Method for monitoring man-made interference in ecological protection red line based on multi-source data fusion

Abstract

The invention discloses a method for monitoring man-made interference in an ecological protection red line based on multi-source data fusion. The method comprises the following steps: constructing a matrix property-stress risk-wild property three-in-one dynamic monitoring index system; integrating multi-source heterogeneous data of remote sensing, territorial, forestry, meteorology and the like, and executing standardized preprocessing of coordinate unification, scale normalization and boundary clipping; adopting an entropy weight method to initially construct an ecological quality index, and combining an XGBoost algorithm to realize weight adaptive correction; performing time sequence trend robust diagnosis in combination with a Theil-Sen Median method and a Mann-Kendall method, and introducing an LSTM (Long Short Term Memory) model to realize ecological quality change prediction; and on the basis of a grid and county-level double-level evaluation framework, a current situation index, a historical trend and prediction information are fused, and a man-made interference grading early warning result with a space traceability capability is generated. According to the method, the accuracy of human interference identification, the reliability of dynamic tracking and the practicability of early warning response are improved, and quantifiable, traceable and operable technical support is provided for fine supervision and adaptive governance of ecological protection red lines.

Description

Technical Field

[0001] This invention relates to the field of ecological and environmental protection monitoring technology, and in particular to a method for monitoring human interference within ecological protection red lines based on multi-source data fusion. Background Technology

[0002] Currently, both domestically and internationally, a comprehensive monitoring system has been established in the field of ecological and environmental monitoring, primarily based on remote sensing technology and combined with ground observation and model simulation. Significant progress has been made in areas such as forest cover change, soil erosion, and biodiversity assessment. However, directly applying existing technologies to the ecological protection red line—a special area with strict control attributes—exposes a systemic deficiency that is severely disconnected from the needs of red line supervision.

[0003] Existing technologies generally employ evaluation index systems applicable to large-scale ecological functional zones or general ecosystems, lacking targeted designs for the core objective of ecological protection red lines: "protecting the authenticity and integrity of natural ecosystems." These general indicators struggle to effectively deconstruct the coupling relationship between human activity disturbances and natural ecological processes, particularly exhibiting weak ability to capture typical human disturbance signals such as abnormal nighttime light growth, sudden local changes in population density, and gradual alterations in land use intensity. Consequently, monitoring results fail to accurately reflect the core regulatory requirements within the red line area: "whether human activities have crossed boundaries and whether ecological functions have been damaged."

[0004] Significant technological gaps exist in the multi-source data fusion process: Heterogeneous data such as remote sensing images, land surveys, meteorological observations, and socio-economic data differ significantly in coordinate systems, spatial resolution, and temporal references. Existing methods often employ simple overlay or local correction processing, lacking a standardized preprocessing mechanism covering coordinate transformation, scale normalization, boundary clipping, and quality verification. This leads to systematic errors such as spatial misalignment and scale distortion during data fusion, severely weakening the reliability and spatiotemporal comparability of the evaluation results.

[0005] Traditional ecological quality assessment methods often employ expert scoring or the analytic hierarchy process (AHP) for indicator weight allocation, which is highly subjective and difficult to adapt to the differences in ecological characteristics across different regions. While some methods utilize objective weighting methods such as entropy weighting, they calculate weights solely based on data variation characteristics, failing to consider the spatial heterogeneity of ecological sensitivity across regions. This results in the inadequate strengthening of indicator weights in highly ecologically sensitive areas, reducing the accuracy of monitoring methods in identifying ecologically vulnerable zones. Furthermore, existing weighting systems lack dynamic adjustment mechanisms, cannot adaptively optimize based on historical ecological response patterns, and struggle to accurately capture ecological anomalies caused by human interference.

[0006] The aforementioned deficiencies make it difficult for existing technologies to support the refined regulatory requirements of ecological protection red lines, which include "accurately identifying sources of disturbance, dynamically tracking evolution processes, and scientifically warning of potential risks." There is an urgent need to build a monitoring method with a highly targeted indicator system, rigorous and standardized data integration, an objective and dynamic assessment process, and management-friendly output results, so as to provide a solid technical foundation for the scientific management and adaptive governance of ecological protection red lines. Summary of the Invention

[0007] The purpose of this invention is to provide a method for monitoring human interference within ecological protection red lines based on multi-source data fusion.

[0008] The problem this invention aims to solve is that existing ecological protection red line monitoring technologies suffer from defects such as the generalization of indicator systems, inaccuracy in multi-source data fusion, subjective and static assessment processes, weak trend diagnosis and prediction capabilities, and a disconnect between monitoring results and administrative management. These defects make it difficult to accurately identify hidden human interference, dynamically track ecological degradation processes, and achieve forward-looking early warnings, thus hindering the effectiveness of refined supervision and adaptive governance of red lines.

[0009] The technical solution adopted for the monitoring method of human interference within the ecological protection red line based on multi-source data fusion is as follows: S1: Construct a three-dimensional dynamic monitoring index system for ecological quality in ecological protection red line areas, which integrates matrix, stress risk, and wildness. The matrix dimension characterizes the structure and functional basis of the ecosystem, the stress risk dimension focuses on the comprehensive impact of human activities and natural stresses, and the wildness dimension reflects the integrity and authenticity of natural ecological processes. S2: Collect multi-source heterogeneous datasets that strictly correspond to each indicator in the aforementioned indicator system, covering hyperspectral / multispectral remote sensing images, time-series optical satellite data, basic geographic information, results of the third national land survey and annual change survey, comprehensive monitoring data of forestry and grassland ecology, environmental and meteorological monitoring data from environmental protection departments, and socio-economic statistical data; S3: Perform a standardized preprocessing procedure on the multi-source heterogeneous dataset, including uniformly projecting and transforming all data to the CGCS2000 geographic coordinate system, resampling to a standard spatial resolution of 1km×1km, and performing masking and cropping according to the vector boundary of the ecological protection red line to generate an analysis dataset with completely consistent spatial range, coordinate system and scale. S4: Based on the entropy weight method, information entropy is calculated and objective weights are assigned to the preprocessed indicator data. A basic ecological quality index reflecting the regional ecological baseline state is constructed by weighted synthesis. S5: Using the original indicator data and initial weight parameters involved in the calculation of the basic ecological quality index as input features, the XGBoost machine learning algorithm is introduced to construct a weight optimization model. Through iterative training of the model, the adaptive correction of the weight threshold is achieved, and the corrected ecological quality index that better fits the actual regional ecological process is output. S6: By combining the Theil-Sen Median trend slope estimation method and the Mann-Kendall nonparametric significance test, the slope of change of the modified ecological quality index over many years is calculated and the statistical significance is determined, so as to complete the quantitative diagnosis of the dynamic change trend of ecological quality. S7: Construct a long short-term memory neural network time series prediction model, using the modified ecological quality index sequence of historical periods as training input, and perform time series extrapolation and trend simulation of the ecological quality index for the target prediction period. S8: Establish a two-tiered assessment framework consisting of 1km×1km gridded monitoring units and county-level administrative units. At the grid unit scale, calculate characteristic parameters such as ecological quality index, trend slope, and prediction bias. At the county-level unit scale, generate comprehensive assessment results through area-weighted aggregation. Comprehensively correct the ecological quality index, trend diagnosis conclusions, and prediction values. Conduct comprehensive ecological quality assessment of ecological protection red line areas according to preset threshold rules, and generate early warning results for the intensity of human interference.

