scale normalized features while preserving useful information from the original features. Learnable offset parameters, dimensions and Consistency is used to fine-tune the shift of normalized features, improving the flexibility of model fitting.

[0133] During model training, a hybrid loss function is used to optimize the model parameters. The hybrid loss function is a weighted sum of the mean absolute error (MAE) loss and the root mean square error (RMSE) loss, and the formula is as follows:

[0134] ;

[0135] , ;

[0136] Loss: Global mixed loss value, obtained by weighted fusion of MAE and RMSE, comprehensively measuring the deviation between predicted and actual values, suitable for high-altitude ecological indicator prediction (paying attention to both overall error and penalizing extreme biases). 0.6, 0.4: Loss weighting coefficients, representing the contribution ratio of MAE loss and RMSE loss to the global loss, respectively, highlighting the constraint of MAE on overall error, while penalizing larger prediction biases (such as prediction errors caused by abrupt changes in ecological indicators due to extreme weather) through RMSE. Mean Absolute Error (MAE) measures the average absolute deviation between the predicted and actual values. It is robust to outliers and is suitable for the normal fluctuations of high-altitude ecological indicators. Root Mean Square Error (RMSE) measures the square root of the mean square deviation between the predicted and actual values. It penalizes larger deviations more significantly, preventing serious deviations in the prediction of key indicators such as the thickness of the active permafrost layer and hydrological runoff. : Number of predicted samples, corresponding to the number of predicted data entries for a certain batch of alpine ecological indicators. : No. The true values ​​of each sample, that is, the measured data of ecological indicators in high-altitude inland drainage areas (such as the measured thickness of the active permafrost layer and the rate of change in vegetation cover). : No. The predicted value of each sample, that is, the predicted result of the alpine ecological index output by the model.

[0137] S4.3: Scenario-Specific Trend Prediction

[0138] Predictive indicators include core indicators such as the thickness of the active permafrost layer, the rate of change in vegetation cover, hydrological runoff, changes in lake water levels, and the evolution of soil fertility.

[0139] Output results: Predicted values ​​and confidence intervals (confidence levels) The confidence intervals are calculated using the Monte Carlo simulation method. This involves performing 500 random sampling predictions on the parameters of the prediction model, generating 500 sets of prediction results, and then calculating the standard deviation of these results using statistical distribution. To ensure the reliability of the confidence interval, and based on the standard deviation The formula for generating prediction confidence intervals with a specific confidence level is:

[0140] ;

[0141] Confidence interval (confidence level) (), indicating the probable range of future predicted values ​​for high-altitude ecological indicators, providing an uncertain reference for ecological protection decisions. : The predicted value of a certain high-altitude ecological indicator (such as the predicted change in lake water level or the evolution of soil fertility). 1.645: The standard normal distribution quantile corresponding to the 90% confidence level, used to calculate the confidence interval boundary based on the standard deviation of the prediction results. The standard deviation of the prediction results measures the dispersion of multiple predictions and reflects the uncertainty of the model's predictions (the smaller the standard deviation, the more stable the prediction).

[0142] Scenario adaptation: For different ecosystems such as wetlands, meadows, deserts, and lakes in the "alpine inland drainage basin", a differentiated model parameter fine-tuning strategy is adopted to improve the predictive accuracy.

[0143] like Figure 5 The diagram shown is a schematic representation of a module of an ecological quality observation and positioning system for high-altitude inland drainage basins provided in an embodiment of the present invention. The system includes:

[0144] The data acquisition module 210 is used to set up a three-dimensional observation network in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field and observation fields of various ecological types. It also sets up sampling points and monitoring instruments covering hydrological, soil, meteorological and ecological elements in each observation field, and performs continuous data acquisition in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset.

[0145] Data processing module 220 is used to format and normalize the multi-ecosystem collaborative monitoring dataset to construct a structured ecological quality observation standard database for high-altitude inland drainage areas.

[0146] The intelligent analysis module 230 is used to input the time-series monitoring data from the standard database into a pre-trained correlation analysis model based on the adaptive spatiotemporal graph attention network (ASTNet). The correlation analysis model extracts the dynamic spatial correlation features between observation field nodes, the multi-scale periodic features and the nonlinear trend features in the time-series data, and adaptively fuses the dynamic spatial correlation features, multi-scale periodic features and nonlinear trend features to generate a comprehensive feature vector for each observation field node, and outputs a quantitative result representing the collaborative relationship between different observation field nodes.

