Ecological organic fertilizer grassland restoration effect evaluation method based on multi-source data

Abstract

The invention discloses an ecological organic fertilizer grassland restoration effect evaluation method based on multi-source data, and the method comprises the steps: collecting grassland monitoring data and environment data of a preset restoration region, carrying out the extraction of ecological vegetation features of the grassland monitoring data and the environment data through combining a soil vegetation atmosphere coupling mechanism, and obtaining the ecological features of the grassland, performing three-level space-time collaborative classification on the grassland ecological characteristics to obtain time node data; performing spectral texture physiological feature coupling on the time node data through three-dimensional convolution to obtain fusion features, and constructing a three-dimensional index system based on humus, carbon sink and biodiversity to obtain an ecological grassland restoration index; and according to the fusion features and the ecological grassland restoration index, constructing an ecological organic fertilizer grassland restoration effect evaluation model based on machine learning, fusing mobile phone signaling data and animal husbandry points to optimize the ecological organic fertilizer grassland restoration effect evaluation model, and obtaining an evaluation result.

Description

Technical Field

[0001] This invention relates to the field of grassland restoration assessment, and more particularly to a method for assessing the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data. Background Technology

[0002] Grassland ecosystems, as an important component of terrestrial ecosystems, play crucial ecological roles such as soil and water conservation, carbon sequestration and oxygen release, and maintaining biodiversity. However, affected by multiple factors such as overgrazing, climate change, and human disturbance, grasslands in many parts of the world are facing degradation problems, with prominent phenomena such as declining soil fertility, reduced vegetation cover, and ecological function decline, which seriously restrict ecological balance and the sustainable development of animal husbandry.

[0003] Organic fertilizers, with their advantages of improving soil structure, increasing soil humus content, and promoting vegetation growth, have become an important technical means for grassland restoration. However, current grassland restoration effectiveness assessments have significant shortcomings: traditional assessments often rely on single monitoring data, lacking integration of multi-dimensional data on soil, vegetation, and atmosphere; the assessment indicator system is not comprehensive enough, failing to fully consider core ecological elements such as humus conversion, carbon sequestration capacity, and biodiversity; and dynamic influencing factors such as human disturbance have not been effectively quantified, resulting in insufficient accuracy and timeliness of assessment results.

[0004] With the development of technologies such as satellite remote sensing, UAV mapping, and IoT sensing, acquiring multi-source data has become possible. However, how to effectively integrate data of different scales and types and construct a scientific evaluation model remains a technical challenge in the field of grassland restoration assessment. Therefore, there is an urgent need to establish a grassland restoration effectiveness evaluation method that integrates multi-source data and takes into account both core ecological indicators and dynamic disturbance factors, providing scientific support for the optimization of grassland ecological restoration programs and sustainable management. Summary of the Invention

[0005] The purpose of this invention is to provide a method for evaluating the effectiveness of ecological organic fertilizer in grassland restoration based on multi-source data.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Grassland monitoring data and environmental data are collected from the pre-defined restoration area, and the grassland monitoring data and environmental data are preprocessed. The grassland monitoring data includes soil physicochemical data and vegetation ecological data. The vegetation ecological data includes UAV hyperspectral data, satellite remote sensing data, biomass, spectral reflectance, and indicator species surveys. The environmental data includes meteorological data, topographic data, and human disturbance data. The soil physicochemical data includes humus content, pH value and ion concentration, and organic matter conversion rate. By combining the soil-vegetation-atmosphere coupling mechanism, ecological vegetation features are extracted from the grassland monitoring data and the environmental data to obtain grassland ecological features, and time node data are obtained by performing three-level spatiotemporal collaborative classification of the grassland ecological features. By coupling spectral texture and physiological features of the time node data through three-dimensional convolution, a fusion feature is obtained, and a three-dimensional index system based on humus, carbon sink, and biodiversity is constructed to obtain the ecological grassland restoration index. Based on the fusion features and the ecological grassland restoration index, a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration is constructed. The model is then optimized by integrating mobile phone signaling data and livestock point data. The data to be evaluated is input into the optimized model, and the evaluation results are output.