[0010] Furthermore, the S1-based dynamic monitoring index system for ecological quality, which integrates matrix-stress risk-wildness characteristics and addresses the ecological baseline state and anthropogenic disturbance response features of ecological protection redline areas, also includes: The matrix dimension was characterized by normalized vegetation index, air quality index, carbon storage, leaf area index, water quality index, and aboveground biomass. The stress risk dimension uses annual rainfall, topographic slope, nighttime light index, population density, and land use intensity index as characterization parameters. Wildness dimension was characterized by vegetation cover, net primary productivity, landscape sprawl index, and Shannon diversity index. The three types of indicators are logically coupled: matrix indicators provide ecological quality benchmarks, human activity-related parameters in stress risk indicators serve as direct input variables for human disturbances, including nighttime light index, population density, and land use intensity index, and wildness indicators serve as output feedback variables for the ecosystem's response to disturbances.

[0011] Furthermore, the collection of multi-source heterogeneous datasets in S2 that strictly correspond to each indicator in the indicator system also includes: For matrix-related indicators, data sources include multispectral and hyperspectral remote sensing image sequences suitable for vegetation parameter inversion, regional air quality and water quality monitoring reports regularly released by environmental monitoring departments, and basic biomass remote sensing inversion data required for carbon sink assessment. For the stress risk dimension indicators, time-series optical satellite image sequences with nighttime light detection capabilities were collected from data sources, population spatial distribution statistics released by authoritative statistical departments were retrieved, annual land use change vector results approved by natural resources authorities were obtained, and rainfall observation records and digital elevation model data from basic geographic information provided by meteorological authorities were collected simultaneously. For the wildness dimension indicators, long-term remote sensing images covering the entire monitoring period were collected from data sources and integrated with thematic data on forest and grassland resource inventory and ecological monitoring released by forestry authorities; All collected data must meet three core requirements: in terms of time attributes, the data must cover the complete time series required for ecological monitoring and assessment and maintain consistency in the time benchmarks of each data source; in terms of spatial attributes, the data must fully cover the entire area of ​​the target ecological protection red line and include necessary buffer zones; and in terms of data authority, standardized and quality-verified operational data products issued by national or provincial competent authorities should be given priority. Based on the vector boundary of the ecological protection red line, the collected data are initially classified and cataloged according to spatial location and time label, forming a structured data set that strictly corresponds to the three-dimensional indicators of matrix-stress risk-wildness.

[0012] Furthermore, the standardization preprocessing procedure for the multi-source heterogeneous dataset in step S3 also includes: Identify the original coordinate system type of each data source, and convert all raster images and vector boundary data to the CGCS2000 National Geodetic Coordinate System according to the coordinate transformation specifications promulgated by the national surveying and mapping geographic information authority. To address the differences in the original resolution of raster data from different sources, and based on the unified spatial scale standard preset by the monitoring task, continuous ecological parameter data are smoothed and resampled using bilinear interpolation, while categorical attribute data are preserved using the nearest neighbor method to maintain the integrity of category boundaries. Using the ecological protection red line vector boundary approved by the natural resources authority as a mask template, the intersection and cropping operation is performed on all resampled raster data through raster-vector overlay operation, retaining only the effective pixels inside the red line boundary, and simultaneously removing redundant data and invalid value areas outside the boundary, forming a data subset that perfectly matches the spatial range of the monitoring target. The transformed dataset undergoes triple verification, including coordinate system consistency verification, resolution compliance verification, and mask boundary matching verification. A backtracking correction process is initiated for data with abnormal verification results.

[0013] Furthermore, S4 also includes: The preprocessed indicator data are subjected to indicator attribute discrimination and standardization: positive and negative indicators are clearly distinguished according to the ecological significance of the indicators. The increase of positive indicators indicates the improvement of ecological quality, while the increase of negative indicator values ​​indicates the decline of ecological quality. The range standardization method is used to perform dimensionless transformation to eliminate the influence of dimensional differences and orders of magnitude, and generate a standardized indicator data matrix. The proportion distribution sequence of each indicator in all spatial evaluation units is calculated based on the standardized data matrix, and the information entropy value of each indicator is calculated accordingly. The smaller the information entropy value, the higher the degree of dispersion of the indicator in spatial distribution and the richer the ecological state discrimination information contained therein. The difference coefficient of each indicator is calculated based on the information entropy value, defined as 1 minus the information entropy value, and then the difference coefficient is converted into the objective weight coefficient of each indicator through normalization. Weighting formula based on ecological sensitivity Adjustments were made to the objective weighting coefficients, among which... Here, S represents the weights of the original indicators calculated using the entropy weight method, and S is the ecological sensitivity index calculated based on topography, vegetation type, biodiversity, etc. These are ecological sensitivity thresholds determined through historical data statistics. The sensitivity adjustment coefficient, determined based on regional ecological characteristics, ranges from 0.5 to 2.0. The weights of the indicators are weighted by ecological sensitivity. The ecological sensitivity index L is calculated using a multi-factor weighted superposition method. ,in The comprehensive weight of the i-th indicator is... Assign a sensitivity level to the i-th indicator using a five-level grading method (insensitive, slightly sensitive, moderately sensitive, highly sensitive, extremely sensitive), and assign grading values ​​according to the magnitude of the indicator value; The standardized values ​​of each indicator are linearly weighted and synthesized with the adjusted objective weight coefficients to generate a basic ecological quality index with continuous values. This index quantitatively represents the comprehensive ecological baseline status of the ecological protection redline area at the current monitoring time point and serves as the initial input benchmark for subsequent machine learning adaptive correction.

[0014] Furthermore, S5 introduces the XGBoost machine learning algorithm to construct a weight optimization model, including: Standardized indicator data from multiple time-series nodes within the historical monitoring period are selected as input feature vectors, and the comprehensive ecological quality assessment level, which has been jointly verified by multiple departments or supported by long-term ground verification, is used as the target variable label to form training sample pairs with strict correspondence between features and labels. Configure the XGBoost regression model framework, match the input feature dimensions with the continuous features of the target variable, and iteratively build the decision tree ensemble through the gradient boosting mechanism: each iteration focuses on minimizing the prediction residual of the preceding model, learning the nonlinear correlation between the features of each indicator and the comprehensive state of ecological quality, and incorporating regularization constraints during training to suppress the risk of overfitting. Gain-type importance scores for each indicator feature are extracted from the convergent model. The importance score is based on the total amount of objective function optimization brought about by the feature splitting in all decision trees. The score value is weighted and fused with the initial weights obtained by the entropy weight method. The fusion coefficient is dynamically determined according to the performance of the model validation set. After normalization, a corrected set of indicator weight coefficients is generated. The corrected weighting coefficients are applied to the standardized indicator data of the current monitoring period, and the corrected ecological quality index is generated through linear weighted synthesis.