[0147] The trend prediction module 240 is used to input the comprehensive feature vector and / or the time-series monitoring data in the standard database into a prediction model constructed based on the time-series attention gating unit (TAGU), to perform short-term, medium-term and long-term trend predictions on at least one core ecological indicator among the frozen soil active layer thickness, vegetation cover change rate and hydrological runoff, and output prediction results with confidence intervals.

[0148] The ecological quality observation and positioning system for high-altitude inland drainage basins provided in this embodiment of the invention can execute the ecological quality observation and positioning method for high-altitude inland drainage basins provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the method. For detailed process, please refer to the relevant operations of the aforementioned method embodiments, which will not be repeated here.

[0149] This invention provides a method and system for ecological quality observation and positioning in high-altitude inland drainage basins. By establishing a three-dimensional observation network and collecting comprehensive data on hydrology, soil, meteorology, and ecology, a standardized database is constructed. Then, using an adaptive spatiotemporal graph attention network, spatial correlation, seasonal cycles, and frequency domain trend features are deeply integrated. This not only generates comprehensive features for prediction but also directly quantifies and outputs a dynamic synergistic relationship matrix and key driving factors among ecosystems. Finally, through a time-series attention-gated unit prediction model specifically designed for high-altitude data, accurate multi-period predictions with confidence intervals for core ecological indicators are achieved. This forms an integrated technical system of comprehensive observation, deep correlation analysis, and reliable prediction, solving the problems of fragmented observations, weak correlation analysis, and inaccurate long-term predictions in existing technologies.

[0150] To verify the effectiveness of the method described in the embodiments of the present invention, comparative experiments were conducted. Figure 6The paper presents the performance evolution characteristics of the algorithm of this invention and five other algorithms, namely TimesNet (temporal modeling network), iTransformer (inverse transform temporal network), FDFSAN (feature decomposition filtering self-attention network), CNN (convolutional neural network) and GCN (graph convolutional neural network), in the task of predicting the core indicators of ecological quality in high-altitude inland drainage areas, showing the root mean square error (RMSE) as a function of the number of training iterations (10 to 100). Overall, the RMSE of all algorithms showed a monotonically decreasing trend with increasing training epochs, reflecting the continuous improvement of the model's ability to gradually fit and generalize the characteristics of high-altitude ecological time-series data. Among them, the algorithm of this invention consistently maintained the best performance: in the initial training stage (10 epochs), its RMSE was 2.8%, significantly lower than TimesNet (3.2%), iTransformer (3.5%), FDFSAN (3.8%), CNN (4.2%), and GCN (4.0%), demonstrating its ability to quickly capture multi-dimensional features of high-altitude inland drainage basins; as the number of training epochs increased, the performance advantage of the algorithm of this invention further expanded, with the RMSE dropping to 1.0% after 100 training epochs, a reduction of 41.18%, 44.44%, 52.38%, 60.00%, and 56.52% compared to the five comparison algorithms, respectively. One key achievement is attributed to the integration of an adaptive spatial graph encoder (ASGE), a seasonal capture module (SCM), and a frequency domain decomposition module (FDD), which achieves a deep fusion of spatial dynamic correlations, temporal periodic patterns, and nonlinear fluctuation characteristics. Furthermore, the prediction architecture based on a temporal attention gating unit (TAGU) effectively adapts to the long temporal dependence and strong seasonality of alpine ecological data. In contrast, CNN and GCN, lacking targeted cross-modal fusion mechanisms and long-term temporal modeling capabilities, consistently lag behind in performance. While TimesNet, iTransformer, and FDFSAN have achieved some accuracy improvements through temporal modeling optimization, they have failed to fully exploit the synergistic relationships between ecosystems in alpine inland drainage basins, thus making it difficult to surpass the comprehensive performance of the algorithm presented in this invention. This fully verifies the scientific validity and superiority of this invention in the task of predicting ecological quality in alpine inland drainage basins.

[0151] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and relevant parts can be referred to the method section.