[0007] Furthermore, the method for obtaining the ecological characteristics of the grassland includes: Based on the soil, vegetation, and atmosphere continuum theory, core driving parameters were selected, including humus content, active aluminum content, soil moisture conductivity, red-edge band reflectivity, interannual NDVI coefficient of variation, and grazing intensity. Soil physicochemical characteristics, such as humus conversion efficiency, soil carbon pool activity index, and ion balance index, were obtained by feature extraction from soil physicochemical data. Physiological and structural characteristics of vegetation were extracted. Physiological characteristics included red edge position shift and photochemical reflectance index; structural characteristics included leaf area index and community height. A three-dimensional feature matrix of soil, vegetation, and atmosphere was constructed to obtain fused coupling features, which include soil-driven features, vegetation response features, and environmental regulation features. Soil-driven features include humus conversion efficiency, soil carbon pool activity index, and ion balance index. Vegetation response features include REP offset, PRI value, LAI, and community height. Environmental regulation features include slope, soil moisture, and grazing intensity. The quarterly change rate of NDVI was calculated, and the water-limiting period was identified by combining soil moisture data. The Mann-Kendall test was used to analyze the 5-year NDVI time series data and synchronously correlated the annual increase of humus to obtain the time dimension characteristics. The spatial variation coefficient of humic content and LAI was calculated using grids as units. Based on the coupled feature matrix, K-means clustering was used to divide functional zones and obtain spatial dimensional features. The functional zones include high-efficiency remediation zone, medium-efficiency remediation zone and low-efficiency remediation zone. The reflectance of the red-edge band, the normalized differential red-edge index, and the chlorophyll absorption characteristics were extracted as spectral features; the contrast and entropy values ​​were calculated based on the gray-level co-occurrence matrix of the RGB image to obtain texture features. The soil physicochemical characteristics, physiological characteristics, structural characteristics, fusion and coupling characteristics, temporal dimension characteristics, spatial dimension characteristics, spectral characteristics, and texture characteristics are output as grassland ecological characteristics.

[0008] Furthermore, the method for obtaining time node data by performing three-level spatiotemporal collaborative classification of the grassland ecological characteristics includes: For macro satellite data: the Fmask algorithm is used to mark cloud pixels, the cloud-covered area is filled with linear interpolation, the maximum value of the effective images each month is calculated to synthesize NDVI, and a monthly macro vegetation cover map with a resolution of 30m is generated. The average NDVI of the 12 months before restoration is used as the benchmark to calculate the relative change rate of each month. For meso-level UAV data: using satellite imagery as a benchmark, SIFT feature matching is employed to control the UAV image registration error within one pixel; pseudo-invariant feature regions in the UAV imagery are selected, and reflectance is corrected to the satellite spectral response function through linear regression, ensuring that the reflectance deviation in the red-edge band is ≤3%; the UAV hyperspectral data is downsampled to 10m resolution and overlaid with the Sentinel-2 quarterly NDVI map to generate a composite layer of spectrum and vegetation index; For micro-level IoT data: the 3σ criterion is used to filter out sensor outliers, retaining time periods with a valid data ratio >90%; the moving window mean is calculated for each quarter's soil moisture data to generate a quarterly-scale moisture dataset; ground quadrat spectral data and UAV hyperspectral pixels are correlated through spatial buffer analysis; each quadrat contains at least 16 UAV pixels; Multiple hyperspectral indices were calculated, and texture parameters were extracted from UAV true-color images using gray-level co-occurrence matrix. Based on soil moisture data, water stress index was calculated and coupled with NDVI change rate to generate vegetation physiological state index. Hyperspectral indices, texture parameters and vegetation physiological state index were used as features to be fused. A convolutional long short-term memory network is used to perform three-dimensional feature fusion of the features to be fused, a three-level classification rule is constructed for dynamic classification decision-making, and the fused features of the four quarters of each year are clustered in time series to extract the cluster centers as key time node data. A Bayesian network is used to integrate the three-level classification results, conditional probabilities are defined, key time node data are matched with the three-level classification rules to generate a quarterly dynamic classification map, and the data corresponding to the time nodes in the quarterly dynamic classification map are output as time node data.