[0015] Furthermore, the combined application of the Theil-Sen Median trend slope estimation method and the Mann-Kendall nonparametric significance test in S6 also includes: For the multi-year time series of the modified ecological quality index of each spatial evaluation unit, the Theil-SenMedian method is used to calculate the trend slope: traverse all non-repeating time point pairs in the time series, calculate the set of slope values ​​of the index change between each pair of time points, and take the median of the set as the robust slope estimate of the ecological quality change of the unit. Perform Mann-Kendall nonparametric tests on the same time series simultaneously: construct standardized test statistics based on the rank relationship of the series data, determine the direction of change by the sign of the statistic, and determine the statistical significance of the trend by comparing its absolute value with the theoretical distribution critical value; The slope sign obtained by Theil-Sen Median is used as the basis for determining the direction of change, and the significance conclusion of the Mann-Kendall test is used as the basis for determining the reliability of the trend. Qualitative trend diagnosis conclusions are generated according to the preset logical rules. The output is a structured dataset containing trend direction indicators, significance status indicators, and integrated diagnostic conclusions.

[0016] Furthermore, the construction of the long short-term memory neural network time-series prediction model in S7 includes: For each spatial evaluation unit, the modified ecological quality index continuously acquired within the historical monitoring period is organized into a one-dimensional time series in chronological order, and divided into a model training set and an independent validation set according to the principle of time continuity. A long short-term memory neural network architecture is constructed, which includes an input gate, a forget gate, an output gate, and a cell state unit. The input gate filters the effective information at the current moment, the forget gate selectively removes redundant historical memories, and the output gate integrates the current state and long-term memory to generate prediction basis. The time series is constructed into supervised learning sample pairs of historical window input and next time step target value using a sliding window method. The gradient optimization algorithm is used to iteratively train the model parameters. A validation set is introduced to monitor the training process. The training is automatically terminated when the validation error does not improve significantly for several consecutive rounds. Input the latest monitoring cycle's corrected ecological quality index time series segment into the trained model, and the model will generate the ecological quality index prediction value for each node within the target prediction period by time step by time. The predicted trend direction is compared with the historical trend diagnosis conclusions obtained from S6. If a significant contradiction is found, a manual review mechanism is triggered. After the verification is passed, the structured prediction results are output.

[0017] Furthermore, S8, which involves conducting a comprehensive ecological quality assessment of the ecological protection red line area based on preset threshold rules and generating a graded early warning result for the intensity of human interference, also includes: The basic monitoring grid units are formed by dividing the area into regular geographical grids. At the same time, the administrative management units are defined by the county-level administrative division boundaries approved by the natural resources authorities. A mapping table of grid units and county-level units is established through spatial topology operations to clarify the county-level units to which each grid unit belongs and the proportion of its effective analysis area within that unit. For each grid cell, the modified ecological quality index output by S5, the trend diagnosis conclusion output by S6, and the index sequence characteristics of the predicted period output by S7 are integrated to form a three-dimensional feature parameter set of current status level, historical evolution, and future trend. The area-weighted average method was used to generate the county-level comprehensive index for the modified ecological quality index; the spatial proportion statistical method was used to generate the county-level comprehensive trend judgment level for the trend diagnosis conclusion; and the sequence direction consistency analysis method was used to generate the county-level predicted trend judgment result for the predicted trend characteristics. The system has a pre-defined judgment logic that combines the county-level comprehensive index, the trend comprehensive judgment level, and the predicted trend judgment result. The rule base covers multiple combination logic scenarios. Based on the comprehensive assessment results and in accordance with the preset interference intensity mapping rules, a list of early warnings for human interference intensity at the county-level administrative unit scale and a thematic map are generated. Simultaneously, the early warning level is back-linked to the set of grid units that constitute the county-level unit, and the grid areas that trigger medium- and high-level early warnings are spatially highlighted and the list is exported.

[0018] The beneficial effects of this invention are: by coupling the matrix-stress risk-wildness three-in-one indicator system with multi-source heterogeneous data, a complete logical closed loop is constructed from ecological baseline characterization, human interference input identification to ecosystem response feedback, so that the monitoring process closely matches the core control objective of "protecting the authenticity and integrity of the natural ecology by the ecological protection red line".

[0019] At the index construction level, the entropy weight method objectively quantifies the information content of each parameter based on the information entropy of the indicator, so that the weight allocation is based on the inherent variation characteristics of the data. The XGBoost algorithm embeds the regional historical ecological response pattern into the weight correction process through the dynamic fusion of feature importance score and initial weight. The synergy of the two not only eliminates the arbitrariness of expert subjective weighting, but also strengthens the index's ability to sensitively capture regional-specific human interference signals.

[0020] At the level of weight optimization and ecological sensitivity adaptation, an ecological sensitivity weight adjustment formula is introduced. This formula can dynamically adjust the indicator weights according to the degree of regional ecological sensitivity, so that the monitoring indicators in highly ecologically sensitive areas can be reasonably strengthened, thereby improving the accuracy of identifying human interference in ecologically fragile areas. The ecological sensitivity index is calculated using a multi-factor weighted superposition method. This formula comprehensively considers the superposition effects of topographic factors, ecological environment factors, hydrological factors, and land use factors, ensuring the comprehensiveness and scientific nature of ecological sensitivity assessment. Through the synergistic effect of the above formula system, the weight system has both an objective data-driven basis and incorporates regional ecological sensitivity characteristics, reserving technical interfaces for dynamic optimization. This effectively solves the prominent problems of strong subjectivity and poor regional adaptability in the weight allocation of existing technologies.

[0021] At the dynamic analysis level, the combined application of the Theil-Sen Median and Mann-Kendall tests robustly identifies the true trend of ecological quality evolution over time using nonparametric statistical methods, effectively removing interference from natural fluctuations and locking in signals of continuous degradation caused by human activities. The LSTM prediction model learns the laws of ecological evolution based on historical sequences, enabling reasonable extrapolation of future changes. The combination of the two forms a dynamic assessment dual-engine of historical diagnosis and future prediction. The grid and county-level dual-level assessment framework organically integrates the precise positioning capabilities at the micro scale with the administrative management needs at the macro scale. Through the logical coupling of multi-dimensional parameters, it generates graded early warning results for human interference with spatial source tracing capabilities. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for monitoring human interference within ecological protection red lines based on multi-source data fusion. Detailed Implementation

[0023] The present invention will be further described clearly and completely below, but the scope of protection of the present invention is not limited thereto.

