Restoration area full life cycle evaluation method based on multi-source remote sensing and ground monitoring

Abstract

The invention relates to the technical field of ecological restoration assessment, in particular to a restoration area full life cycle assessment method based on multi-source remote sensing and ground monitoring, and the method comprises the steps: firstly obtaining three types of multi-source data of satellite remote sensing, unmanned aerial vehicle / near-ground remote sensing and ground monitoring, and generating a unified reference data set through space-time alignment, quality control and restoration processing; multi-modal features are extracted and fused, and a spatial-temporal feature sequence representing the ecological state is formed; life cycle stages are automatically divided through the model, an evaluation index set is dynamically matched, and stage self-adaptive ecological health degree scores are calculated; and finally, predicting an ecological state change trend, judging a degradation risk and outputting spatialized early warning information. According to the invention, automatic and precise evaluation and degradation risk early warning of the life cycle of the ecological system in the restoration area can be realized.

Description

Technical Field

[0001] This application relates to the field of ecological restoration assessment technology, and in particular to a method for assessing the entire life cycle of a restoration area based on multi-source remote sensing and ground monitoring. Background Technology

[0002] In ecological restoration projects, accurately assessing the evolutionary process, health status, and degradation risk of the ecosystem in the restoration area is a core prerequisite for optimizing restoration strategies and ensuring restoration effectiveness. Existing ecological restoration assessment methods largely rely on single remote sensing indicators (such as the Normalized Difference Vegetation Index, NDVI) or phased manual quadrat surveys, which have significant limitations: First, multi-source data (satellite, UAV, ground sensor) differ in spatiotemporal scale and data format, making joint analysis and deep fusion difficult; second, data quality is easily affected by factors such as cloud cover, sensor malfunction, and seasonal fluctuations, resulting in insufficient stability and reliability of assessment results; third, the division of ecological life cycle stages largely depends on expert subjective experience, lacking standardized and automated identification models, making it difficult to adapt to the general assessment needs of different types of restoration areas; fourth, traditional assessment indicator systems are fixed and singular, unable to be dynamically adjusted according to different stages, resulting in insufficient specificity and accuracy of health scores, and a lack of early warning capabilities for ecological degradation risks. To address these issues, this application proposes an ecological assessment method that integrates multi-source data and adapts to life cycle stages, enabling full-cycle dynamic monitoring, accurate assessment, and risk warning of the ecosystem in the restoration area. Summary of the Invention

[0003] To facilitate automated and precise assessment of the ecosystem life cycle and early warning of degradation risks in restoration areas, this application provides a method for assessing the entire life cycle of restoration areas based on multi-source remote sensing and ground monitoring.

[0004] A method for assessing the entire life cycle of remediation areas based on multi-source remote sensing and ground monitoring includes: Acquire multi-source data of the target remediation area within the assessment time window. The multi-source data includes satellite remote sensing data, UAV or near-ground remote sensing data, and ground monitoring data. Spatiotemporal alignment, quality control, and data repair are performed on multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scale. Based on multi-source benchmark datasets, multimodal feature extraction and fusion are performed to generate multimodal spatiotemporal feature sequences that comprehensively represent the ecological state of the restoration area; Based on multimodal spatiotemporal feature sequences, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through a stage identification model; Based on the identified life cycle stages, dynamically select a set of ecological evaluation indicators that are appropriate for the target life cycle stage; Based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, a stage-adaptive ecological health score is calculated. Based on multimodal spatiotemporal feature sequences, life cycle stages, and ecological health scores, the changing trend of the ecological status of the restoration area is predicted; Based on the changing trends, determine whether there is a risk of degradation and output spatial ecological risk early warning information.

[0005] Optionally, spatiotemporal alignment, quality control, and data repair processes are performed on the multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scales, including: Temporal interpolation and spatial resampling are performed on satellite remote sensing data, UAV or near-ground remote sensing data and ground monitoring data, and unified to the preset analysis time step and spatial resolution grid. Identify and label cloud-occluded pixels, sensor outliers, and spatially discontinuous areas in the data; For cloud-occluded pixels and spatially discontinuous areas, multi-temporal data and adjacent spatial information are integrated for collaborative interpolation and repair. Radiometric normalization and topographic correction were performed on the restored data from different sources to eliminate the effects of differences in observation conditions and topography. All corrected data are mapped to the same spatiotemporal grid to form a multi-source benchmark dataset.

[0006] Optionally, for cloud-occluded pixels and spatially discontinuous areas, collaborative interpolation and repair by fusing multi-temporal data and neighboring spatial information includes: Construct a spatiotemporal cube centered on the target pixel. The spatiotemporal cube contains multi-period data of the target pixel in the time dimension and neighboring pixel data in the spatial dimension. If the target pixel is completely missing in a single time phase due to cloud obstruction, the spatiotemporal kriging interpolation method is used to estimate it using the value of the effective pixel within the spatiotemporal cube. If the target pixel is missing in multiple consecutive time phases, the surface phenological change curve is introduced as a priori constraint, and iterative repair is carried out in combination with the temporal pattern of similar pixel groups.

[0007] Optionally, based on multi-source benchmark datasets, multi-modal feature extraction and fusion are performed to generate a multi-modal spatiotemporal feature sequence that comprehensively represents the ecological state of the restoration area, including: Extract time-series vegetation index, surface temperature, surface deformation, and land use type characteristics from satellite remote sensing data; Extract high-resolution vegetation cover, canopy height, and micro-topographic erosion features from UAV or near-ground remote sensing data; Extract time-series data of soil physicochemical properties, hydrological parameters, and biodiversity survey characteristics from ground monitoring data; All extracted features are standardized and denoised, and then aligned according to the time series. An attention mechanism was used to fuse time-series data of vegetation index, land surface temperature and soil physicochemical properties across modes to generate coupled features characterizing environmental stress. By spatially correlating and integrating high-resolution vegetation cover, canopy height, and biodiversity survey characteristics, composite features representing ecosystem structure are generated. Trend analysis and overlay of surface deformation characteristics and micro-topographic erosion characteristics are performed to generate risk basement characteristics that characterize surface stability. Based on hydrological parameters and land use type characteristics, we calculate the hydrological regulation capacity change index and generate functional response characteristics that characterize ecological functions. By integrating coupling features, composite features, risk-based features, and functional response features, a multimodal spatiotemporal feature sequence is formed.

[0008] Optionally, based on multimodal spatiotemporal feature sequences, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through a stage identification model, including: Multimodal spatiotemporal feature sequences are input into a time-series segmentation model to initially detect time points where significant changes in ecological status occur; Based on time nodes, the entire time series is divided into multiple consecutive time periods; Extract the mean, trend, and fluctuation features of the multimodal spatiotemporal feature sequence within each time period; Based on mean characteristics, trend characteristics, and fluctuation characteristics, a discriminative feature vector for the corresponding time period is formed by combining them. The discriminant feature vector for each time period is input into the stage classifier, and the confidence score of each time period belonging to different life cycle stages is output. The category with the highest confidence level is selected as the target lifecycle stage.

[0009] Optionally, based on mean characteristics, trend characteristics, and volatility characteristics, the discriminant feature vector for the corresponding time period can be formed by combining the following: The mean, trend, and volatility characteristics are normalized. According to the preset feature order, the normalized mean feature, trend feature and fluctuation feature are concatenated to generate an initial feature vector. Principal component analysis is used to reduce the dimensionality of the initial eigenvectors, remove redundant information, and retain the principal components. The dimensionality-reduced feature vectors are used as discriminant feature vectors, where the dimension of the discriminant feature vectors is lower than that of the initial feature vectors.

[0010] Optionally, based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequences, a stage-adaptive ecological health score can be calculated: Establish a mapping relationship library between life cycle stages and ecological evaluation indicators, with different core and auxiliary indicators corresponding to different life cycle stages in the mapping relationship library; Based on the identified target life cycle stage, the corresponding core indicators and auxiliary indicators are matched from the mapping relationship library to form an initial ecological evaluation indicator set; The actual observed values ​​of each initial indicator in the initial ecological evaluation index set are calculated based on multimodal spatiotemporal feature sequences. Obtain the missing rate corresponding to the actual observation value, and select the initial indicators with a missing rate less than or equal to the preset missing threshold to form the final ecological evaluation indicator set; Trend stability analysis was performed on the spatiotemporal characteristic sequences corresponding to each target indicator in the final ecological evaluation indicator set, and the corresponding spatiotemporal characteristic trend stability was obtained. Based on the stability of spatiotemporal characteristics, the effectiveness coefficient of the target indicator is determined; Based on the effectiveness coefficient, match the influence weight of the corresponding target indicator; The actual observed values ​​of each target indicator are classified and quantitatively scored. Based on the impact weights and quantitative scores of each target indicator, a phase-adaptive comprehensive score for ecological health is calculated.

[0011] Optionally, based on the effectiveness coefficient, the influence weights of the corresponding target indicators may include: Obtain historical monitoring data corresponding to the target life cycle stage, and extract the numerical sequence of each target indicator in the historical monitoring data; Trend decomposition is performed on the numerical sequence to separate the long-term trend component, seasonal cycle component, and random fluctuation component. Based on the long-term trend component, calculate the average rate of change of each target indicator during the target life cycle stage. Based on the seasonal cyclical component, the cyclical stability coefficient of each target indicator is calculated within the target life cycle stage. Based on random fluctuation components, calculate the anti-interference capability coefficient of each target indicator during the target life cycle stage; The average rate of change, periodic stability coefficient, and anti-interference ability coefficient are dimensionless and then input into the weight allocation function. Based on the output of the weight allocation function and the effectiveness coefficient, the initial influence weights of each target indicator are generated. The initial influence weights are normalized to generate the final influence weights.