[0009] Furthermore, a method for obtaining fused features by coupling spectral texture and physiological features of the time-node data through three-dimensional convolution includes: A quantitative correlation between physiological characteristics and spectral / texture features was established by using a sequential chain mechanism of soil humus content, root water conductivity, and vegetation canopy structure; structural equation modeling was introduced to quantify the path coefficients of the chain. To address the temporal resolution differences between UAV spectral and soil sensor data, the DTW algorithm is employed to align high-frequency time-series data of physiological features with low-frequency data of spectral / texture features. The UAV hyperspectral data was used to eliminate atmospheric scattering by the dark target method, and atmospheric correction satellite data was used to select key bands from 125 bands based on the continuous projection algorithm to standardize physiological characteristics. The UAV hyperspectral imagery and satellite imagery were registered to the UTM coordinate system using a random sampling consistency algorithm. Soil sampling points were used to generate physiological feature raster maps through inverse distance weighted interpolation. Using the UAV flight date as the reference time node, the soil sensor data were aligned to this node using the DTW algorithm to form a quarterly physiological feature sequence. The spectral features, texture features, and physiological features are input into a convolutional long short-term memory network to extract coupled features. The weight vector of the coupled features is calculated using the sigmoid activation function to enhance the contribution of the red edge band and soil moisture. A 2D spatial mask is generated to assign 1.2 times the weight to the features of densely vegetated areas. Principal component analysis was used to reduce the dimensionality of the 128-dimensional coupled features to 32-dimensional features, and then the features were fused with the three-dimensional index system to form fused features.

[0010] Furthermore, a method for constructing a three-dimensional index system based on humus, carbon sinks, and biodiversity to obtain the ecological grassland restoration index includes: The three-dimensional indicators are process, state, and response dynamic indicators. Process indicators include soil organic carbon mineralization rate, nitrogen conversion efficiency, and carbon input from root exudates. State indicators include interannual NDVI coefficient of variation, soil organic carbon density, and dominant species coverage ratio. Response dynamic indicators include vegetation recovery rate under drought stress, functional diversity index, and microbial biomass carbon / organic carbon ratio. The process indicators, state indicators, and response dynamic indicators are standardized. A judgment matrix is ​​constructed based on the importance scores of the indicators by ecological restoration experts, and the subjective weights of the process, state, and response indicators are calculated. The objective weights are calculated based on the information entropy of the indicators, and the subjective and objective weights are weighted and summed to obtain the combined weights. The ecological grassland restoration index is obtained by weighting and summing the three-dimensional indicators using combined weights.

[0011] Furthermore, the method for constructing a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on the fusion features and the ecological grassland restoration index includes: An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on extreme gradient boosting and random forest is constructed. Logistic regression is used as the meta-model, fusing the predicted ecological grassland restoration index output by extreme gradient boosting with the random forest correction coefficient. The expression is as follows:

[0012] in This is a preliminary prediction of the ecological grassland restoration index. These are the random forest correction coefficients. The integrated ecological grassland restoration index; When the integrated ecological grassland restoration index is between 0.8 and 1, the restoration level is excellent; when the integrated ecological grassland restoration index is between 0.6 and 0.8, the restoration level is good; when the integrated ecological grassland restoration index is between 0.4 and 0.6, the restoration level is medium; when the integrated ecological grassland restoration index is below 0.4, the restoration level is poor. The fused grassland restoration index and the weighted sum of the loss function are used as the objective function. SMOTE oversampling is used to divide the dataset into training and test sets. The hyperparameters are adjusted by Bayesian optimization and iterated until the error between the grassland restoration index in the test set and the predicted value of the grassland restoration index is minimized.

[0013] Furthermore, the method for integrating mobile signaling data with livestock point data to optimize the ecological organic fertilizer grassland restoration effectiveness evaluation model includes: By acquiring mobile phone signaling data and livestock point data, fusing livestock point density estimates, and calculating spatial distribution intensity, the following methods are employed:

[0014] in Spatial distribution intensity, Let i be the weight of the i-th livestock point. For kernel function bandwidth, Let i be the coordinates of the i-th point of interest. The number of livestock farms, Spatial distribution intensity; Based on the spatiotemporal distribution of active users in mobile signaling, a monthly interference intensity curve is constructed, expressed as:

[0015] in The time t disturbance index, This refers to the number of active users in the current month. This is the historical average for the same period. A spatiotemporal weighted average is used to fuse spatial distribution intensity and temporal interference index to obtain a comprehensive interference index, in which the entropy weight method is used to determine the weights; Pearson correlation coefficient was used to analyze the relationship between the comprehensive disturbance index and vegetation indicators, and restoration areas were divided according to the comprehensive disturbance index value. A comprehensive disturbance index is added to the input features of the ecological organic fertilizer grassland restoration effectiveness evaluation model to form a fusion feature set of ecological features and disturbance features; An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration was trained using a fusion feature set. The evaluation results before and after correction were compared using ground-measured data. The comprehensive interference index threshold was recalculated quarterly based on newly collected mobile phone signaling and point of interest data.