[0024] 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, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Example 1 The technical solution adopted for the monitoring method of human interference within the ecological protection red line based on multi-source data fusion is as follows: S1: Construct a three-dimensional dynamic monitoring index system for ecological quality in ecological protection red line areas, which integrates matrix, stress risk, and wildness. The matrix dimension characterizes the structure and functional basis of the ecosystem, the stress risk dimension focuses on the comprehensive impact of human activities and natural stresses, and the wildness dimension reflects the integrity and authenticity of natural ecological processes. S2: Collect multi-source heterogeneous datasets that strictly correspond to each indicator in the aforementioned indicator system, covering hyperspectral / multispectral remote sensing images, time-series optical satellite data, basic geographic information, results of the third national land survey and annual change survey, comprehensive monitoring data of forestry and grassland ecology, environmental and meteorological monitoring data from environmental protection departments, and socio-economic statistical data; S3: Perform a standardized preprocessing procedure on the multi-source heterogeneous dataset, including uniformly projecting and transforming all data to the CGCS2000 geographic coordinate system, resampling to a standard spatial resolution of 1km×1km, and performing masking and cropping according to the vector boundary of the ecological protection red line to generate an analysis dataset with completely consistent spatial range, coordinate system and scale. S4: Based on the entropy weight method, information entropy is calculated and objective weights are assigned to the preprocessed indicator data. A basic ecological quality index reflecting the regional ecological baseline state is constructed by weighted synthesis. S5: Using the original indicator data and initial weight parameters involved in the calculation of the basic ecological quality index as input features, the XGBoost machine learning algorithm is introduced to construct a weight optimization model. Through iterative training of the model, the adaptive correction of the weight threshold is achieved, and the corrected ecological quality index that better fits the actual regional ecological process is output. S6: By combining the Theil-Sen Median trend slope estimation method and the Mann-Kendall nonparametric significance test, the slope of change of the modified ecological quality index over many years is calculated and the statistical significance is determined, so as to complete the quantitative diagnosis of the dynamic change trend of ecological quality. S7: Construct a long short-term memory neural network time series prediction model, using the modified ecological quality index sequence of historical periods as training input, and perform time series extrapolation and trend simulation of the ecological quality index for the target prediction period. S8: Establish a two-tiered assessment framework consisting of 1km×1km gridded monitoring units and county-level administrative units. At the grid unit scale, calculate characteristic parameters such as ecological quality index, trend slope, and prediction bias. At the county-level unit scale, generate comprehensive assessment results through area-weighted aggregation. Comprehensively correct the ecological quality index, trend diagnosis conclusions, and prediction values. Conduct comprehensive ecological quality assessment of ecological protection red line areas according to preset threshold rules, and generate early warning results for the intensity of human interference.

[0026] refer to Figure 1 The diagram shows a flowchart of a method for monitoring human interference within the ecological protection red line based on multi-source data fusion.

[0027] Furthermore, the S1-based dynamic monitoring index system for ecological quality, which integrates matrix-stress risk-wildness characteristics and addresses the ecological baseline state and anthropogenic disturbance response features of ecological protection redline areas, also includes: Based on the core management objectives and ecological function positioning of the ecological protection red line, the monitoring indicators are divided into three logically related dimensions: The matrix dimension focuses on the underlying structure and basic service functions of the ecosystem, selecting Normalized Difference Vegetation Index (NDVI), Air Quality Index (AQI), Carbon Storage, Leaf Area Index (LAI), Water Quality Index (DQI), and Aboveground Biomass as characterization parameters to quantify vegetation cover, air quality, carbon sequestration capacity, photosynthetic activity, water cleanliness, and biomass accumulation. NDI data is sourced from NASA MOD13A3 V6.1 monthly composite products, using the annual maximum value synthesis method to calculate the annual maximum NDVI value pixel-by-pixel. AQI data is sourced from the 1km resolution product of the National Qinghai-Tibet Plateau Data Center. Carbon Storage data is sourced from the InVEST carbon storage module. LAI data is sourced from NASA MCD15A2H V6.1, using the time-series maximum value synthesis method to extract the maximum value from May to September of the growing season. Water Quality Index data is sourced from monthly high-resolution water quality data. Aboveground biomass data is sourced from the GEE platform's 30-meter resolution annual China forest aboveground biomass dataset.

[0028] The stress risk dimension focuses on the comprehensive impact of external stressors on the ecosystem, selecting annual rainfall, topographic slope, nighttime light index, population density, and land use intensity index as characterization parameters. Among them, the nighttime light index and population density directly map the spatial distribution and intensity of human activities, the land use intensity index reflects the degree of change in land development and use patterns, and annual rainfall and topographic slope are used to correct for the interference of regional natural background differences on ecological responses. The annual rainfall data is sourced from the National Tibetan Plateau Data Center at a resolution of 0.0083333° (approximately 1 km); the topographic slope data is sourced from the ASTERGDEM 30-meter resolution digital elevation model; the nighttime light index data is sourced from global NPP-VIIRS-like long-term nighttime illumination data; the population density data is sourced from the ORNL LandScan 1km resolution product; and the land use intensity index data is sourced from annual land cover data.

[0029] The Wildness dimension focuses on the natural succession status and authenticity of ecosystems, selecting vegetation cover, net primary productivity, landscape sprawl index, and Shannon diversity index as characterization parameters to assess vegetation's natural recovery capacity, ecosystem energy fixation efficiency, landscape patch connectivity, and biodiversity maintenance levels. The systematic decline in the values ​​of this dimension's indicators can serve as an indirect chain of evidence for the sustained effects of anthropogenic disturbances. Vegetation cover data is sourced from the 250-meter resolution product of the National Tibetan Plateau Scientific Data Center; net primary productivity data is sourced from the NASA MOD17A3HGFv061 dataset; and landscape sprawl index and Shannon diversity index data are sourced from land use / cover classification data.

[0030] The matrix indicators provide a benchmark for ecological quality. Anthropogenic parameters in the stress risk indicators, such as nighttime light index, population density, and land use intensity index, serve as direct input variables for anthropogenic disturbances. Wildness indicators serve as output feedback variables for the ecosystem's response to disturbances. The three form a closed-loop monitoring logic chain of baseline state - disturbance input - system response.

[0031] Furthermore, the collection of multi-source heterogeneous datasets in S2 that strictly correspond to each indicator in the indicator system also includes: Establish an indicator-data source mapping table to clarify the data type, data attributes, and authoritative source channels required for each monitoring indicator: For matrix-related dimensional indicators, multispectral and hyperspectral remote sensing image sequences suitable for vegetation parameter inversion, regional air quality and water quality monitoring reports regularly released by environmental monitoring departments, and biomass remote sensing inversion basic data required for carbon sink assessment are collected from the aforementioned data sources.

[0032] For the stress risk dimension indicators, time-series optical satellite image sequences with nighttime light detection capabilities are collected from the data sources to characterize the intensity of human activities. Population spatial distribution statistics released by the statistical department are retrieved, and annual change vector results of land use approved by the natural resources department are obtained. Simultaneously, rainfall observation records and digital elevation model data from basic geographic information provided by the meteorological department are collected.

[0033] For the wildness dimension indicators, long-term remote sensing images covering the entire monitoring period are collected from the data sources to calculate vegetation dynamic parameters, and thematic data on forest and grassland resource inventory and ecological monitoring released by the forestry authorities are integrated to support the analysis of biodiversity-related indicators.

[0034] Data collection quality control must be implemented. All collected data must meet three core requirements: in terms of time attributes, the data must cover the complete time series required for ecological monitoring and assessment, and maintain consistency in the time base of each data source; in terms of spatial attributes, the data must fully cover the entire target ecological protection red line area, including necessary buffer zones, using a 1km×1km grid as the basic assessment unit, ensuring that the data coverage includes the ecological protection red line and the surrounding 5km buffer zone; in terms of data authority, standardized and quality-verified operational data products issued by national or provincial competent authorities should be given priority; and data element information, including collection time, spatial range, data version, and issuing agency, must be structurally registered and traceable.