[0012] Optionally, based on the output of the weighting function and the effectiveness coefficient, the initial influence weights for each target indicator are generated, including: The average rate of change, periodic stability coefficient, and anti-interference ability coefficient are normalized to generate a normalized dynamic vector. The normalized dynamic vector is combined with the effectiveness coefficient to generate a feature combination of the target index; The feature combination is input into the weight allocation function; The weighting function outputs the initial weight values ​​of the target index through a mapping relationship.

[0013] In summary, this application includes the following beneficial technical effects: 1. Through a series of processes such as spatiotemporal alignment, quality control, and data repair, the inconsistencies in spatiotemporal scale and data quality between satellite remote sensing, UAV remote sensing, and ground monitoring data were resolved, forming a unified multi-source benchmark dataset, which provides high-quality data support for subsequent evaluation. 2. Based on multimodal spatiotemporal feature sequences and stage identification models, the automatic division of ecological life cycle stages is realized, overcoming the subjective defects of traditional methods that rely on expert experience, improving the standardization and reliability of stage identification, and being able to output the confidence level of each stage, thus enhancing the credibility of the results. 3. Based on the ecological characteristics of different life cycle stages, the corresponding evaluation index set is dynamically selected, and the weights are allocated by combining the effectiveness coefficient and historical change characteristics of the indicators. This achieves stage-adaptive health scoring. Compared with the evaluation method of fixed index system, the scoring results are more targeted and accurate, and can truly reflect the ecological health level at different stages. 4. By integrating multi-dimensional feature data to predict the trend of ecological status changes, it is possible to identify ecological degradation risks in advance and output spatial early warning information, clarify the spatial distribution and level of risks, provide a scientific basis for optimizing and adjusting restoration strategies and preventing and controlling risks, and help ensure the long-term effectiveness of ecological restoration projects. 5. This method is applicable to various ecological restoration projects such as mine restoration, wetland restoration, and desertification control. It can flexibly adjust parameters such as data sources, assessment time windows, and evaluation indicators according to the characteristics of different restoration areas and assessment needs, and has strong applicability and scalability. Attached Figure Description

[0014] Figure 1 This is a main flowchart of a method for assessing the entire life cycle of a remediation area based on multi-source remote sensing and ground monitoring, according to an embodiment of this application. Figure 2 This is a flowchart illustrating the steps involved in generating a multimodal spatiotemporal feature sequence that comprehensively represents the ecological state of the restoration area. Figure 3It is a flowchart of the steps to divide and output the life cycle stage of the repair area as the target life cycle stage. Detailed Implementation

[0015] Reference Figure 1 A method for assessing the entire life cycle of a remediation area based on multi-source remote sensing and ground monitoring, comprising steps S101 to S108: Step S101: Obtain multi-source data of the target remediation area within the evaluation time window. The multi-source data includes satellite remote sensing data, UAV or near-ground remote sensing data, and ground monitoring data.

[0016] Specifically, in this embodiment, the target restoration area refers to a specific geographical area requiring ecological restoration effect assessment and life cycle monitoring. Its boundaries can be determined based on the planning scope of the restoration project, the integrity of the ecosystem, or administrative boundaries, such as the ecological restoration area after mining or the restoration area of ​​a degraded wetland. The assessment time window refers to a continuous time interval set for ecological life cycle assessment. It can be flexibly set according to the cycle of the restoration project and the assessment requirements. For example, short-term restoration projects can be set for 1-3 years, and long-term restoration projects can be set for 5-10 years. The time window needs to cover the key nodes of ecosystem succession to ensure that the complete process of ecological state from initial restoration to stable development can be captured. Satellite remote sensing data refers to Earth surface observation data obtained by sensors (such as multispectral sensors, thermal infrared sensors, synthetic aperture radar, etc.) carried by artificial Earth satellites, including but not limited to vegetation indices (such as NDVI, EVI), surface temperature, surface deformation / roughness, land use change, etc. Its advantages are wide coverage, continuous time series, and the ability to provide macro-scale ecological state information. Unmanned aerial vehicle (UAV) or near-ground remote sensing data refers to regional data acquired by UAVs equipped with high-resolution cameras, LiDAR, multispectral sensors, or near-ground (e.g., ground-based remote sensing equipment). This data primarily includes high-resolution vegetation classification data, canopy height and structural parameters, micro-topography, and erosion characteristics. It enables fine feature extraction at the micro- and meso-scales, compensating for the insufficient spatial resolution of satellite remote sensing data. Ground monitoring data refers to raw data directly reflecting the ecosystem's state, acquired through deploying sensor networks and conducting field surveys within the restoration area. This data includes soil moisture, soil temperature, soil nutrients (such as nitrogen, phosphorus, and potassium content), soil heavy metal content, hydrological and meteorological data (precipitation and evaporation), groundwater level, biodiversity (species abundance and community structure), and biomass. These data directly reflect the core physiological and environmental parameters of the ecosystem. These three types of data complement each other, forming the foundation for a comprehensive data source characterizing the ecological state of the restoration area.

[0017] Step S102: Perform spatiotemporal alignment, quality control, and data repair on the multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scale.

[0018] Specifically, in this embodiment, spatiotemporal alignment refers to adjusting multi-source data from different sources and with different spatiotemporal resolutions to a unified time scale and spatial framework, eliminating inconsistencies in time and space, and ensuring that the data can be jointly analyzed. Quality control refers to identifying and processing errors, outliers, missing values, and other issues in multi-source data through a series of technical means to ensure the accuracy, completeness, and reliability of the data. Data repair refers to using reasonable interpolation, fitting, or other completion methods to restore the integrity and continuity of the data, addressing defects such as missing values ​​and anomalies. Spatiotemporal scale unification refers to adjusting the time interval (time scale) and spatial resolution (spatial scale) of multi-source data to a consistent standard, such as a unified time step of 1 month and a spatial resolution of 10 meters. The multi-source benchmark dataset refers to a standardized dataset with consistent spatiotemporal scale, reliable data quality, and direct usability for subsequent feature extraction and evaluation analysis, formed after spatiotemporal alignment, quality control, and data repair processing. Due to inherent differences in the acquisition methods and spatiotemporal resolutions of satellite remote sensing, UAV remote sensing, and ground monitoring data (e.g., satellite data may have a spatial resolution of 10 meters to 1 kilometer and a temporal resolution of 1 to 16 days; UAV data can achieve centimeter-level spatial resolution, but has a small coverage area and an inconsistent acquisition cycle; ground monitoring data is based on fixed-point observations, with high temporal resolution but limited spatial coverage), standardization processing is necessary to eliminate these differences and lay the foundation for subsequent feature fusion and evaluation. The processing must strictly adhere to the principles of spatiotemporal consistency and data reliability to ensure that the benchmark dataset accurately reflects the spatiotemporal distribution characteristics of the ecological state of the restoration area.

[0019] Step S103: Based on the multi-source benchmark dataset, perform multi-modal feature extraction and fusion to generate a multi-modal spatiotemporal feature sequence that comprehensively represents the ecological state of the restoration area.

[0020] Specifically, in this embodiment, multimodal feature extraction refers to extracting feature variables that reflect different attributes of the ecosystem from different types of data sources (satellite remote sensing, UAV remote sensing, and ground monitoring) within a multi-source benchmark dataset. For example, vegetation growth characteristics are extracted from satellite remote sensing data, and soil environmental characteristics are extracted from ground monitoring data. Different types of features constitute different modalities. Multimodal feature fusion refers to integrating feature variables from different modalities using certain mathematical methods and models to eliminate information redundancy, explore the complementarity between features of different modalities, and form a comprehensive feature that can fully and accurately represent the state of the ecosystem. Multimodal spatiotemporal feature sequence refers to a feature sequence that integrates features from different modalities and includes information on changes over time. That is, each time step corresponds to a set of comprehensive features that can reflect the dynamic changes in the state of the ecosystem over time. Comprehensive representation of the ecological state of the restoration area means that this feature sequence can comprehensively cover multiple dimensions of the ecosystem, such as structure, function, and environmental stress, and truly reflect the overall condition and changing trends of the ecosystem in the restoration area. Multi-source benchmark datasets contain raw data of different types and dimensions. These need to be transformed into feature variables that reflect the essential attributes of the ecosystem through feature extraction. Furthermore, multimodal fusion techniques are used to integrate information from different types of features, forming a comprehensive feature sequence that can fully and accurately characterize the ecological state. Feature extraction requires tailored methods designed for the characteristics of different data sources, while multimodal fusion must fully explore the complementarity of features across different modalities, avoid information redundancy, and enhance the representational power of the features.

[0021] Step S104: Based on the multimodal spatiotemporal feature sequence, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through the stage identification model.

[0022] Specifically, in this embodiment, the stage identification model refers to a model built based on artificial intelligence algorithms such as machine learning and deep learning, capable of automatically identifying the life cycle stages of an ecosystem. This model takes multimodal spatiotemporal feature sequences as input and learns the feature patterns of different life cycle stages to classify and identify the current stage. Life cycle stages refer to different developmental stages with significant differences in characteristics within the restoration area. Based on the characteristics of ecological restoration, they are typically divided into stages such as soil stabilization (initial improvement of the soil environment, providing basic conditions for plant growth), pioneer plant stage (the establishment of pioneer plants that are tolerant of poor soil and grow rapidly, leading to a gradual increase in vegetation coverage), community succession stage (various plants begin to compete for growth, increasing biodiversity, and gradually complicating the community structure), and functional stability stage (the ecosystem structure is intact, its functions are stable, and it can self-sustain). The target life cycle stage refers to the specific life cycle stage currently occupied by the restoration area, as identified by the stage identification model.

[0023] Step S105: Based on the identified life cycle stage, dynamically select a set of ecological evaluation indicators that are appropriate for the target life cycle stage.