[0016] The beneficial effects of this invention are: This invention is a method for evaluating the effectiveness of ecological organic fertilizer in grassland restoration based on multi-source data. Compared with existing technologies, this invention has the following technical advantages: This invention employs a series of steps including preprocessing, ecological vegetation feature extraction, three-level spatiotemporal collaborative classification, coupling of spectral texture and physiological features, model construction, and model optimization. It comprehensively integrates multi-source data, covering soil, vegetation, atmosphere, and human interference, thus avoiding biased assessments. Feature extraction is precise, with deep coupling of multi-dimensional features to capture core ecological response patterns. The scientific three-dimensional indicator system considers the restoration process, state, and response, reflecting overall effectiveness. The optimized machine learning model incorporates corrections for interfering factors, resulting in higher assessment accuracy. The spatiotemporal collaboration is highly efficient, supporting dynamic assessment, adapting to multiple scenarios, and possessing strong universality. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the ecological organic fertilizer grassland restoration effectiveness evaluation method based on multi-source data of the present invention. Detailed Implementation

[0018] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0019] The present invention provides a method for evaluating the effectiveness of ecological organic fertilizer in grassland restoration based on multi-source data, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Grassland monitoring data and environmental data are collected from the pre-defined restoration area, and the grassland monitoring data and environmental data are preprocessed. The grassland monitoring data includes soil physicochemical data and vegetation ecological data. The vegetation ecological data includes UAV hyperspectral data, satellite remote sensing data, biomass, spectral reflectance, and indicator species surveys. The environmental data includes meteorological data, topographic data, and human disturbance data. The soil physicochemical data includes humus content, pH value and ion concentration, and organic matter conversion rate. In the actual assessment, a grassland degradation restoration area was selected as the pre-set assessment area. This area has an altitude of 1050-1200m, a slope of 3°-8°, and belongs to the temperate continental climate zone, with an average annual precipitation of 320mm and an average annual temperature of 2.3℃. The restoration area is 500hm². 2 It is divided into 3 restoration sub-regions (Region A, 180 hm²). 2 Area B, 170 hm 2 Area C, 150 hm 2 ) and 1 unrepaired control area (50 hm) 2 The remediation measure is to apply weathered lignite-based organic fertilizer (humic acid content 72%) at a rate of 2250 kg / hm². 2 Apply in mid-May each year and mow once a year (late August). Five soil samples from the 0-20cm soil layer were collected diagonally within a 1m×1m quadrat. The samples were mixed and sieved through a 2mm sieve, and then separated into fresh and air-dried samples. Three replicate plots were set up in each remediation area, and two additional sampling points were required in the control area. Humus content: determined by the burning method, with simultaneous recording of rapid detection results by the soaking method; pH value and ion concentration: pH value was determined by the potentiometric method, and the content of exchangeable calcium and magnesium ions and active aluminum was detected by atomic absorption spectrometry; Organic matter conversion rate: the carbon-nitrogen ratio was determined by an elemental analyzer, and the 34.8% carbon loss due to microbial decomposition was deducted from the results of indoor culture experiments. UAV hyperspectral imaging: Using a Cucumber S185 airborne imager, at a flight altitude of 100m, with a 70% forward / lateral overlap, 2.5cm resolution orthophotos were acquired. Key bands included red edges for vegetation vitality monitoring and near-infrared biomass retrieval. Satellite remote sensing: Landsat-8 / 9 (30m, monthly NDVI) and Sentinel-2 (10m, quarterly vegetation cover). Radiometric normalization was required to eliminate sensor differences. Biomass: All vegetation within the sample plots was collected using the harvesting method, and both fresh and dry weights were measured. The average fresh weight of the un-weeded areas was approximately 180 g / m². 2 Approximately 100g / m² of residue remains after mowing. 2 ; Spectral reflectance: Data were collected 1m vertically from the ground using an ASD HandHeld 2 spectrometer (350-1075nm), with 5 measurements taken for each quadrat and the average taken. The data were then matched and verified with UAV spectral data; Indicator species survey: Coverage of dominant species such as Stipa kryori and Artemisia frigida were recorded, and community composition was statistically analyzed using the transect method. Meteorological and topographic data: Automatic weather stations record precipitation and temperature; Beidou positioning instruments acquire altitude and slope; soil moisture sensors monitor the moisture content of the 0-10cm soil layer to correct for the moisture effect on vegetation indices. Human-caused interference data: Grazing intensity: A spatiotemporal interference model was constructed using POI (points of interest) and mobile phone signaling data to quantify sheep units / hm. 2 Stress on vegetation; Management measures records: log of agricultural activities such as organic fertilizer application amount, fertilization time, and weeding frequency; Soil physicochemical data: Sampling points were set up in a 100m×100m grid in each sub-region and control area, for a total of 65 sampling points. At each sampling point, five soil samples from the 0-20cm soil layer were collected diagonally within a 1m×1m quadrat. After mixing, the samples were sieved through a 2mm sieve and separated into fresh and air-dried samples. The humus content was determined using the ignition method. The average values ​​before remediation in areas A, B, and C were 18.2 g / kg, 17.5 g / kg, and 16.8 g / kg, respectively, while the control area had 21.3 g / kg. The pH value was determined using the potentiostatic method, ranging from 7.8 to 8.5. The exchangeable calcium and magnesium ion concentrations were detected by atomic absorption spectrometry. The average calcium ion concentration was 185 mg / kg, the average magnesium ion concentration was 42 mg / kg, and the average active aluminum content was 12.3 mg / kg. The carbon-nitrogen ratio was determined by elemental analysis. Combined with the indoor culture experiment, after deducting 34.8% for microbial decomposition carbon loss, the organic matter conversion rate was calculated, with an average of 31.2% before remediation.