[0035] Constructing an initial data association index: Based on the vector boundary range of the ecological protection red line, the collected data of various types are initially classified and cataloged according to spatial location and time label. Using ArcGIS's mask extraction tool, various types of data are cropped to the range of the ecological protection red line. Data of different resolutions are resampled and unified to a spatial resolution of 1km×1km to form a structured data set that strictly corresponds to the three dimensions of matrix-stress risk-wildness indicators. This provides a clear data input framework for subsequent standardized preprocessing, ensuring that multi-source data are accurately anchored to the monitoring targets at the logical level, and avoiding data redundancy or missing key indicator data.

[0036] Furthermore, the standardization preprocessing procedure for the multi-source heterogeneous dataset in step S3 also includes: Implement spatial coordinate system standardization: Identify the original coordinate system type of each data source, such as WGS84, Beijing 54, etc., and according to the coordinate transformation specifications promulgated by the national surveying and mapping geographic information authority, use the standard seven-parameter or grid correction method to accurately transform all raster images and vector boundary data to the CGCS2000 national geodetic coordinate system, eliminating spatial position offsets caused by coordinate datum differences; for national data, use the nationally published seven-parameter transformation model, and for provincial data, use the CGCS2000 grid correction file provided by the provincial surveying and mapping geographic information bureau; for high-resolution images, use the rational polynomial coefficient model combined with ASTER GDEM for orthorectification, and use control points extracted from 1m images and road / shoreline corner points for geometric fine correction.

[0037] Spatial resolution normalization is performed: Addressing the differences in original resolution between raster data from different sources, a unified spatial scale standard is pre-defined for the monitoring task. For continuous ecological parameter data such as vegetation index and temperature, bilinear interpolation is used for smooth resampling, calculating the weighted average of the four neighboring pixels surrounding the target pixel. For categorical attribute data such as land use type, the nearest neighbor method is used to preserve the integrity of category boundaries, directly assigning the target pixel the value of the nearest original pixel. This ensures that all raster data pixels are strictly aligned in scale and that attribute information distortion is minimized.

[0038] Precise spatial constraint processing is carried out: using the vector boundary of the ecological protection red line approved by the natural resources authority as a mask template, a topological check is performed on the vector boundary to repair any possible gaps, overlaps, or other errors, ensuring that the boundary is closed and free of self-intersections; through raster-vector overlay operations, an intersection clipping operation is performed on all resampled raster data, retaining only valid pixels inside the red line boundary. The center point method is used to determine whether a pixel is retained if its center point is inside the red line boundary, otherwise it is discarded; simultaneously, redundant data and invalid value areas outside the boundary are removed to form a data subset that perfectly matches the spatial range of the monitoring target.

[0039] A preprocessing quality verification mechanism is established: the transformed dataset undergoes triple verification—coordinate system consistency verification (e.g., back-calculation verification of random sampling points), resolution compliance verification (e.g., statistical analysis of pixel size), and mask boundary fit verification (e.g., comparison of red line vectors overlaid with cropped raster edges). A backtracking correction process is initiated for verification anomalies to ensure that the output analysis dataset achieves strict uniformity in three dimensions: spatial reference, scale granularity, and range boundary. This lays an unambiguous spatial data foundation for subsequent multi-indicator collaborative calculations.

[0040] Furthermore, S4 also includes: The preprocessed indicator data are subjected to indicator attribute discrimination and standardization: positive and negative indicators are clearly distinguished according to their ecological significance. An increase in positive indicators indicates an improvement in ecological quality, while an increase in negative indicator values ​​indicates a decline in ecological quality. The range standardization method is used to perform dimensionless transformation to eliminate the influence of dimensional differences and orders of magnitude, and a standardized indicator data matrix is ​​generated.

[0041] Based on the standardized data matrix, the weight distribution sequence of each indicator in all spatial evaluation units is calculated, and the information entropy value of each indicator is calculated accordingly. The smaller the information entropy value, the higher the degree of dispersion of the indicator in spatial distribution and the richer the ecological state discrimination information contained therein.

[0042] The difference coefficient of each indicator is calculated based on the information entropy value, which is defined as 1 minus the information entropy value. The difference coefficient is then converted into the objective weight coefficient of each indicator through normalization. This weight allocation process is driven entirely by the inherent variation characteristics of the data, effectively avoiding the subjective arbitrariness of expert experience in weighting.

[0043] Weighting formula based on ecological sensitivity Adjustments were made to the objective weighting coefficients, among which... Here, S represents the weights of the original indicators calculated using the entropy weight method, and S is the ecological sensitivity index calculated based on topography, vegetation type, biodiversity, etc. These are ecological sensitivity thresholds determined through historical data statistics. The sensitivity adjustment coefficient, determined based on regional ecological characteristics, ranges from 0.5 to 2.0. The weights of the indicators are calculated by weighting them according to their ecological sensitivity.

[0044] The ecological sensitivity index L is calculated using a multi-factor weighted superposition method. ,in The comprehensive weight of the i-th indicator is... Assign a sensitivity level to the i-th indicator using a five-level grading method (insensitive, slightly sensitive, moderately sensitive, highly sensitive, extremely sensitive), and assign grading values ​​according to the magnitude of the indicator value; The factors include topographic factors, ecological environment factors, hydrological factors, and land use factors; topographic factors include elevation, slope, aspect, and topographic relief; ecological environment factors include NDVI (Normalized Difference Vegetation Index), NPP (Net Primary Productivity), and vegetation cover; hydrological factors include water buffer zones and distance from water bodies; and land use factors include land use type and soil properties. The sensitivity level is not sensitive, with a value range of [0,2); the sensitivity level is slightly sensitive, with a value range of [2,4); the sensitivity level is moderately sensitive, with a value range of [4,6); the sensitivity level is highly sensitive, with a value range of [6,8); and the sensitivity level is not extremely sensitive, with a value range of [8,10]. The comprehensive weight of the i-th indicator Using the analytic hierarchy process (AHP), expert scores are used to construct a judgment matrix and calculate the weights of each indicator.

[0045] The standardized values ​​of each indicator are linearly weighted and synthesized with their adjusted objective weight coefficients to generate a basic ecological quality index with continuous values. This index quantitatively represents the comprehensive ecological baseline status of the ecological protection red line area at the current monitoring time point and serves as the initial input benchmark for the subsequent machine learning adaptive correction process. This ensures that the weight system has both an objective data-driven basis and reserves technical interfaces for dynamic optimization.

[0046] Furthermore, S5 introduces the XGBoost machine learning algorithm to construct a weight optimization model, including: Constructing a supervised learning training sample set: Standardized indicator data from multiple time-series nodes within a historical monitoring period are selected as input feature vectors. Simultaneously, the comprehensive ecological quality assessment level, jointly verified by multiple departments such as natural resources and ecological environment authorities or supported by long-term ground verification, is used as the target variable label, forming training sample pairs with strict correspondence between features and labels. Spatially matching the 1km×1km grid data of each time-series node with the ecological protection red line vector boundary ensures that each grid unit has a corresponding indicator value at each time-series node. The data is divided into training set, validation set, and test set in an 8:1:1 ratio.

[0047] The XGBoost regression model framework is configured to match the input feature dimensions with the continuous features of the target variable. A decision tree ensemble is iteratively built through a gradient boosting mechanism: each iteration focuses on minimizing the prediction residuals of the preceding model, learning the nonlinear correlation between the features of each indicator and the overall state of ecological quality, and incorporating regularization constraints during training to suppress the risk of overfitting; the learning rate is set to 0.1, the maximum tree depth is 6, the sampling ratio is 0.8, the feature sampling ratio is 0.8, the minimum loss reduction is 0.1, the L2 regularization coefficient is 1, and the minimum weight of the child node is 1.