[0024] Specifically, in this embodiment, dynamic selection refers to automatically matching the corresponding ecological evaluation indicator set based on the current target life cycle stage of the restoration area. Different life cycle stages use different evaluation indicator sets, avoiding the shortcomings of fixed and singular indicator systems in traditional assessment methods, and improving the pertinence and accuracy of the assessment. An ecological evaluation indicator set refers to a collection of indicators used to assess the health of an ecosystem. Each indicator corresponds to a specific attribute of the ecosystem (such as soil quality, vegetation growth, biodiversity, ecological function, etc.). The selection of the indicator set must adhere to the principles of pertinence (meeting the ecological characteristics of the current stage), scientific rigor (objectively reflecting the ecological state), and operability (data is available and quantifiable). Because the characteristics and restoration goals of ecosystems differ at different life cycle stages, their evaluation focuses also vary: for example, the core of the soil stabilization stage is the improvement of the soil environment, and evaluation indicators should focus on soil physicochemical properties (such as soil nutrient content and heavy metal content); the core of the pioneer plant stage is the establishment and growth of vegetation, and evaluation indicators should focus on vegetation cover and pioneer plant biomass; the core of the community succession stage is the increasing complexity of community structure and the enhancement of biodiversity, and evaluation indicators should focus on biodiversity and vegetation community structure; the core of the functional stabilization stage is the stability and self-maintenance of ecological functions, and evaluation indicators should focus on ecological functions (such as hydrological regulation capacity and carbon sequestration capacity). Therefore, it is necessary to dynamically select corresponding core and auxiliary indicators based on the characteristics of the target life cycle stage to form an ecological evaluation indicator set.

[0025] Step S106: Based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, calculate the stage-adaptive ecological health score.

[0026] Specifically, in this embodiment, the stage-adaptive ecological health score is calculated using a specific scoring model based on the ecological evaluation index set of the current life cycle stage, combined with relevant feature data from multimodal spatiotemporal feature sequences. This score reflects the quantitative health status of the ecosystem at that stage (e.g., 0-100 points). Stage adaptation is reflected in the fact that the selection of indicators, indicator weights, and scoring standards in the scoring model are all matched to the current life cycle stage. Different scoring systems exist for different stages, ensuring that the scoring results objectively and accurately reflect the ecological health level of that stage. The ecological health score directly reflects the health status of the ecosystem. A higher score indicates a more complete ecosystem structure, more stable function, and better restoration effects; a lower score indicates structural defects, functional impairment, or degradation risk in the ecosystem, requiring adjustments to restoration strategies.

[0027] Step S107: Based on multimodal spatiotemporal feature sequences, life cycle stages, and ecological health scores, predict the changing trend of the ecological status of the restoration area.

[0028] Specifically, in this embodiment, the trend of ecological status change refers to the direction of the health status development of the restoration area's ecosystem over a future period (e.g., 1-3 years), including four trend types: continuous improvement, stability, slow degradation, and rapid degradation. The prediction process employs time series prediction models or machine learning prediction models, using multimodal spatiotemporal feature sequences (reflecting historical ecological status changes), life cycle stages (reflecting the current stage), and ecological health scores (reflecting the current health level) as model inputs. By learning the correspondence between "inputs" and "change trends" in historical data, the model predicts future trends. Commonly used prediction models include Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), Prophet time series prediction models, and gradient boosting regression models. For example, if the multimodal spatiotemporal feature sequences show a continuous upward trend in vegetation cover, biodiversity, etc., and the current stage is community succession with an ecological health score of 75 (good), the prediction model may output a trend prediction result of "continuous improvement in ecological status over the next 1-2 years, with the health score expected to rise to around 85."

[0029] Step S108: Based on the changing trend, determine whether there is a risk of degradation and output spatial ecological risk early warning information.

[0030] Specifically, in this embodiment, degradation risk refers to the possibility of structural damage, functional degradation, and a decline in health score in the future development of an ecosystem. Based on the type and intensity of the change trend, it can be divided into four levels: no risk, low risk, medium risk, and high risk. The judgment rule is as follows: if the change trend is "continuous improvement," it is judged as no risk; if it is "stable," it is judged as low risk; if it is "slow degradation," it is judged as medium risk; and if it is "rapid degradation," it is judged as high risk. Spatial ecological risk early warning information maps degradation risk levels onto the spatiotemporal grid of the restoration area according to spatial distribution patterns, generating risk early warning thematic maps. This clarifies the risk level and spatial distribution characteristics of different areas. For example, the northern area of ​​the restoration area is predicted to be high risk (rapid degradation), the central area is medium risk (slow degradation), and the southern area is no risk (continuous improvement). The thematic map can intuitively display the risk status of each area. The early warning information also includes risk cause analysis (such as the risk of degradation due to the rebound of heavy metal content in the soil in the northern region) and recommended countermeasures (such as strengthening soil remediation in the northern region and increasing vegetation cover). The output early warning information can be in the form of charts, reports and other forms to provide accurate and intuitive decision support for restoration project managers, so as to achieve early warning and proactive prevention and control of ecological degradation risks.

[0031] In one embodiment of this example, step S102 performs spatiotemporal alignment, quality control, and data repair on the multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scale, including steps S201 to S205: Step S201: Perform temporal interpolation and spatial resampling on satellite remote sensing data, UAV or near-ground remote sensing data and ground monitoring data, and unify them to the preset analysis time step and spatial resolution grid.

[0032] Specifically, in this embodiment, temporal interpolation is used when data is missing or the time intervals are inconsistent in the time series. Based on the changing trends of existing data, mathematical methods such as linear interpolation, spline interpolation, and Fourier interpolation are used to estimate the data at the missing time points, ensuring the time series is continuous and the intervals are consistent. Spatial resampling adjusts the spatial resolution of the data to a preset standard resolution. This is achieved by recalculating the pixel values ​​of the original data to generate raster data at the new resolution. Commonly used spatial resampling methods include bilinear interpolation, nearest neighbor interpolation, and cubic convolution interpolation. The analysis time step is a preset, uniform time interval used for ecological assessment. It needs to be determined based on the rate of change of the ecosystem. For example, the peak vegetation growth period can be set to 1 month, and the dormant period can be set to 3 months to ensure that dynamic changes in the ecological state can be captured. Spatial resolution grids divide the target remediation area into several square or rectangular grid cells of the same size. The side length of each grid cell is the spatial resolution. For example, a 10m × 10m grid has a spatial resolution of 10m. The preset spatial resolution needs to be set according to the assessment accuracy requirements, generally between 10m and 50m. For refined assessment scenarios (such as small mine remediation areas), it can be set to 1-5m. Temporal interpolation uses linear interpolation or spline interpolation. For meteorological data with obvious periodicity (such as temperature and precipitation), Fourier interpolation is used to maintain the periodic characteristics. Spatial resampling uses bilinear interpolation (suitable for continuous data such as temperature and vegetation indices) or nearest neighbor interpolation (suitable for categorical data such as land use types) to ensure the accuracy and rationality of the resampled data. Through this step, different source data are unified under the same spatiotemporal framework, solving the problem of inconsistent data scales.

[0033] Step S202: Identify and mark cloud-occluded pixels, sensor outliers, and spatially discontinuous areas in the data.

[0034] Specifically, in this embodiment, cloud-covered pixels are pixels in satellite remote sensing images that are covered by clouds. Because clouds obstruct the sensor's observation of the ground, the raw data of these pixels cannot accurately reflect the ground's ecological state, which is a common quality problem in satellite remote sensing data. Sensor outliers are abnormal values ​​in the data that exceed the normal range due to sensor malfunctions, abnormal observation conditions (such as strong light, atmospheric interference), etc. These values ​​do not match the actual ecological state and will affect the accuracy of the assessment results. Spatial discontinuities are areas where the data shows abrupt changes or breaks in value in spatial distribution, that is, the difference in value between adjacent pixels exceeds the reasonable range. These are commonly found at the boundary between restoration and non-restoration areas, in areas of abrupt topographic changes, or at data stitching points. Cloud-obscured pixels are identified using satellite data quality control bands (such as the QA band in Landsat data) and cloud detection algorithms (such as the Fmask algorithm), marking pixels that are completely or partially obscured by clouds. Sensor anomalies are detected using statistical methods, such as the 3σ criterion (values ​​exceeding the data mean ± 3 standard deviations are considered anomalies) or by detecting sudden aberrations in time series data (such as cumulative sum control charts). Spatial discontinuities are identified by calculating the spatial gradient of the data; when the difference in values ​​between adjacent pixels exceeds a preset threshold (set according to the data type, such as a threshold of 0.3 for the vegetation index), these spatially discontinuous regions are clearly marked, providing targeted targets for subsequent remediation.

[0035] Step S203: For cloud-occluded pixels and spatially discontinuous areas, perform collaborative interpolation and repair by fusing multi-temporal data and neighboring spatial information.

[0036] Specifically, in this embodiment, multi-temporal data refers to remote sensing or monitoring data of the same area acquired at different time points, such as satellite imagery data of the same restoration area acquired in January, February, and March. This data can reflect the temporal variation patterns of the regional ecological state and can be used to supplement missing data. Neighboring spatial information refers to the effective data of adjacent pixels or regions centered on the target pixel or region. Since the ecological states of adjacent regions are often correlated, this data can be used to infer missing values ​​in the target region. Collaborative interpolation and restoration comprehensively utilizes the temporal correlation of multi-temporal data and the spatial correlation of neighboring spatial data, employing interpolation or other mathematical methods to complete and correct missing or abnormal data in cloud-occluded pixels and spatially discontinuous regions, restoring the integrity and continuity of the data. Data loss and discontinuity lead to incomplete ecological state representation, requiring restoration by combining effective information in the spatiotemporal dimensions to ensure data continuity and integrity. The restoration process must fully utilize the spatiotemporal correlation of the data, such as historical and future time-series data of the same pixel and contemporaneous data of adjacent pixels, avoiding errors caused by single-dimensional restoration.

[0037] Step S204: Perform radiometric normalization and topographic correction on the repaired data from different sources to eliminate the effects caused by differences in observation conditions and topography.