[0020] Vegetation ecological data: Using a Curt S185 airborne imager at an altitude of 100m and a 70% forward / lateral overlap, 2.5cm resolution hyperspectral images were acquired. The average reflectance of the red-edge band in the key band was 0.32 in area A, 0.30 in area B, and 0.28 in area C. Landsat-8 / 9 satellite data was used to acquire monthly NDVI at 30m resolution, and Sentinel-2 satellite data was used to acquire quarterly vegetation cover at 10m resolution. The average NDVI for the 12 months prior to restoration was 0.42 in area A, 0.40 in area B, 0.38 in area C, and 0.55 in the control area. All vegetation within the sample plots was collected using a harvesting method, and both fresh and dry weights were measured. The average fresh weight of the un-weeded area before restoration was approximately 175 g / m². 2 Approximately 95g / m² of residue remained after mowing. 2 Spectral reflectance was collected using an ASD HandHeld 2 spectrometer (350-1075nm) at a height of 1m from the ground, with 5 measurements taken for each quadrat and the average value taken. The coverage of dominant species such as Stipa krusei and Artemisia frigida was recorded, and the community composition was statistically analyzed using the transect method. Before restoration, the coverage of Stipa krusei was 18% in area A, 16% in area B, and 15% in area C. Meteorological data: Three automatic weather stations are deployed in the area to record hourly precipitation and temperature. The average precipitation over the past five years is 318 mm, and the average annual temperature is 2.5℃. Soil moisture sensors sample hourly to monitor the moisture content of the 0-10 cm soil layer, with a quarterly average between 12% and 25%. Topographic data: Elevation and slope were obtained using Beidou positioning devices with an accuracy of ±0.5°. The elevation of the area is concentrated between 1060-1180m, and the slope is mainly between 4° and 6°. Human interference data: The distribution of livestock points was obtained through Points of Interest (POIs). There are a total of 23 livestock points in the area, with an average density of 0.046 points / hm². 2 The number of active users is counted based on mobile signaling data. The average number of monthly active users is 860, while the historical average for the same period is 720. Logs of agricultural activities such as organic fertilizer application amount, fertilization time, and weeding frequency are recorded. Preprocessing: Outliers were removed from soil physicochemical data, and indicators such as organic matter conversion rate and exchangeable calcium and magnesium ion concentration were normalized to the [0,1] interval through logarithmic transformation; UAV hyperspectral imagery was matched with ground-measured spectral reflectance after atmospheric and geometric correction; satellite remote sensing data was normalized to eliminate sensor differences and generate a monthly NDVI time series dataset; meteorological data and soil moisture sensor data were aligned through time series matching to the time nodes for vegetation feature extraction; extreme values ​​in soil pH and NDVI were removed using interquartile range. By combining the soil-vegetation-atmosphere coupling mechanism, ecological vegetation features are extracted from the grassland monitoring data and the environmental data to obtain grassland ecological features, and time node data are obtained by performing three-level spatiotemporal collaborative classification of the grassland ecological features. By coupling spectral texture and physiological features of the time node data through three-dimensional convolution, a fusion feature is obtained, and a three-dimensional index system based on humus, carbon sink, and biodiversity is constructed to obtain the ecological grassland restoration index. Based on the fusion features and the ecological grassland restoration index, a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration is constructed. The model is then optimized by integrating mobile phone signaling data and livestock point data. The data to be evaluated is input into the optimized model, and the evaluation results are output.