[0048] Gain-type importance scores for each indicator feature are extracted from the convergent model. The importance score is based on the total amount of objective function optimization brought about by the feature when splitting in all decision trees. This score objectively quantifies the contribution intensity of each indicator to the actual response of regional ecological quality. The higher the value, the greater the contribution of the indicator to the optimization of the objective function when splitting the decision tree, reflecting the strength of the indicator's contribution to the actual response of regional ecological quality. The score value is weighted and fused with the initial weights obtained by the entropy weight method. The fusion coefficient is dynamically determined based on the model validation set performance. After normalization, a corrected set of indicator weight coefficients is generated. The model performance of different fusion coefficient values ​​is tested on the validation set, and the fusion coefficient value with a high correlation coefficient on the validation set is selected.

[0049] The modified weighting coefficients are applied to the standardized indicator data of the current monitoring period, and the modified ecological quality index is generated through linear weighted synthesis. This process ensures that the weight allocation retains the inherent variation information of the data and incorporates the historical response patterns of regional ecological processes, effectively enhancing the index's ability to capture ecological anomalies caused by human interference and its regional applicability, while avoiding assessment bias caused by subjective experience intervention.

[0050] Furthermore, the combined application of the Theil-Sen Median trend slope estimation method and the Mann-Kendall nonparametric significance test in S6 also includes: For the multi-year time series of the modified ecological quality index for each spatial evaluation unit, such as a grid unit, the Theil-Sen Median method is used to calculate the trend slope: traversing all non-repeating time point pairs in the time series, calculating the set of slope values ​​of the index change between each pair of time points, and taking the median of this set as the robust slope estimate of the ecological quality change of the unit; this method has a natural ability to resist individual abnormal fluctuations or outliers that may exist in the time series data, and avoids the distortion of the overall trend judgment by data from a single extreme year.

[0051] The Mann-Kendall nonparametric test is performed synchronously on the same time series: a standardized test statistic is constructed based on the rank relationship of the series data, the sign of the statistic determines the direction of change (upward or downward), and the statistical significance of the trend (significant or insignificant) is determined by comparing its absolute value with the theoretical distribution critical value. This method does not require preconditions such as the data following a normal distribution and is applicable to the non-stationary and nonlinear time series data characteristics commonly found in ecological monitoring. The significance confidence level of the ecological quality index change trend is set at 0.0. 5. At this point, the corresponding critical values ​​are: The normal distribution value is ±1.96; if the normal distribution value is ≤1.65, the significance level is not significant, and the trend change is not statistically significant; if 1.65 < the normal distribution value ≤1.96, the significance level is 90% confidence level, and the trend change has low confidence; if 1.96 < the normal distribution value ≤2.58, the significance level is 95% confidence level, and the trend change has moderate confidence; if the normal distribution value ≥2.58, the significance level is 99% confidence level, and the trend change has high confidence.

[0052] A dual-method result fusion judgment mechanism is established: the slope sign (positive / negative) obtained by Theil-Sen Median is used as the basis for judging the direction of change, and the significance conclusion of the Mann-Kendall test is used as the basis for judging the reliability of the trend. Qualitative trend diagnosis conclusions are generated according to preset logical rules, where the logical rules are as follows: if the slope is positive and the test is significant, it is judged as an improving trend; if the slope is negative and the test is significant, it is judged as a deteriorating trend; and if the test is not significant, it is judged as no significant change.

[0053] Theil-Sen slope symbol Mann-Kendall significance Fusion Diagnostic Conclusion illustrate just >2.58 Significant improvement Ecological quality continues to improve significantly just 1.96-2.58 Significant improvement Ecological quality has improved significantly just 0-1.96 No significant improvement Ecological quality has improved, but not significantly. burden >2.58 Extremely significant degradation Ecological quality continues to deteriorate significantly burden 1.96-2.58 Significant degradation Ecological quality has deteriorated significantly. burden 0-1.96 Insignificant degradation Ecological quality has deteriorated, but not significantly. any ≤0 No significant changes No significant trend change The table above illustrates the mechanism for determining the fusion of results from both methods.

[0054] The output is a structured dataset containing trend direction indicators, significance status indicators, and integrated diagnostic conclusions. This enables the objective and robust identification of the temporal evolution of ecological quality within ecological protection red line areas, effectively distinguishing between continuous ecological changes caused by real human interference and short-term random fluctuations caused by natural fluctuations. This provides a high-confidence temporal diagnostic basis for subsequent human interference attribution analysis and early warning decision-making.

[0055] Furthermore, the construction of the long short-term memory neural network time-series prediction model in S7 includes: For each spatial evaluation unit, the modified ecological quality index continuously acquired within the historical monitoring period is organized into a one-dimensional time series in chronological order, and divided into a model training set and an independent validation set according to the principle of time continuity, to ensure that the training and validation data do not overlap in the time dimension.

[0056] A long short-term memory neural network architecture was constructed, comprising an input gate, a forget gate, an output gate, and cell state units. This architecture dynamically regulates the retention and updating of historical information through a gating mechanism. The input layer has one node, the first hidden layer has 50 LSTM units, and the second hidden layer has 25 LSTM units. The input gate filters the valid information at the current moment, the forget gate selectively removes redundant historical memories, and the output gate integrates the current state and long-term memory to generate prediction data. This enables the model to have the ability to adaptively learn the long-term dependencies and nonlinear fluctuation characteristics in the evolution of ecological quality.

[0057] The time series is constructed into supervised learning sample pairs of historical window input and next time step target value using a sliding window approach. The model parameters are iteratively trained using a gradient optimization algorithm with an adaptive learning rate optimizer of 0.001, a batch size of 16, and 500 training epochs. A validation set is introduced simultaneously to monitor the training process. Training is automatically terminated when the validation error does not improve significantly for several consecutive epochs, effectively suppressing the risk of overfitting and ensuring the model's generalization ability.

[0058] The latest monitoring cycle's revised ecological quality index time series segment is input into the trained model, and the model extrapolates and generates the ecological quality index prediction values ​​for each node within the target prediction period step by step, forming a continuous and smooth time series extrapolation curve.

[0059] Logical verification of prediction results: The predicted trend direction is compared with the historical trend diagnosis conclusion obtained from S6. If there is a significant contradiction, such as an abnormal jump in the predicted value under a historically significant degradation trend, the manual review mechanism is triggered. After the verification is passed, the structured prediction results are output to provide a quantitative basis for the forward-looking early warning of human interference.

[0060] Furthermore, S8, which involves conducting a comprehensive ecological quality assessment of the ecological protection red line area based on preset threshold rules and generating a graded early warning result for the intensity of human interference, also includes: A spatially nested two-level assessment unit system is constructed: basic monitoring grid units are formed by dividing the area into regular geographic grids. A 1km×1km regular geographic grid is used as the basic monitoring unit. At the same time, the administrative management units are defined by the county-level administrative division boundaries approved by the natural resources authorities. A mapping table of grid units and county-level units is established through spatial topology operations to clarify the county-level units to which each grid unit belongs and the proportion of its effective analysis area within that unit.