[0038] Specifically, in this embodiment, radiometric normalization is a process that eliminates inconsistencies in radiance caused by factors such as atmospheric conditions, solar altitude angle, and sensor response differences during the radiative transmission process in remote sensing data acquired by different sensors at different observation times, ensuring data comparability. Topographic correction eliminates the impact of differences in illumination conditions caused by topographic undulations (such as slope and aspect) on remote sensing data, ensuring that the remote sensing values ​​of the same feature under different topographic conditions accurately reflect its inherent attributes, rather than differences caused by topographic factors. Topographic correction uses a C-correction model or an SCS+C correction model, combined with digital elevation model (DEM) data, to correct the radiance of each pixel, ensuring the comparability of remote sensing values ​​for the same feature under different topographic conditions.

[0039] Step S205: Map all corrected data to the same spatiotemporal grid to form a multi-source benchmark dataset.

[0040] Specifically, in this embodiment, the spatiotemporal grid is a regular grid that divides the target restoration area according to a preset spatial resolution and coordinate system. Each grid cell corresponds to a unique spatial location and time node, serving as the foundation for unified data storage and management. The coordinate system is a reference frame used to define spatial locations, such as the WGS-84 coordinate system or the UTM coordinate system, ensuring the accuracy of the data's spatial location. All satellite remote sensing data, UAV remote sensing data, and ground monitoring data, after undergoing temporal interpolation, spatial resampling, data restoration, radiometric normalization, and terrain correction, are mapped onto this grid. Each grid cell corresponds to a set of multidimensional data containing various ecological parameters.

[0041] In one embodiment of this example, step S203, which involves coordinating interpolation and repair of cloud-occluded pixels and spatially discontinuous regions by fusing multi-temporal data and neighboring spatial information, includes steps S301 to S303: Step S301: Construct a spatiotemporal cube centered on the target pixel. The spatiotemporal cube contains multi-period data of the target pixel in the time dimension and neighboring pixel data in the spatial dimension.

[0042] Specifically, in this embodiment, the target pixel is a pixel in the data that is missing, abnormal, or needs repair; it is the object of data repair. A spatiotemporal cube is a data structure that integrates temporal and spatial dimensions. Centered on the target pixel, it extends spatially within a certain range, such as a 3×3 or 5×5 pixel window, and temporally contains multiple consecutive or adjacent time point data, forming a three-dimensional data cube (spatial x × spatial y × time t). The multi-period data in the temporal dimension are the observation data of the target pixel at different time points, reflecting the temporal variation characteristics of the target pixel; the neighboring pixel data in the spatial dimension are the observation data of adjacent pixels around the target pixel at the same or different time points, reflecting the spatial correlation characteristics between the target pixel and its surrounding area. The temporal dimension of the spatiotemporal cube is determined based on the duration of data loss, generally including 3-5 periods of data before and after the period of data loss for the target pixel, ensuring sufficient coverage of temporal information. The spatial dimension is set to a 3×3 or 5×5 pixel window. The window size needs to balance spatial correlation and computational efficiency. For areas with strong spatial heterogeneity (such as areas with complex terrain), a 5×5 window can be used, while for areas with good spatial homogeneity (such as plains and wetlands), a 3×3 window can be used. The construction of the spatiotemporal cube realizes the integration of spatiotemporal context information of the target pixel, providing data support for accurate restoration.

[0043] Step S302: If the target pixel is completely missing in a single time phase due to cloud obstruction, then the spatiotemporal kriging interpolation method is used to estimate the value of the effective pixels within the spatiotemporal cube.

[0044] Specifically, a complete lack of data in a single time phase refers to a situation where, at a specific point in time (e.g., June 2023), no valid observation data was obtained for the target pixel due to cloud cover or other reasons, and the original data for that point in time is completely blank. Valid pixels are those within the spatiotemporal cube that have complete data, no missing data, and no anomalies. The values ​​of these pixels can accurately reflect the ecological state of the region and are the basis for interpolation estimation.

[0045] Step S303: If the target pixel is missing in multiple consecutive time phases, the surface phenological change curve is introduced as a priori constraint, and iterative repair is performed in combination with the temporal pattern of similar pixel groups.

[0046] Specifically, in this embodiment, continuous multi-temporal missing data refers to the absence of valid data for target pixels at multiple consecutive time points (e.g., June, July, and August 2023) due to cloud cover, sensor malfunction, or other reasons. Relying solely on data from adjacent time points or adjacent pixels is insufficient for accurate data restoration. Surface phenological change curves reflect the regular changes in surface vegetation or other land features with the seasons. For example, the periodic changes in phenological processes such as germination, growth, and withering of vegetation are shown in vegetation index time-series data. These curves reflect the natural temporal changes of land features and can serve as prior knowledge constraints for data restoration. Prior constraints are constraints set before data restoration based on existing theoretical knowledge, historical data, or natural laws. They are used to limit the reasonable range of restoration results and ensure that the restoration results conform to natural laws. A similar pixel group is a set of pixels that are highly similar to the target pixel in static features such as terrain (elevation, slope, aspect), soil type, remediation measures, and land use type. The temporal variation patterns of these pixels are highly consistent with those of the target pixel and can serve as a reference for the target pixel's temporal pattern. A temporal pattern is the variation law of pixel data over time, such as trend, periodicity, or volatility. Iterative remediation is a process of gradually optimizing the remediation result through multiple iterative calculations. First, the missing values ​​of the target pixel are initialized based on the similar pixel group and prior constraints. Then, the constraints and similar pixel weights are adjusted according to the initialization results. The calculations are repeated until the remediation result meets the preset convergence condition (e.g., the numerical difference between two adjacent iterations is less than 0.01).

[0047] Reference Figure 2 In one embodiment of this example, step S103, based on a multi-source benchmark dataset, performs multimodal feature extraction and fusion to generate a multimodal spatiotemporal feature sequence that comprehensively represents the ecological state of the restoration area, including steps S401 to S408: Step S401: Extract time-series vegetation index, surface temperature, surface deformation, and land use type characteristics from satellite remote sensing data.

[0048] Specifically, in this embodiment, temporal vegetation indices are indices obtained from continuous time series that reflect information such as vegetation growth status, coverage, and biomass. Commonly used indices include the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Normalized Difference Water Index (NDWI). These indices are calculated using satellite multispectral data, and their temporal variations reflect the growth cycle and dynamic changes of vegetation. Surface temperature is the temperature of the Earth's surface, obtained through inversion from satellite thermal infrared data, and reflects the regional heat distribution, microclimate characteristics, and the intensity of ecological processes such as vegetation transpiration and soil evaporation. Surface deformation is the change in the Earth's surface position (uplift or subsidence) caused by factors such as geological activity, soil subsidence, and vegetation growth. It is extracted using synthetic aperture radar interferometry (InSAR) technology, and the surface subsidence or uplift rate is calculated using the Small Baseline Set (SBAS) method, reflecting changes in soil stability and the geological environment. Land use type characteristics are derived by classifying satellite multispectral data to classify land use patterns in the restoration area, such as vegetation cover areas, water bodies, bare land, and building land. Changes in these types can reflect the impact of restoration projects on land use patterns. During the extraction process, time-series data are smoothed using methods such as Savitzky-Golay filtering to eliminate noise interference and preserve the true temporal change trends.

[0049] Step S402: Extract high-resolution vegetation cover, canopy height, and micro-topographic erosion features from UAV or near-ground remote sensing data.

[0050] Specifically, in this embodiment, high-resolution vegetation cover is the proportion of vegetation cover area to the total area calculated based on high-resolution UAV imagery. Due to the high spatial resolution of UAV imagery (up to centimeter level), it can accurately identify the distribution range and coverage of vegetation, resulting in higher accuracy compared to vegetation cover calculated from satellite remote sensing data. Canopy height is the vertical height of the vegetation canopy, extracted from UAV LiDAR data. LiDAR emits laser pulses and receives reflected signals from ground objects. By calculating the propagation time of the laser pulses, the elevation information of the ground objects is obtained, leading to the maximum height, average height, and standard deviation of the vegetation canopy, reflecting the vertical structure of the vegetation community. Micro-topographic erosion features are parameters reflecting the micro-topographic undulations and erosion degree of the land surface. They are obtained through difference analysis between the digital surface model (DSM) and digital elevation model (DEM) generated from UAV imagery, including parameters such as slope, slope length, gully density, and erosion depth. These features can accurately depict the erosion status and micro-topographic changes of the land surface. These features can accurately depict the micro-ecological structure and topographic conditions of the restoration area, compensating for the insufficient spatial resolution of satellite remote sensing data.

[0051] Step S403: Extract time-series data of soil physicochemical properties, hydrological parameters, and biodiversity survey characteristics from ground monitoring data.

[0052] Specifically, in this embodiment, the soil physicochemical property time-series data are parameter data reflecting the physical and chemical properties of the soil, obtained over a continuous time series. These include soil moisture, soil temperature, soil pH, soil organic matter content, and soil heavy metal content. These parameters directly affect plant growth and the function of the soil ecosystem. Acquired through continuous monitoring at fixed points using ground sensors, they reflect the dynamic changes in the soil environment. Hydrological parameters reflect the regional hydrological cycle, including precipitation, evaporation, groundwater depth, and surface runoff. These are acquired through monitoring equipment such as meteorological stations, water level gauges, and flow meters, reflecting the region's water resources and hydrological processes. Biodiversity survey characteristics are parameters reflecting the richness and structure of the ecosystem's biological community, obtained through field quadrat surveys. These include the number of plant species, dominant species cover, number of animal species, and the total dry weight of organisms per unit area. Surveys are conducted using transect or quadrat methods, with quadrat size determined according to vegetation type (e.g., 1m × 1m for herbaceous plants and 5m × 5m for shrubs). These parameters directly reflect the biological community structure and function of the ecosystem.