[0021] In this embodiment, the method for obtaining the grassland ecological characteristics includes: Based on the soil, vegetation, and atmosphere continuum theory, core driving parameters were selected, including humus content, active aluminum content, soil moisture conductivity, red-edge band reflectivity, interannual NDVI coefficient of variation, and grazing intensity. Soil physicochemical characteristics, including humus transformation efficiency, soil carbon pool activity index, and ion balance index, were obtained by feature extraction from soil physicochemical data. The expressions are as follows:

[0022]

[0023]

[0024] in For humus conversion efficiency, The soil carbon pool activity index, This is the ion balance index. The content of humus after restoration, The humus content of undegraded grassland. Soil microbial biomass carbon content, Soil organic carbon content, This refers to the content of exchangeable calcium per unit mass of soil. This refers to the magnesium content per unit mass of soil. The content of active aluminum per unit mass of soil; Physiological and structural characteristics of vegetation were extracted. Physiological characteristics included red edge position shift and photochemical reflectance index; structural characteristics included leaf area index and community height. A three-dimensional feature matrix of soil, vegetation, and atmosphere was constructed to obtain fused coupling features, which include soil-driven features, vegetation response features, and environmental regulation features. Soil-driven features include humus conversion efficiency, soil carbon pool activity index, and ion balance index. Vegetation response features include REP offset, PRI value, LAI, and community height. Environmental regulation features include slope, soil moisture, and grazing intensity. The quarterly change rate of NDVI was calculated, and the water-limiting period was identified by combining soil moisture data. The Mann-Kendall test was used to analyze the 5-year NDVI time series data and synchronously correlated the annual increase of humus to obtain the time dimension characteristics. Spatial variation coefficients of humic content and LAI were calculated using grid units. Based on the coupled feature matrix, K-means clustering was used to divide functional zones and obtain spatial dimensional features. The functional zones included high-efficiency remediation zone, medium-efficiency remediation zone, and low-efficiency remediation zone. High-efficiency remediation zone: humic conversion efficiency > 25%, photochemical reflectance index > 0.3, leaf area index > 3.0; medium-efficiency remediation zone: humic conversion efficiency 15%-25%, photochemical reflectance index between 0.1 and 0.3, leaf area index between 1.5 and 3.0; low-efficiency remediation zone: humic conversion efficiency < 15%, photochemical reflectance index < 0.1, leaf area index < 1.5. The reflectance of the red-edge band, the normalized differential red-edge index, and the chlorophyll absorption characteristics were extracted as spectral features; the contrast and entropy values ​​were calculated based on the gray-level co-occurrence matrix of the RGB image to obtain texture features. The soil physicochemical characteristics, physiological characteristics, structural characteristics, fusion and coupling characteristics, temporal dimension characteristics, spatial dimension characteristics, spectral characteristics, and texture characteristics are output as grassland ecological characteristics. In actual assessment, soil physicochemical data are standardized: outliers are removed from soil physicochemical data such as humus content, pH value, and ion concentration, and indicators such as organic matter conversion rate and exchangeable calcium and magnesium ion concentration are normalized to the [0,1] interval through logarithmic transformation. Spatiotemporal matching of vegetation ecological data: After atmospheric and geometric correction, the hyperspectral images of UAVs are matched with the measured spectral reflectance of the ground for band matching; satellite remote sensing data are normalized