[0061] Multi-dimensional feature parameters are integrated at the grid cell scale: For each grid cell, the modified ecological quality index representing the current ecological baseline state output by S5, the trend diagnosis conclusion containing the change direction indicator and statistical significance indicator output by S6, and the index sequence features of the prediction period output by S7, such as the overall change direction of the sequence, are integrated to form a three-dimensional feature parameter set of current status level-historical evolution-future trend, providing a data foundation for micro-scale disturbance identification.

[0062] At the county-level unit scale, multi-source parameter collaborative aggregation is performed: the area-weighted average method is used to generate the county-level comprehensive index for the modified ecological quality index; the spatial proportion statistical method is used to generate the county-level trend comprehensive judgment level for the trend diagnosis conclusion, and the proportion of the area of ​​the grid unit showing a significant degradation trend to the total area of ​​the county-level unit red line is calculated; the sequence direction consistency analysis method is used to generate the county-level predicted trend judgment result for the predicted trend characteristics, and the predicted change direction of most grids within the county-level unit is statistically analyzed.

[0063] Construct a comprehensive assessment rule base with multi-dimensional logical coupling: preset judgment logic that combines the county-level comprehensive index, trend comprehensive judgment level, and predicted trend judgment result. For example, when the comprehensive index is lower than the regional benchmark threshold, the proportion of significantly degraded grid area exceeds the preset proportion threshold, and the predicted trend is judged to be continuously declining, a high-level risk judgment is triggered. The rule base covers multiple combination logic scenarios to ensure that the assessment conclusions simultaneously reflect the current ecological status, historical change trajectory, and future evolutionary tendency.

[0064] Implement a tiered early warning and spatial source tracing linkage mechanism: Based on the comprehensive assessment results and according to the preset interference intensity mapping rules, the risk assessment level is mapped to four levels: no interference, slight interference, moderate interference, and severe interference, generating a list of early warnings for human interference intensity at the county-level administrative unit scale and a thematic map; simultaneously, the early warning level is back-linked to the set of grid units that constitute the county-level unit, and the grid areas that trigger medium- and high-level early warnings are spatially highlighted and the list is exported, forming a closed-loop output mechanism of overall county-level risk assessment and precise grid problem location.

[0065] Example 2 This embodiment takes a certain ecological protection red line area as the specific application object, with a total area of ​​349.15 square kilometers, divided into 349 1km×1km grid units. A typical degraded grid unit in the northeast of this area was selected, numbered #187, with center coordinates of 115.98°E, 29.62°N. Time series data of the modified ecological quality index (CEQI) from 2000 to 2023 were extracted; the CEQI was 0.52 in 2000, 0.49 in 2005, 0.45 in 2010, 0.41 in 2015, and 0.38 in 2020.

[0066] Theil-Sen Median slope calculation: Time point pair traversal generated for 5 time points. =10 combinations, namely (0.49-0.52) / 5=-0.0060, (0.45-0.52) / 10=-0.0070, (0.41-0.52) / 15=-0.0073, (0.38-0.52) / 20=-0.0070, (0.45-0.49) / 5=-0.0080, (0.41-0.49) / 10=-0.0080, (0.38-0.49) / 15=-0.0073, (0.41-0.45) / 5=-0.0080, (0.38-0.45) / 10=-0.0070, (0. 38-(0.41) / 5=-0.0060; Median extraction: The slope set is arranged in ascending order as [-0.0080, -0.0080, -0.0080, -0.0073, -0.0073, -0.0070, -0.0070, -0.0070, -0.0060, -0.0060], median=(-0.0073 -0.0070) / 2 = -0.00715; Conclusion: Slope=-0.00715<0, indicating that the ecological quality is declining.

[0067] Mann-Kendall significance test: 2000 (0.52) - all 4 subsequent points < 0.52 → contribution = 0 - 4 = -4; 2005 (0.49) - all 3 subsequent points < 0.49 → contribution = 0 - 3 = -3; 2010 (0.45) - all 2 subsequent points < 0.45 → contribution = 0 - 2 = -2; 2015 (0.41) - 1 subsequent point < 0.41 → contribution = 0 - 1 = -1; 2020 (0.38) - no subsequent points → contribution = 0; S = -4 - 3 - 2 - 1 + 0 = -10; variance is 16.6667, standardized Z-value is -2.695; significance is determined as |Z| = 2.695 > 2.58, reaching a 99% confidence level of significance.

[0068] The Theil-Sen slope is negative, indicating a downward trend; the Mann-Kendall slope is |Z|=2.695, indicating extreme significance; the fusion conclusion is extremely significant.

[0069] The county-level comprehensive index was calculated using an area-weighted average, with input data consisting of 2023 CEQI values ​​from 349 grids, resulting in a value of 0.452. The trend assessment method was based on spatial proportion statistics, with input data consisting of S6 diagnostic results from 349 grids, showing 36 grids with significant / extremely significant degradation, accounting for 10.31%. The predicted trend assessment method was based on sequence direction consistency, with input data consisting of S7 predictions for the 2024–2030 trend, showing 62.5% of grids predicting a decline → continued decline.

[0070] The county-level comprehensive index is 0.452, which is at the lower limit of the medium level and close to the poor threshold of 0.40; the proportion of significantly degraded grids is 10.31%, which is greater than the preset threshold of 5%; the predicted trend is a continuous decline; a moderate interference warning has been triggered.

[0071] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0072] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring anthropogenic interference within ecological protection red lines based on multi-source data fusion, characterized in that, include: S1: Construct a dynamic monitoring indicator system for ecological quality that integrates matrix-stress risk-wildness; S2: Collect multi-source heterogeneous data corresponding to each indicator in the indicator system; S3: Perform standardized preprocessing on the multi-source heterogeneous data, including coordinate system one, spatial resolution resampling, and mask clipping based on the ecological protection red line vector boundary, to generate an analysis dataset with consistent spatial range, coordinate system, and scale. S4: Based on the entropy weight method, objective weights are assigned to the preprocessed index data, and the basic ecological quality index is obtained by weighted synthesis calculation. S5: Using the index data and initial weights used in the calculation of the basic ecological quality index as input features, a weight optimization model is constructed using the XGBoost machine learning algorithm. Through model training, the weights are adaptively corrected to generate the corrected ecological quality index. S6: The Theil-Sen Median trend slope estimation method and the Mann-Kendall nonparametric significance test method are used together to calculate the trend slope and test the significance of the time series of the modified ecological quality index, and generate the ecological quality change trend diagnosis results. S7: Construct a long short-term memory neural network time series prediction model, using historical modified ecological quality index series as input, to predict the ecological quality index for the target period; S8: Establish a two-tiered assessment framework consisting of gridded monitoring units and county-level administrative units. Calculate the ecological quality index, trend slope, and predicted value at the grid unit scale. Generate a comprehensive assessment result through area-weighted aggregation at the county-level unit scale. Determine the intensity classification of human disturbance based on preset threshold rules.

2. The method for monitoring human interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, S1 further includes: The parameters of the matrix dimension include normalized vegetation index, air quality index, carbon storage, leaf area index, water quality index, and aboveground biomass. The parameters of the stress risk dimension include annual rainfall, terrain slope, nighttime light index, population density, and land use intensity index. The parameters of the wildness dimension include vegetation cover, net primary productivity, landscape spread index, and Shannon diversity index. The nighttime light index, population density, and land use intensity index in the stress risk dimension are set as human-dominated input variables, the parameters of the wildness dimension are set as ecosystem response output variables, and the parameters of the matrix dimension are set as baseline parameters.