[0053] Step S404: Standardize and denoise all extracted features and align them according to the time series.

[0054] Specifically, in this embodiment, standardization transforms the values ​​of different feature variables into a unified numerical range with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in units and numerical ranges between different features and ensuring that each feature has equal weight and comparability during the fusion process. Commonly used standardization methods include Z-score standardization and min-max standardization. Denoising eliminates random noise in the feature data caused by sensor errors, observational interference, and other factors, retaining effective signals that truly reflect changes in ecological status. Wavelet thresholding is used for time-series features, and median filtering is used for spatial features. Time series alignment uses a unified time step of the multi-source benchmark dataset as a standard to adjust the time series of different features to the same time node. For example, soil physicochemical property time series data and vegetation index time series data are all adjusted to the 1st of each month, ensuring consistency of features in the time dimension and laying the foundation for the subsequent construction of spatiotemporal feature sequences.

[0055] Step S405: Employ an attention mechanism to perform cross-modal feature fusion on time-series data of vegetation index, surface temperature, and soil physicochemical properties to generate coupled features characterizing environmental stress.

[0056] Specifically, cross-modal feature fusion integrates feature data from different modalities (such as vegetation growth modality, microclimate modality, and soil environment modality), mines the correlations between features of each modality, and forms a fused feature that can comprehensively reflect information from multiple dimensions. Environmental stress is an external pressure factor that affects the normal growth and function of an ecosystem, such as drought (insufficient soil moisture), high temperature (excessively high surface temperature), and soil nutrient deficiency. These factors inhibit vegetation growth and affect the health of the ecosystem. Coupled features are formed by fusing multiple interrelated modal features, forming a feature that can comprehensively reflect their synergistic effects. For example, the coupled feature of time-series vegetation index, surface temperature, and soil moisture can comprehensively reflect the combined impact of environmental stress on vegetation growth. First, time-series data of vegetation index, surface temperature, and soil physicochemical properties are treated as three independent modal feature sequences, and the temporal dependency features of each modality are extracted using a bidirectional long short-term memory network (BiLSTM). Then, a multi-head attention mechanism is introduced to calculate the attention weights between different modal features, quantifying the importance of each modal feature for environmental stress assessment (e.g., the attention weight of soil moisture will significantly increase during drought). Finally, the temporal dependency features of each modality are weighted and fused based on the attention weights to generate coupled features.

[0057] Step S406: Spatially correlate and integrate high-resolution vegetation cover, canopy height and biodiversity survey characteristics to generate composite features characterizing ecosystem structure.

[0058] Specifically, in this embodiment, spatial correlation analyzes the correlation and matching relationships between different spatial features. This involves calculating spatial correlation coefficients, such as the Pearson correlation coefficient, to determine the mutual influence between high-resolution vegetation cover, canopy height, and biodiversity (e.g., areas with higher canopy heights often have richer biodiversity). Ecosystem structure refers to the composition, spatial distribution, and abiotic environment configuration of the ecosystem's biological community, including horizontal vegetation cover (vegetation cover), vertical structure (canopy height), and biological community structure (biodiversity). The integrity of ecosystem structure is a crucial indicator of ecosystem health. Composite features are formed by integrating multiple spatial features reflecting different aspects of ecosystem structure, creating a comprehensive and integrated characteristic variable that characterizes the ecosystem's structural status. The integration process employs a spatial weighted fusion method, using spatial correlation coefficients as weights to weight and sum the three features to generate a composite feature. This composite feature comprehensively reflects the ecosystem's vegetation structure (horizontal cover and vertical height) and biological community structure, demonstrating the ecosystem's structural integrity and complexity.

[0059] Step S407: Perform trend analysis and overlay on surface deformation characteristics and micro-topographic erosion characteristics to generate risk basement characteristics that characterize surface stability.

[0060] Specifically, in this embodiment, trend analysis uses mathematical methods such as linear regression to analyze the changing trends of characteristic data over time, determining whether they are rising, falling, or stable. For example, linear regression is used to fit the relationship between the rate of surface deformation and time to determine the long-term trend of surface subsidence or uplift. Surface stability is the ability of the Earth's surface to maintain its shape and position under the influence of natural factors (such as topography and climate) and anthropogenic factors (such as restoration projects and human activities). Surface deformation (subsidence / uplift) and micro-topographic erosion are key factors affecting surface stability. Risk baseline characteristics are basic environmental features that reflect the risk of degradation faced by an ecosystem. Insufficient surface stability is an important risk source of ecosystem degradation (e.g., severe surface erosion may lead to vegetation degradation and soil loss). Therefore, features characterizing surface stability can serve as baseline information for risk assessment. The feature overlay adopts the normalized weighted overlay method. First, the surface deformation trend and micro-topographic erosion intensity features are normalized (converted into values ​​between 0 and 1). Then, weights are set according to the degree of influence of the two on surface stability (e.g., the weight of surface deformation rate is 0.6 and the weight of micro-topographic erosion intensity is 0.4). Finally, the weighted overlay is performed to generate the risk base feature.

[0061] Step S408: Based on hydrological parameters and land use type characteristics, calculate the hydrological regulation capacity change index and generate functional response characteristics that characterize ecological functions.

[0062] Specifically, in this embodiment, hydrological regulation capacity refers to the regulatory role of an ecosystem on regional hydrological processes, including functions such as water conservation, flood peak reduction, and water quality purification, and is one of the core service functions of an ecosystem. Hydrological regulation capacity change indicators are parameters that reflect changes in the ecosystem's hydrological regulation function, obtained by quantifying the relationship between hydrological parameters (such as precipitation, surface runoff, and groundwater level) and land use types. For example, the degree of improvement in hydrological regulation capacity can be characterized by calculating the "runoff reduction rate" (the difference between runoff under natural conditions and runoff after restoration, divided by the runoff under natural conditions). Ecological functions are the various services and benefits that an ecosystem provides to humans and other organisms, including hydrological regulation, climate regulation, and habitat provision. The integrity of ecological functions is a core manifestation of ecosystem health. Functional response characteristics are characteristic variables that reflect the response of ecosystem functions to restoration measures or environmental changes, i.e., whether ecological functions are improved or maintained. Based on hydrological parameters and land use characteristics, a hydrological regulation capacity assessment model, such as the hydrological regulation module of the InVEST model, is established to calculate the hydrological regulation capacity change index and generate functional response characteristics. These characteristics can accurately reflect the dynamic changes in the hydrological regulation function of the ecosystem in the restoration area.

[0063] Step S409: Integrate coupling features, composite features, risk-based features, and functional response features to form a multimodal spatiotemporal feature sequence.

[0064] Specifically, in this embodiment, integration involves combining the different types and dimensions of fusion features (coupling features, composite features, risk-based features, and functional response features) extracted in the previous steps according to a time series, forming a multidimensional feature sequence containing time-dimensional information. Each time step (e.g., monthly) corresponds to a feature vector composed of four fusion features, and the feature vectors of all time steps are arranged in chronological order, thus forming a multimodal spatiotemporal feature sequence. This sequence contains comprehensive information on different modal features (environmental stress, ecological structure, surface stability, and ecological function) while retaining dynamic change information over time. It can comprehensively and accurately characterize the state change process of the restoration area's ecosystem throughout the entire assessment time window, providing core feature support for subsequent life cycle stage identification, health scoring, and risk prediction.

[0065] Reference Figure 3 In one embodiment of this example, based on the multimodal spatiotemporal feature sequence, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through a stage identification model, including steps S501 to S506: Step S501: Input the multimodal spatiotemporal feature sequence into the time-series segmentation model to initially detect time nodes where significant changes in the ecological state occur.

[0066] Specifically, in this embodiment, the temporal segmentation model is used to identify nodes in time-series data that have undergone significant changes. This model can capture abrupt changes in the time dimension of multimodal spatiotemporal feature sequences, which often correspond to transitions in the ecosystem's life cycle stages. Significant changes are substantial alterations in the ecosystem's state characteristics that exceed normal fluctuation ranges, marking the entry into a new stage. Time nodes are specific points in time when significant changes occur. These nodes divide the entire assessment time window into multiple consecutive periods, within which the ecosystem's state characteristics are relatively stable, corresponding to a potential life cycle stage. Commonly used temporal segmentation models include sliding window-based segmentation models, Bayesian change point detection models, and temporal segmentation networks in deep learning (such as TCN). After inputting the multimodal spatiotemporal feature sequences into the temporal segmentation model, the model analyzes the trend changes and fluctuation amplitudes of the feature sequences to output the time nodes where significant changes in the ecological state occur.

[0067] Step S502: Based on the time nodes, divide the entire time series into multiple consecutive time periods.

[0068] Specifically, in this embodiment, the entire time series refers to the multimodal spatiotemporal feature sequence within the evaluation time window, that is, the complete time series from the start time to the end time of the evaluation. A continuous period is the time interval between two adjacent significant change time nodes. Within each continuous period, the state characteristics of the ecosystem are relatively stable, and no substantial stage transition occurs, laying the foundation for the subsequent extraction of feature patterns for each stage.

[0069] Step S503: Extract the mean, trend, and fluctuation features of the multimodal spatiotemporal feature sequence within each time period.

[0070] Specifically, in this embodiment, the mean characteristic is the average value of each characteristic variable in the multimodal spatiotemporal characteristic sequence within each time period, reflecting the average level of the ecosystem state within that time period. The trend characteristic is the changing trend of the multimodal spatiotemporal characteristic sequence within each time period, calculated using methods such as linear regression and nonlinear fitting, reflecting the development direction of the ecosystem state at that stage. The fluctuation characteristic is the dispersion and fluctuation amplitude of the multimodal spatiotemporal characteristic sequence within each time period, obtained by calculating parameters such as standard deviation, coefficient of variation, and fluctuation frequency, reflecting the stability of the ecosystem state at that stage.