to eliminate sensor differences and generate monthly NDVI time series datasets. Environmental data coupling: meteorological data and soil moisture sensor data are aligned through time series and matched to the time nodes for vegetation feature extraction; The red edge position shift was calculated using hyperspectral differential spectroscopy. The REP of healthy vegetation is located at 720-730 nm, and shifts towards shorter wavelengths under stress. The photochemical reflectance index reflects the photosynthetic efficiency of vegetation. The leaf area index was retrieved based on the NDVI of UAV hyperspectral images and the red edge vegetation index. The community height was obtained by generating a digital surface model through stereo matching of UAV visible light images and subtracting the digital elevation model. The grid is 5m × 5m. The Joman-Kendall test trend test statistic is greater than 1.96, indicating a significant improvement in vegetation. Based on the soil, vegetation, and atmosphere continuum theory, humus content, active aluminum content, soil moisture conductivity, red-edge reflectance, NDVI interannual variation coefficient, and grazing intensity were selected as core driving parameters. The mean value of the NDVI interannual variation coefficient before restoration was 18.5%, and the mean value of soil moisture conductivity was 0.08 cm / s. Soil physicochemical characteristics: Humus conversion efficiency, soil carbon pool activity index, and ion balance index were calculated. Before remediation, the humus conversion efficiency in area A was 12.3%, the soil carbon pool activity index was 0.28, and the ion balance index was 15.2; in area B, the values ​​were 11.8%, 0.26, and 14.8, respectively; and in area C, the values ​​were 11.5%, 0.25, and 14.5, respectively. Vegetation physiological and structural characteristics: The red edge position shift was calculated using hyperspectral differential spectroscopy, with a mean value of -6.2 nm before restoration; the mean value of photochemical reflectance index (PRI) was 0.08; the leaf area index was derived from the inversion of NDVI and red edge vegetation index using UAV hyperspectral data, with a mean value of 1.2; the community height was obtained by subtracting the digital elevation model from the stereo matching of UAV visible light imagery, with a mean value of 28 cm. Integration and Coupling Features: Construct a three-dimensional feature matrix of soil, vegetation, and atmosphere, including soil-driven features, vegetation response features, and environmental regulation features; soil-driven features include humus conversion efficiency, soil carbon pool activity index, and ion balance index; vegetation response features include REP offset, PRI value, LAI, and community height; environmental regulation features include slope, soil moisture, and grazing intensity. Time dimension characteristics: The quarterly change rate of NDVI was calculated, and the soil moisture data was combined to identify June-July as the water limitation period; the Mann-Kendall test was used to analyze the 5-year NDVI time series data, and the statistic was 1.72, which did not show a significant improvement trend. The annual increase of humus was synchronously correlated, with a mean of 0.8 g / kg. Spatial dimensional characteristics: The spatial variation coefficient of humic content and LAI was calculated using a 5m×5m grid. Based on the coupled feature matrix, K-means clustering was used to divide functional zones. Zone A had 28% high-efficiency remediation area, 52% medium-efficiency remediation area, and 20% low-efficiency remediation area; Zone B had 25%, 50%, and 25% respectively; and Zone C had 22%, 48%, and 30% respectively. Spectral and textural features: Reflectance in the red-edge band, normalized differential red-edge index, and chlorophyll absorption features were extracted as spectral features, with a mean normalized differential red-edge index of 0.22. Contrast and entropy were calculated based on the gray-level co-occurrence matrix of the RGB image, with a mean contrast of 45.2 and a mean entropy of 1.8, to obtain texture features.