3. The method for monitoring human interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, The S2 step of collecting multi-source heterogeneous datasets that strictly correspond to each indicator in the indicator system also includes: For the matrix dimension, multispectral and hyperspectral remote sensing images, environmental monitoring data, and biomass remote sensing inversion data were collected from data sources. Regarding the dimension of stress risk, data sources include nighttime light remote sensing images, population spatial distribution data, annual land use change vector data from national land surveys, rainfall observation data, and digital elevation model data. For the wild dimension, long-term remote sensing images and forest and grassland resource monitoring data were collected from data sources; The collected data meets the following requirements: time coverage of the complete monitoring cycle and consistency of time base of each data source; spatial coverage of the entire ecological protection red line area and buffer zone; and data source is standardized data products issued by national or provincial competent authorities that have undergone quality inspection. The collected data are categorized and cataloged according to the vector boundary of the ecological protection red line, based on spatial location and time label, forming a structured data set corresponding to the three dimensions of matrix-stress risk-wildness.

4. The method for monitoring human interference within the ecological protection red line based on multi-source data fusion as described in claim 1, characterized in that, The standardization preprocessing of the multi-source heterogeneous data in step S3 also includes: Convert raster images and vector boundary data from various data sources to the CGCS2000 coordinate system. Raster data were resampled at a uniform spatial resolution, with continuous ecological parameter data using bilinear interpolation and categorical attribute data using the nearest neighbor method. Using the vector boundary of the ecological protection red line approved by the natural resources authority as a mask, the resampled raster data is cropped to retain the effective pixels within the boundary. The cropped dataset is checked for coordinate system consistency, resolution compliance, and mask boundary fit. Any abnormal data is corrected by backtracking.

5. The method for monitoring anthropogenic interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, S4 further includes: The preprocessed index data are attribute-discriminated to distinguish between positive and negative indicators. The range standardization method is then used to perform dimensionless processing on the positive and negative indicators to generate a standardized index data matrix. Based on the standardized indicator data matrix, the weight distribution of each indicator in all spatial evaluation units is calculated, and the information entropy value of each indicator is calculated according to the weight distribution. The difference coefficient of each indicator is calculated based on the information entropy value. The difference coefficient is 1 minus the information entropy value. The difference coefficient is then normalized to obtain the objective weight coefficient of each indicator. Weighting formula based on ecological sensitivity Adjustments were made to the objective weighting coefficients, among which... Here, S represents the weights of the original indicators calculated using the entropy weight method, and S is the ecological sensitivity index calculated based on topography, vegetation type, biodiversity, etc. These are ecological sensitivity thresholds determined through historical data statistics. This is a sensitivity adjustment coefficient determined based on regional ecological characteristics. The weights of the indicators are weighted by ecological sensitivity. The ecological sensitivity index L is calculated using a multi-factor weighted superposition method. ,in The comprehensive weight of the i-th indicator is... Assign a value to the sensitivity level of the i-th indicator; The basic ecological quality index is generated by linearly weighting the standardized values ​​of each indicator with the adjusted objective weight coefficients.

6. The method for monitoring anthropogenic interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, The S5 uses the XGBoost machine learning algorithm to construct a weight optimization model, including: Training samples were constructed using standardized indicator data from multiple time-series nodes within a historical monitoring period as input features and the comprehensive ecological quality assessment level as the target variable. Train the XGBoost regression model, iteratively build decision tree ensembles through gradient boosting, and apply regularization constraints during training; The gain-type importance scores of each indicator feature are extracted from the trained model. The gain-type importance scores are then weighted and fused with the initial weights obtained by the entropy weight method. The fusion coefficients are dynamically determined based on the performance of the model validation set. The fusion results are then normalized to generate the corrected indicator weight coefficients. The corrected index weighting coefficients are linearly weighted and synthesized with the standardized index data of the current monitoring period to generate the corrected ecological quality index.

7. The method for monitoring anthropogenic interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, The S6 method, which combines the Theil-Sen Median trend slope estimation method with the Mann-Kendall nonparametric significance test, also includes: For the time series of the modified ecological quality index of each spatial evaluation unit, the Theil-Sen Median method is used to calculate the trend slope, including calculating the slope values ​​between all non-repeating time point pairs in the time series and taking the median. Perform the Mann-Kendall nonparametric test on the same time series, including constructing the test statistic based on the rank of the series data, determining the direction of change based on the sign of the test statistic, and determining the significance of the trend based on the comparison between the absolute value of the test statistic and the preset critical value. Based on the sign of the trend slope and the significance conclusion of the Mann-Kendall test, a trend diagnosis conclusion is generated according to preset rules. The trend diagnosis conclusion includes a trend direction indicator and a significance status indicator.

8. The method for monitoring human interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, The long short-term memory neural network time-series prediction model constructed in S7 includes: For each spatial evaluation unit, the historical modified ecological quality index time series is divided into a training set and a validation set; The time series is converted into supervised learning samples using a sliding window method, and a long short-term memory neural network is constructed and trained. During the training process, the parameters are updated through a gradient optimization algorithm, and an early stopping strategy is adopted based on the performance on the validation set. Input the time series segment of the corrected ecological quality index for the current monitoring period into the trained long short-term memory neural network to generate the ecological quality index prediction sequence for the target prediction period. The trend direction of the predicted sequence is compared with the trend diagnosis conclusion of claim S6. If there is a contradiction, a manual review process is triggered. After the review is passed, the prediction result is output.

9. The method for monitoring human interference within ecological protection red lines based on multi-source data fusion as described in claim 1, characterized in that, S8, which generates a comprehensive evaluation result at the county-level unit scale through area-weighted aggregation and determines the intensity classification of human interference according to a preset threshold rule, also includes: Establish monitoring grid units divided by regular grids and county-level administrative units defined by county-level administrative division boundaries approved by the natural resources authorities. Construct a mapping relationship between grid units and county-level administrative units through spatial topology operations. The mapping relationship includes the county-level administrative unit to which each grid unit belongs and its area percentage. For each grid cell, the modified ecological quality index described in S5, the trend diagnosis conclusion described in S6, and the predicted trend characteristics described in S7 are integrated to form three-dimensional feature parameters; Based on the attribution mapping relationship, the county-level comprehensive index is calculated by the area-weighted average method for the modified ecological quality index, the county-level trend comprehensive judgment level is determined by the spatial proportion statistical method for the trend diagnosis conclusion, and the county-level predicted trend judgment result is determined by the sequence direction consistency analysis method for the predicted trend characteristics. Based on the preset condition combination judgment logic, the county-level comprehensive index, the county-level trend comprehensive judgment level, and the county-level predicted trend judgment result are comprehensively judged. Based on the comprehensive judgment results and the preset interference intensity mapping rules, a list of early warnings for human interference intensity and a thematic map of county-level administrative units are generated. The early warning level is then back-linked to the corresponding grid unit, and the grid areas with medium and high level early warnings are spatially labeled and the list is exported.

Images