[0071] Step S504: Based on the mean characteristics, trend characteristics, and fluctuation characteristics, combine them to form the discriminative feature vector for the corresponding time period.

[0072] Specifically, in this embodiment, the discriminant feature vector is a high-dimensional vector formed by concatenating the mean feature, trend feature, and fluctuation feature of each time period in a certain order. This vector contains all the key information of the ecosystem state pattern during that time period and serves as the input data for the stage classifier to perform classification and recognition. For example, if the mean feature contains the mean of four fusion features (mean of coupled feature, mean of composite feature, mean of risk basis feature, and mean of functional response feature), the trend feature contains the slope of four fusion features, and the fluctuation feature contains the standard deviation of four fusion features, then the dimension of the discriminant feature vector is 4 + 4 + 4 = 12 dimensions.

[0073] Step S505: Input the discriminative feature vector of each time period into the stage classifier and output the confidence level of each time period belonging to different life cycle stages.

[0074] Specifically, in this embodiment, the stage classifier is a classification model built based on machine learning algorithms, used to classify the input discriminant feature vector into different life cycle stages. Commonly used classifiers include Support Vector Machine (SVM), Random Forest (RF), and neural networks in deep learning (such as CNN and MLP). This classifier is trained using a training dataset to learn the feature patterns of different stages, thereby achieving classification prediction for unknown stages. The confidence score is the output of the stage classifier, representing the probability that the discriminant feature vector of each time period belongs to a certain life cycle stage. The value ranges from 0 to 1; the higher the confidence score, the greater the probability that the time period belongs to that stage.

[0075] Step S506: Select the category with the highest confidence level as the target lifecycle stage.

[0076] Specifically, in this embodiment, the confidence levels of different lifecycle stages output by the classifier for each time period are compared, and the stage with the highest confidence value is selected as the lifecycle stage corresponding to that time period, i.e., the target lifecycle stage. If the entire evaluation time window contains only one continuous time period, then the target lifecycle stage of that time period is the current lifecycle stage of the remediation area; if it contains multiple continuous time periods, then each time period corresponds to a target lifecycle stage, which can reflect the dynamic process of the remediation area.

[0077] In one embodiment of this example, step S504, based on mean characteristics, trend characteristics, and fluctuation characteristics, combines to form a discriminative feature vector for the corresponding time period, including steps S601 to S604: Step S601: Normalize the mean, trend, and volatility characteristics.

[0078] Specifically, in this embodiment, the normalization process here is similar to the standardization process in step S404. The purpose is to eliminate the differences in numerical range and dimensions between different types of features (mean, trend, fluctuation), ensure that each feature has equal importance in the discriminative feature vector, and avoid the classification result being dominated by some features with excessively large numerical ranges.

[0079] Step S602: According to the preset feature order, the normalized mean feature, trend feature and fluctuation feature are concatenated to generate an initial feature vector.

[0080] Specifically, in this embodiment, the preset feature order is a pre-defined feature arrangement order before model training. For example, it is arranged in the order of "mean of coupled features → mean of composite features → mean of risk basis features → mean of functional response features → trend of coupled features → trend of composite features → ... → fluctuation of functional response features" to ensure that the discriminant feature vectors of all time periods have a unified structure, which facilitates unified processing and learning by the stage classifier. The initial feature vector is a high-dimensional feature vector that has not undergone dimensionality reduction after concatenation, and its dimension is equal to the sum of the dimensions of the mean feature, trend feature, and fluctuation feature.

[0081] Step S603: Reduce the dimensionality of the initial feature vector by principal component analysis, remove redundant information and retain principal components.

[0082] Specifically, in this embodiment, Principal Component Analysis (PCA) is a commonly used dimensionality reduction algorithm. It transforms high-dimensional feature vectors into low-dimensional principal component vectors through linear transformation. Principal components are linear combinations of the original features, maximizing the preservation of information (variance) while eliminating redundant information (correlation) between features. Redundant information refers to highly correlated parts between different features; this information does not contribute additionally to classification and recognition but instead increases the computational complexity and risk of overfitting. Principal components are the low-dimensional feature vectors obtained after dimensionality reduction. Each principal component corresponds to a comprehensive index of the original features, and the first few principal components typically explain most of the information in the original features (e.g., over 95%). By using PCA to reduce the initial high-dimensional feature vectors to low-dimensional principal component vectors, key information is preserved while simplifying the model structure and improving classification efficiency and accuracy.

[0083] Step S604: Use the dimensionality-reduced feature vector as the discriminant feature vector, where the dimension of the discriminant feature vector is lower than the dimension of the initial feature vector.

[0084] Specifically, in this embodiment, the dimensionality-reduced feature vector is the principal component vector extracted by principal component analysis. Its dimension is determined according to the variance ratio explained by the principal components. For example, if the current three principal components can explain more than 95% of the information of the original features, then these three principal components are used as the discriminant feature vector, and the dimension is reduced from 12 dimensions to 3 dimensions.

[0085] In one embodiment of this example, based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, steps S701 to S709 are used to calculate the stage-adaptive ecological health score: Step S701: Establish a mapping relationship library between life cycle stages and ecological evaluation indicators. Different core indicators and auxiliary indicators correspond to different life cycle stages in the mapping relationship library.

[0086] Specifically, in this embodiment, the mapping relationship database is a database that stores the correspondence between different life cycle stages and corresponding ecological evaluation indicators, serving as the basis for dynamically selecting the evaluation indicator set. Core indicators are those that most directly and critically reflect the health status of the ecosystem at that stage, playing a decisive role and having a higher weight. Auxiliary indicators supplement the core indicators, further refining the ecological health assessment, and have a relatively lower weight.

[0087] Step S702: Based on the identified target life cycle stage, match the corresponding core indicators and auxiliary indicators from the mapping relationship library to form an initial ecological evaluation indicator set.

[0088] Specifically, in this embodiment, the initial ecological evaluation index set is obtained by directly matching from the mapping relationship library according to the target life cycle stage, and includes a set of all core indicators and auxiliary indicators for that stage.

[0089] Step S703: Calculate the actual observed values ​​of each initial indicator in the initial ecological evaluation indicator set based on the multimodal spatiotemporal feature sequence.

[0090] Specifically, in this embodiment, the actual observed value is the specific value of each initial evaluation index at the current evaluation time node (or the current time period), which is calculated or extracted from relevant feature data in the multimodal spatiotemporal feature sequence.

[0091] Step S704: Obtain the missing rate corresponding to the actual observation value, and select the initial index with a missing rate less than or equal to the preset missing threshold to form the final ecological evaluation index set.

[0092] Specifically, in this embodiment, the missing rate is the proportion of actual observed values ​​of each initial indicator that are missing within the evaluation time window. Data loss may be due to monitoring equipment malfunctions, limitations in observation conditions (such as continuous cloud cover), etc. Indicators with excessively high missing rates cannot accurately reflect the true state of the ecosystem and will affect the reliability of the health score. The preset missing threshold is a pre-set critical value for the missing rate used to screen indicators, for example, set to 20%. That is, indicators with a missing rate exceeding 20% ​​are judged as invalid indicators and removed from the initial indicator set; indicators with a missing rate less than or equal to 20% are retained. The final ecological evaluation indicator set is the set of valid indicators remaining after removing invalid indicators with excessively high missing rates. The indicator data in this set is of reliable quality and can provide effective support for the health score.

[0093] Step S705: Perform trend stability analysis on the spatiotemporal feature sequences corresponding to each target indicator in the final ecological evaluation indicator set, and obtain the corresponding spatiotemporal feature trend stability.

[0094] Specifically, in this embodiment, the target indicators are each specific indicator in the final ecological evaluation indicator set. Spatiotemporal characteristic trend stability refers to whether the spatiotemporal characteristic sequence of the target indicator shows a stable trend over time; that is, whether the indicator value exhibits a pattern of continuous increase, continuous decrease, or stable fluctuation, rather than random and drastic fluctuations. Higher trend stability indicates more stable ecological attributes reflected by the indicator and stronger indicator effectiveness. Trend stability analysis is achieved by calculating parameters such as the trend slope, fluctuation amplitude, and autocorrelation coefficient of the indicator's spatiotemporal characteristic sequence. For example, a trend slope close to 0 and a small fluctuation amplitude (small standard deviation) indicate a stable indicator trend; a large absolute value of the trend slope and a large fluctuation amplitude indicate an unstable indicator trend.

[0095] Step S706: Determine the effectiveness coefficient of the target indicator based on the stability of spatiotemporal characteristic trends.

[0096] Specifically, in this embodiment, the effectiveness coefficient is a coefficient that quantifies the effectiveness of the target indicator, ranging from 0 to 1. A higher effectiveness coefficient indicates a more stable spatiotemporal trend in the indicator, more accurately reflecting the health status of the ecosystem, and thus a higher weight in the health score. Conversely, a lower effectiveness coefficient indicates an unstable indicator trend, poorer effectiveness, and thus a lower weight. The effectiveness coefficient is determined based on the analysis results of the spatiotemporal trend stability, for example, using the following rules: when the trend stability parameter (such as the normalized value of the combined trend slope and fluctuation amplitude) is between 0.8 and 1.0, the effectiveness coefficient is 0.9-1.0; between 0.6 and 0.8, the effectiveness coefficient is 0.7-0.9; between 0.4 and 0.6, the effectiveness coefficient is 0.5-0.7; between 0.2 and 0.4, the effectiveness coefficient is 0.3-0.5; and between 0 and 0.2, the effectiveness coefficient is 0-0.3.

[0097] Step S707: Based on the effectiveness coefficient, match the influence weight of the corresponding target indicator.