[0152] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0153] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for monitoring and locating the ecological quality of high-altitude inland drainage basins, characterized in that, Includes the following steps: A three-dimensional observation network was established in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field, and observation fields of various ecological types. Sampling points and monitoring instruments covering hydrological, soil, meteorological, and ecological elements were set up in each observation field. Continuous data collection was carried out in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset. The multi-ecosystem collaborative monitoring dataset was formatted and normalized to construct a structured ecological quality observation standard database for high-altitude inland drainage areas. The time-series monitoring data from the standard database is input into a pre-trained association analysis model based on an adaptive spatiotemporal graph attention network. The association analysis model extracts dynamic spatial association features between observation field nodes, multi-scale periodic features, and nonlinear trend features from the time-series data. The dynamic spatial association features, multi-scale periodic features, and nonlinear trend features are then adaptively fused to generate a comprehensive feature vector for each observation field node. Simultaneously, based on the learning results of the model, a quantitative result representing the collaborative relationship between different observation field nodes is output. The adaptive spatiotemporal graph attention network includes at least: an adaptive spatial graph encoder for extracting dynamic spatial correlation features between observation field nodes and outputting a spatial feature vector; a seasonal capture module for extracting multi-scale periodic features from the input time series data and outputting a seasonal feature vector; a frequency domain decomposition module for extracting nonlinear frequency domain features from the input time series data and outputting a frequency domain feature vector; and a feature fusion layer for adaptively fusing the spatial feature vector, seasonal feature vector, and frequency domain feature vector to generate a comprehensive feature vector characterizing the state of observation field nodes and potential correlations between nodes. The comprehensive feature vector and / or the time-series monitoring data in the standard database are input into the prediction model constructed based on the time-series attention gating unit to make short-term, medium-term and long-term trend predictions for at least one core ecological indicator among the frozen soil active layer thickness, vegetation cover change rate and hydrological runoff, and output prediction results with confidence intervals. The prediction model built based on the temporal attention gating unit adopts an encoder-decoder architecture and includes at least: The location coding layer is used to embed location information into the input time-series data; The encoder, consisting of four layers of stacked temporal attention gating units, is used to encode the position-encoded input sequence and extract deep features. The decoder, consisting of two stacked layers of temporal attention gating units, is used to decode based on encoded features and generate the hidden state representation of the predicted sequence. The output layer is used to map the hidden state representation output by the decoder to the final ecological index prediction value. The temporal attention gating unit integrates a gating loop unit and a multi-head self-attention mechanism to collaboratively capture local short-term dependencies and global long-term dependencies in temporal data.

2. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, The three-dimensional observation network specifically includes: 1 comprehensive observation field, 4 evapotranspiration observation fields respectively located in alpine wetlands, meadows, deserts and around lakes, 3 meteorological observation fields, 5 hydrological observation fields, and 7 ecological observation fields respectively covering alpine wetlands, shrublands, meadows, deserts, alpine vegetation, non-vegetated areas and lake shorelines.

3. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, The adaptive spatial graph encoder calculates the feature similarity between nodes of each observation field through a learnable node embedding matrix, and dynamically generates an adaptive adjacency matrix through nonlinear activation and normalization processing to quantify the time-varying spatial correlation strength between nodes of different observation fields. Using the adaptive adjacency matrix and the input features of the observation field nodes, at least one layer of graph convolution operation is performed to aggregate the neighborhood features of each observation field node in order to learn and output the spatial feature vector.

4. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, The seasonal capture module performs one-hot encoding on the intraday time period, intraweek date and season contained in the time information corresponding to the input data to obtain three independent sparse encoding vectors. The three sparse coding vectors are concatenated to form a high-dimensional feature vector that integrates multi-scale time period information. Through a learnable linear transformation layer, the spliced ​​high-dimensional feature vector is mapped into a low-dimensional dense seasonal embedding vector, which is then output as the seasonal feature vector to explicitly characterize the multi-scale periodic evolution of alpine ecosystems.

5. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, The frequency domain decomposition module applies Discrete Fourier Transform to the input standardized time-series monitoring data, transforming it from a time-domain representation to a frequency-domain representation. By analyzing the frequency domain representation, high-frequency fluctuation components and low-frequency trend components in the time series data are separated and extracted. Based on the high-frequency fluctuation component and the low-frequency trend component, the frequency domain feature vector is constructed and output to enhance the capture of nonlinear features in the time series data.

6. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, The feature fusion layer generates the comprehensive feature vector in the following manner: The system receives the spatial feature vector output by the adaptive spatial map encoder, the seasonal feature vector output by the seasonal capture module, and the frequency domain feature vector output by the frequency domain decomposition module. The spatial feature vector, seasonal feature vector, and frequency domain feature vector are weighted linearly combined using a set of learnable fusion weight parameters, and a learnable bias term is added. The result of the linear combination is input into the sigmoid activation function to achieve dynamic adaptive adjustment and fusion of the contribution weights of spatial correlation features, time periodic features and nonlinear trend features, and output the comprehensive feature vector.

7. The method for ecological quality observation and positioning in high-altitude inland drainage basins according to claim 1, characterized in that, During the training process, the prediction model constructed based on the temporal attention gating unit uses a hybrid loss function to optimize the model parameters. The hybrid loss function is a weighted sum of the mean absolute error loss and the root mean square error loss. When making trend predictions, the Monte Carlo simulation method is used to perform multiple random sampling predictions on the prediction model. The standard deviation of the predicted values ​​is calculated based on the statistical distribution of the multiple prediction results, and a prediction confidence interval with a specific confidence level is generated accordingly.

8. A system for monitoring and locating the ecological quality of a high-altitude, cold-region inland drainage basin, characterized in that, include: The data acquisition module is used to deploy a three-dimensional observation network in the target high-altitude inland drainage area, including a comprehensive observation field, an evapotranspiration observation field, a meteorological observation field, a hydrological observation field, and observation fields of various ecological types. Within each observation field, sampling points and monitoring instruments covering hydrological, soil, meteorological, and ecological elements are deployed. Continuous data acquisition is carried out in a hierarchical and classified manner according to a preset strategy to form a multi-ecosystem collaborative monitoring dataset. The data processing module is used to format and normalize the multi-ecosystem collaborative monitoring dataset to construct a structured ecological quality observation standard database for high-altitude inland drainage areas. The intelligent analysis module is used to input time-series monitoring data from the standard database into a pre-trained correlation analysis model based on an adaptive spatiotemporal graph attention network. The correlation analysis model extracts dynamic spatial correlation features between observation field nodes, multi-scale periodic features, and nonlinear trend features from the time-series data. It then adaptively fuses these features to generate a comprehensive feature vector for each observation field node and outputs a quantitative result representing the collaborative relationship between different observation field nodes. The adaptive spatiotemporal graph attention network includes at least: an adaptive spatial graph encoder for extracting dynamic spatial correlation features between observation field nodes and outputting a spatial feature vector; a seasonal capture module for extracting multi-scale periodic features from the input time-series data and outputting a seasonal feature vector; a frequency domain decomposition module for extracting nonlinear frequency domain features from the input time-series data and outputting a frequency domain feature vector; and a feature fusion layer for adaptively fusing the spatial feature vector, seasonal feature vector, and frequency domain feature vector to generate a comprehensive feature vector representing the state of observation field nodes and potential correlations between nodes. The trend prediction module is used to input the comprehensive feature vector and / or the time-series monitoring data in the standard database into the prediction model constructed based on the time-series attention gating unit, to perform short-term, medium-term and long-term trend predictions on at least one core ecological indicator among the frozen soil active layer thickness, vegetation cover change rate and hydrological runoff, and output prediction results with confidence intervals. The prediction model built based on the temporal attention gating unit adopts an encoder-decoder architecture and includes at least: The location coding layer is used to embed location information into the input time-series data; The encoder, consisting of four layers of stacked temporal attention gating units, is used to encode the position-encoded input sequence and extract deep features. The decoder, consisting of two stacked layers of temporal attention gating units, is used to decode based on encoded features and generate the hidden state representation of the predicted sequence. The output layer is used to map the hidden state representation output by the decoder to the final ecological index prediction value. The temporal attention gating unit integrates a gating loop unit and a multi-head self-attention mechanism to collaboratively capture local short-term dependencies and global long-term dependencies in temporal data.

Images