[0098] Specifically, in this embodiment, the influence weight is the proportion of each target indicator in the overall ecological health score, with the sum of the weights being 1. The higher the weight, the greater the influence of the indicator on the ecological health. The matching process of influence weights combines the effectiveness coefficient of the indicator with the importance of the indicator in the current life cycle stage (core indicator or auxiliary indicator). The basic weight of core indicators is higher than that of auxiliary indicators. The basic weights are then adjusted according to the effectiveness coefficients to finally obtain the influence weight of each target indicator.

[0099] Step S708: Classify and quantify the actual observed values ​​of each target indicator.

[0100] Specifically, in this embodiment, the grading is based on the ecological significance and restoration goals of the target indicators, dividing their actual observed values ​​into several health levels, such as excellent, good, moderate, poor, and very poor. Each level corresponds to a specific numerical range, and the grading criteria can be: excellent (≥60%), good (40%-60%), moderate (20%-40%), poor (10%-20%), and very poor (<10%). For the "vegetation coverage" indicator during the functional stability period, the excellent standard may be raised to ≥80%. Quantitative scoring assigns a corresponding quantitative score (e.g., 0-100 points) to each health level, such as excellent (90-100 points), good (70-89 points), moderate (50-69 points), poor (30-49 points), and very poor (0-29 points). Based on the health level to which the actual observed value of each target indicator belongs, a corresponding quantitative score is assigned to obtain a single score for each indicator.

[0101] Step S709: Based on the impact weights and quantitative scores of each target indicator, calculate the stage-adaptive comprehensive score of ecological health.

[0102] Specifically, in this embodiment, the comprehensive ecological health score is calculated using a weighted summation method. This involves multiplying the quantitative score of each target indicator by its corresponding final influence weight, and then summing all the products to obtain the comprehensive score. The calculation formula is: Comprehensive Score = Σ (Quantitative score of a certain indicator × Final influence weight of that indicator). The comprehensive score ranges from 0 to 100 points, with higher scores indicating better ecosystem health: 90-100 points indicate excellent ecosystem structure and stable function; 70-89 points indicate good ecosystem condition with minor defects; 50-69 points indicate moderate ecosystem stability with room for improvement; 30-49 points indicate poor ecosystem with significant structural defects or functional impairment, facing the risk of degradation; and 0-29 points indicate severe ecosystem degradation requiring urgent restoration measures.

[0103] In one embodiment of this example, based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, steps S801 to S807 are used to calculate the stage-adaptive ecological health score: Step S801: Obtain historical monitoring data corresponding to the target life cycle stage, and extract the numerical sequence of each target indicator in the historical monitoring data.

[0104] Specifically, in this embodiment, historical monitoring data refers to other remediation areas of the same type and at the same target life cycle stage as the current remediation area, or monitoring data accumulated by the current remediation area in historical periods. This data contains long-term observation records of each target indicator and can reflect the historical change patterns of the indicators. The numerical sequence is extracted from the historical monitoring data, and is a sequence formed by arranging the observation values ​​of each target indicator at different time points in chronological order.

[0105] Step S802: Perform trend decomposition on the numerical sequence to separate the long-term trend component, seasonal cycle component and random fluctuation component.

[0106] Specifically, in this embodiment, trend decomposition employs time series analysis to divide the numerical sequence of the target indicator into three independent components, each reflecting different types of change patterns. The long-term trend component refers to the sustained upward, downward, or stable trend of the numerical sequence over a relatively long period, unaffected by short-term fluctuations and seasonal factors. The seasonal cycle component refers to the periodic fluctuations of the numerical sequence with seasonal changes. The random fluctuation component refers to the irregular fluctuations in the numerical sequence caused by random factors. Commonly used trend decomposition methods include Time Series Decomposition (STL) and X-12-ARIMA decomposition.

[0107] Step S803: Based on the long-term trend component, calculate the average rate of change of each target indicator during the target life cycle stage.

[0108] Specifically, in this embodiment, the average rate of change is the average rate of change of the long-term trend component of the target indicator within the target life cycle stage, reflecting the long-term development rate of the ecological attribute represented by the indicator. The calculation method is: Average rate of change = (Long-term trend component value at the end of the stage - Long-term trend component value at the beginning of the stage) / Stage duration. For example, if the initial value of the long-term trend component of a target indicator during the pioneer plant stage (lasting 3 years) is 0.5, and the final value is 0.8, then the average rate of change is (0.8 - 0.5) / 3 = 0.1 / year, indicating that the ecological attribute represented by the indicator improves by an average of 0.1 per year. A positive average rate of change indicates an improving trend in the ecological attribute; a negative rate indicates a deteriorating trend; and a rate close to 0 indicates a stable trend.

[0109] Step S804: Based on the seasonal cyclical component, calculate the cyclical stability coefficient of each target indicator within the target life cycle stage.

[0110] Specifically, in this embodiment, the periodic stability coefficient is a parameter that quantifies the seasonal periodic fluctuation stability of a target indicator, reflecting the adaptability of the ecological attributes represented by the indicator to seasonal changes. The calculation method is as follows: First, calculate the standard deviation of the seasonal periodic component. Then, subtract the ratio of the standard deviation to the maximum value of the periodic component from 1 (normalization processing) to obtain the periodic stability coefficient, which ranges from 0 to 1. For example, if the standard deviation of the seasonal periodic component of an indicator is 0.1 and the maximum value is 0.5, then the periodic stability coefficient = 1 - (0.1 / 0.5) = 0.8. The closer the periodic stability coefficient is to 1, the smaller the seasonal fluctuation of the indicator, the stronger the adaptability of the ecological attributes to seasonal changes, and the more stable the indicator; conversely, the closer the coefficient is to 0, the greater the seasonal fluctuation and the worse the stability of the indicator.

[0111] Step S805: Based on the random fluctuation component, calculate the anti-interference capability coefficient of each target indicator during the target life cycle stage.

[0112] Specifically, in this embodiment, the anti-interference capability coefficient is a parameter that quantifies the ability of a target indicator to resist random interference, reflecting the stability of the ecological attributes represented by the indicator when facing sudden disturbances. The calculation method is as follows: First, calculate the average value of the absolute values ​​of the random fluctuation components (average fluctuation amplitude). Then, subtract the ratio of the average fluctuation amplitude to the maximum value of the indicator's numerical sequence from 1 (normalization processing) to obtain the anti-interference capability coefficient, which ranges from 0 to 1. For example, if the average fluctuation amplitude of a certain indicator's random fluctuation component is 0.05 and the maximum value of its numerical sequence is 0.9, then the anti-interference capability coefficient = 1 - (0.05 / 0.9) ≈ 0.944. The closer the anti-interference capability coefficient is to 1, the less affected the indicator is by random interference, and the stronger the anti-interference capability of the ecological attributes; conversely, the closer the coefficient is to 0, the greater the impact of random interference on the indicator, and the weaker the anti-interference capability.

[0113] Step S806: Dimensionless processing is performed on the average rate of change, periodic stability coefficient, and anti-interference ability coefficient, and then input into the weight allocation function.

[0114] Specifically, in this embodiment, the dimensionless processing transforms the average rate of change, periodic stability coefficient (0-1), and anti-interference ability coefficient (0-1) into a unified dimensionless numerical range (e.g., 0-1), eliminating dimensional differences between different parameters and ensuring they can be comprehensively calculated in the weight allocation function. For the average rate of change, min-max normalization is used to map it to the 0-1 interval, where positive values ​​(improvement trend) are mapped to 0.5-1, negative values ​​(deterioration trend) are mapped to 0-0.5, and zero values ​​are mapped to 0.5. The weight allocation function is pre-constructed and can integrate the average rate of change, periodic stability coefficient, anti-interference ability coefficient, and effectiveness coefficient to output a mathematical function or model (e.g., multiple linear regression model, machine learning model) that initially influences the weights of the indicators. This function is calibrated using training data to ensure that the output weights can reasonably reflect the influence of each parameter on the importance of the indicator. The three dimensionless parameters are input into the weight allocation function to obtain preliminary weight calculation results.

[0115] Step S807: Generate the initial influence weights of each target indicator based on the output of the weight allocation function and the effectiveness coefficient.

[0116] Specifically, in this embodiment, the initial influence weight is obtained by multiplying the output of the weight allocation function with the validity coefficient determined in step S706 using a weighted average. The formula is: Initial influence weight = Output of weight allocation function × Validity coefficient. This process further combines the current validity (validity coefficient) of the indicator with its historical change characteristics (average rate of change, periodic stability, and anti-interference ability), ensuring that the initial weight reflects both the historical performance of the indicator and the reliability of the current data. For example, if the output of the weight allocation function is 0.25 and the validity coefficient is 0.9, then the initial influence weight = 0.25 × 0.9 = 0.225.

[0117] Step S808: Normalize the initial influence weights to generate the final influence weights.

[0118] Specifically, normalization refers to dividing the initial influence weight of all target indicators by the sum of all initial influence weights, so that the sum of the final influence weights is 1. This ensures that the weights of each indicator can be weighted and summed to calculate the comprehensive health score. The calculation formula is: Final influence weight = Initial influence weight of a certain indicator / Sum of initial influence weights of all indicators. For example, if the initial influence weights of three target indicators are 0.225, 0.175, and 0.1, the sum is 0.5, and the corresponding final influence weights are 0.225 / 0.5=0.45, 0.175 / 0.5=0.35, and 0.1 / 0.5=0.2, with a sum of 1. The final influence weight objectively reflects the relative importance of each target indicator in the ecological health score at the current life cycle stage.

[0119] In one embodiment of this example, based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, the stage-adaptive ecological health score is calculated in steps S901 to S904: Step S901: Normalize the average rate of change, periodic stability coefficient, and anti-interference capability coefficient to generate a normalized dynamic vector.

[0120] Specifically, in this embodiment, the normalized dynamic vector is a three-dimensional vector formed by combining the average rate of change, periodic stability coefficient, and anti-interference ability coefficient in a preset order after dimensionless processing, such as [normalized average rate of change, normalized periodic stability coefficient, normalized anti-interference ability coefficient]. This vector can comprehensively reflect the dynamic change characteristics (average rate of change), periodic stability, and anti-interference ability of the target indicator, providing structured data for subsequent feature combination and weight calculation. The normalization process adopts the min-max normalization method, mapping all three coefficients to the [0,1] interval to ensure that the numerical range of each element in the vector is consistent, which facilitates subsequent combination and model input.

[0121] Step S902: Combine the normalized dynamic vector with the effectiveness coefficient to generate the feature combination of the target index.

[0122] Specifically, in this embodiment, the feature combination is formed by concatenating a normalized dynamic vector (three-dimensional) with an effectiveness coefficient (one-dimensional), resulting in a four-dimensional feature vector, such as [normalized average rate of change, normalized periodic stability coefficient, normalized anti-interference ability coefficient, effectiveness coefficient]. This feature combination comprehensively integrates the dynamic change characteristics (average rate of change, periodic stability, anti-interference ability) and current effectiveness (effectiveness coefficient) of the target indicator, fully reflecting the comprehensive attributes of the target indicator and providing comprehensive and accurate input features for the weight allocation function.

[0123] Step S903: Input the feature combination into the weight allocation function.

[0124] Specifically, in this embodiment, after the feature combination of the target indicator is input into the function, the function can output the initial weight prediction value corresponding to the indicator based on the learned mapping relationship.

[0125] Step S904: The weight allocation function outputs the initial weight values ​​of the target index through the mapping relationship.

[0126] Specifically, in this embodiment, the mapping relationship refers to the correspondence between the feature combination and the initial weight of the indicator, which is learned by the weight allocation function through training. This relationship can be linear or non-linear, depending on the intrinsic association between the features and the weights. For example, if the feature combination of a target indicator is [0.8, 0.75, 0.9, 0.85], the weight allocation function will output an initial weight value of 0.28 after calculating through the internal mapping relationship.

[0127] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for assessing the entire life cycle of a remediation area based on multi-source remote sensing and ground monitoring, characterized in that, include: Acquire multi-source data of the target remediation area within the assessment time window. The multi-source data includes satellite remote sensing data, UAV or near-ground remote sensing data, and ground monitoring data. Spatiotemporal alignment, quality control, and data repair are performed on multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scale. Based on multi-source benchmark datasets, multimodal feature extraction and fusion are performed to generate multimodal spatiotemporal feature sequences that comprehensively represent the ecological state of the restoration area; Based on multimodal spatiotemporal feature sequences, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through a stage identification model; Based on the identified life cycle stages, dynamically select a set of ecological evaluation indicators that are appropriate for the target life cycle stage; Based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, a stage-adaptive ecological health score is calculated. Based on multimodal spatiotemporal feature sequences, life cycle stages, and ecological health scores, the changing trend of the ecological status of the restoration area is predicted; Based on the changing trends, determine whether there is a risk of degradation and output spatial ecological risk early warning information.

2. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 1, characterized in that, Spatiotemporal alignment, quality control, and data repair are performed on multi-source data to generate a multi-source benchmark dataset with unified spatiotemporal scales, including: Temporal interpolation and spatial resampling are performed on satellite remote sensing data, UAV or near-ground remote sensing data and ground monitoring data, and unified to the preset analysis time step and spatial resolution grid. Identify and label cloud-occluded pixels, sensor outliers, and spatially discontinuous areas in the data; For cloud-occluded pixels and spatially discontinuous areas, multi-temporal data and adjacent spatial information are integrated for collaborative interpolation and repair. Radiometric normalization and topographic correction were performed on the restored data from different sources to eliminate the effects of differences in observation conditions and topography. All corrected data are mapped to the same spatiotemporal grid to form a multi-source benchmark dataset.

3. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 2, characterized in that, For cloud-occluded pixels and spatially discontinuous areas, collaborative interpolation and restoration are performed by fusing multi-temporal data and neighboring spatial information, including: Construct a spatiotemporal cube centered on the target pixel. The spatiotemporal cube contains multi-period data of the target pixel in the time dimension and neighboring pixel data in the spatial dimension. If the target pixel is completely missing in a single time phase due to cloud obstruction, the spatiotemporal kriging interpolation method is used to estimate it using the value of the effective pixel within the spatiotemporal cube. If the target pixel is missing in multiple consecutive time phases, the surface phenological change curve is introduced as a priori constraint, and iterative repair is carried out in combination with the temporal pattern of similar pixel groups.

4. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 1, characterized in that, Based on multi-source benchmark datasets, multimodal feature extraction and fusion are performed to generate a multimodal spatiotemporal feature sequence that comprehensively represents the ecological state of the restoration area, including: Extract time-series vegetation index, surface temperature, surface deformation, and land use type characteristics from satellite remote sensing data; Extract high-resolution vegetation cover, canopy height, and micro-topographic erosion features from UAV or near-ground remote sensing data; Extract time-series data of soil physicochemical properties, hydrological parameters, and biodiversity survey characteristics from ground monitoring data; All extracted features are standardized and denoised, and then aligned according to the time series. An attention mechanism was used to fuse time-series data of vegetation index, land surface temperature and soil physicochemical properties across modes to generate coupled features characterizing environmental stress. By spatially correlating and integrating high-resolution vegetation cover, canopy height, and biodiversity survey characteristics, composite features representing ecosystem structure are generated. Trend analysis and overlay of surface deformation characteristics and micro-topographic erosion characteristics are performed to generate risk basement characteristics that characterize surface stability. Based on hydrological parameters and land use type characteristics, we calculate the hydrological regulation capacity change index and generate functional response characteristics that characterize ecological functions. By integrating coupling features, composite features, risk-based features, and functional response features, a multimodal spatiotemporal feature sequence is formed.

5. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 4, characterized in that, Based on multimodal spatiotemporal feature sequences, the life cycle stage of the repair area is automatically divided and output as the target life cycle stage through a stage identification model, including: Multimodal spatiotemporal feature sequences are input into a time-series segmentation model to initially detect time points where significant changes in ecological status occur; Based on time nodes, the entire time series is divided into multiple consecutive time periods; Extract the mean, trend, and fluctuation features of the multimodal spatiotemporal feature sequence within each time period; Based on mean characteristics, trend characteristics, and fluctuation characteristics, a discriminative feature vector for the corresponding time period is formed by combining them. The discriminant feature vector for each time period is input into the stage classifier, and the confidence score of each time period belonging to different life cycle stages is output. The category with the highest confidence level is selected as the target lifecycle stage.

6. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 5, characterized in that, Based on mean characteristics, trend characteristics, and fluctuation characteristics, the discriminant feature vector for the corresponding time period is formed by combining the following: The mean, trend, and volatility characteristics are normalized. According to the preset feature order, the normalized mean feature, trend feature and fluctuation feature are concatenated to generate an initial feature vector. Principal component analysis is used to reduce the dimensionality of the initial eigenvectors, remove redundant information, and retain the principal components. The dimensionality-reduced feature vectors are used as discriminant feature vectors, where the dimension of the discriminant feature vectors is lower than that of the initial feature vectors.

7. The method for full life-cycle assessment of remediation areas based on multi-source remote sensing and ground monitoring according to claim 1, characterized in that, Based on the ecological evaluation index set and the corresponding multimodal spatiotemporal feature sequence, a stage-adaptive ecological health score is calculated: Establish a mapping relationship library between life cycle stages and ecological evaluation indicators, with different core and auxiliary indicators corresponding to different life cycle stages in the mapping relationship library; Based on the identified target life cycle stage, the corresponding core indicators and auxiliary indicators are matched from the mapping relationship library to form an initial ecological evaluation indicator set; The actual observed values ​​of each initial indicator in the initial ecological evaluation index set are calculated based on multimodal spatiotemporal feature sequences. Obtain the missing rate corresponding to the actual observation value, and select the initial indicators with a missing rate less than or equal to the preset missing threshold to form the final ecological evaluation indicator set; Trend stability analysis was performed on the spatiotemporal characteristic sequences corresponding to each target indicator in the final ecological evaluation indicator set, and the corresponding spatiotemporal characteristic trend stability was obtained. Based on the stability of spatiotemporal characteristics, the effectiveness coefficient of the target indicator is determined; Based on the effectiveness coefficient, match the influence weight of the corresponding target indicator; The actual observed values ​​of each target indicator are classified and quantitatively scored. Based on the impact weights and quantitative scores of each target indicator, a phase-adaptive comprehensive score for ecological health is calculated.

8. The method according to claim 7, characterized in that, Based on the effectiveness coefficient, the influence weights of the corresponding target indicators include: Obtain historical monitoring data corresponding to the target life cycle stage, and extract the numerical sequence of each target indicator in the historical monitoring data; Trend decomposition is performed on the numerical sequence to separate the long-term trend component, seasonal cycle component, and random fluctuation component. Based on the long-term trend component, calculate the average rate of change of each target indicator during the target life cycle stage. Based on the seasonal cyclical component, the cyclical stability coefficient of each target indicator is calculated within the target life cycle stage. Based on random fluctuation components, calculate the anti-interference capability coefficient of each target indicator during the target life cycle stage; The average rate of change, periodic stability coefficient, and anti-interference ability coefficient are dimensionless and then input into the weight allocation function. Based on the output of the weight allocation function and the effectiveness coefficient, the initial influence weights of each target indicator are generated. The initial influence weights are normalized to generate the final influence weights.

9. The method according to claim 8, characterized in that, Based on the output of the weight allocation function and the effectiveness coefficient, the initial influence weights for each target indicator are generated, including: The average rate of change, periodic stability coefficient, and anti-interference ability coefficient are normalized to generate a normalized dynamic vector. The normalized dynamic vector is combined with the effectiveness coefficient to generate a feature combination of the target index; The feature combination is input into the weight allocation function; The weighting function outputs the initial weight values ​​of the target index through a mapping relationship.

Images