[0025] In this embodiment, the method for obtaining time node data by performing three-level spatiotemporal collaborative classification of the grassland ecological characteristics includes: For macro satellite data: the Fmask algorithm is used to mark cloud pixels, the cloud-covered area is filled with linear interpolation, the maximum value of the effective images each month is calculated to synthesize NDVI, and a monthly macro vegetation cover map with a resolution of 30m is generated. The average NDVI of the 12 months before restoration is used as the benchmark to calculate the relative change rate of each month. For meso-level UAV data: using satellite imagery as a benchmark, SIFT feature matching is employed to control the UAV image registration error within one pixel; pseudo-invariant feature regions in the UAV imagery are selected, and reflectance is corrected to the satellite spectral response function through linear regression, ensuring that the reflectance deviation in the red-edge band is ≤3%; the UAV hyperspectral data is downsampled to 10m resolution and overlaid with the Sentinel-2 quarterly NDVI map to generate a composite layer of spectrum and vegetation index; For micro-level IoT data: the 3σ criterion is used to filter out sensor outliers, retaining time periods with a valid data ratio >90%; the moving window mean is calculated for each quarter's soil moisture data to generate a quarterly-scale moisture dataset; ground quadrat spectral data and UAV hyperspectral pixels are correlated through spatial buffer analysis; each quadrat contains at least 16 UAV pixels; Multiple hyperspectral indices were calculated, and texture parameters were extracted from UAV true-color images using gray-level co-occurrence matrix. Based on soil moisture data, water stress index was calculated and coupled with NDVI change rate to generate vegetation physiological state index. Hyperspectral index, texture parameters, and vegetation physiological state index were used as features to be fused. Water stress index = 1 - actual water content / field capacity. A convolutional long short-term memory network is used to perform three-dimensional feature fusion of the features to be fused, and a three-level classification rule is constructed for dynamic classification decision-making. The fused features are then subjected to temporal clustering for the four quarters of each year, and the cluster centers are extracted as key time node data. The first-level classification is based on the monthly change rate of satellite NDVI, the second-level classification is based on the recognition of UAV hyperspectral features, and the third-level classification is based on IoT data. A Bayesian network is used to integrate the three-level classification results, conditional probabilities are defined, key time node data are matched with the three-level classification rules, a quarterly dynamic classification map is generated, and the data corresponding to the time nodes in the quarterly dynamic classification map is output as time node data. In practical evaluation, the Convolutional Long Short-Term Memory (CLSTM) network includes an output layer, an input layer, and a time step. The input layer contains a 3-channel tensor and a time step. The 3-channel tensor contains 12 spectral features, 6 texture features, and 2 physiological features. The time step is 4, corresponding to 4 quarters. The ConvLSTM layer contains 2 hidden layers and a 3D convolutional kernel to extract spatiotemporal and feature coupling information. The hidden layers have 64 and 32 neurons respectively. The 3D convolutional kernel is 3×3×3. The output layer outputs 3 types of feature vectors through a softmax activation function, with a classification probability threshold of 0.5. Training optimization uses Monte Carlo cross-validation with a learning rate of 0.001 and 100 iterations to achieve a validation set accuracy > 85%. The 3 types of feature vectors are degenerate, in repair, and stable. If the ΔNDVI of a primary classification is less than -10% for two consecutive months, it indicates degradation; if -10% ≤ ΔNDVI ≤ 20%, it indicates recovery; if the ΔNDVI is greater than 20% for two consecutive quarters, it indicates stability. Secondary classification uses a support vector machine, which takes the red-edge band reflectance and texture features as input to classify the dominant species. In the three-level classification, when the water stress index is <0.3 and NDVI >0.6, it is a soil-vegetation coordinated type; when 0.3 ≤ water stress index <0.5 or 0.4 ≤ NDVI ≤0.6, it is a slightly restrictive type; when the water stress index is ≥0.5 or NDVI <0.4, it is a severely restrictive type. Macro satellite data: Cloud pixels were labeled using the Fmask algorithm, and cloud-covered areas were filled using linear interpolation. The maximum NDVI was calculated for each month's valid images to generate a 30m resolution monthly macro vegetation cover map. Using the average NDVI of the 12 months prior to restoration as a benchmark, the relative change rate for each month was calculated. In the 6th month after restoration, the relative change rate for area A was 15.2%, for area B it was 13.8%, and for area C it was 12.5%. Mid-level UAV data: Based on satellite imagery, SIFT feature matching was used to control the registration error of UAV imagery within 1 pixel; pseudo-invariant feature regions were selected, and reflectance was corrected to the satellite spectral response function through linear regression, with a reflectance deviation of ≤2.8% in the red-edge band; UAV hyperspectral data was downsampled to 10m resolution and overlaid with Sentinel-2 quarterly NDVI map to generate a composite layer of spectrum and vegetation index; Micro-IoT data: Sensor outliers are filtered using the 3σ criterion, retaining time periods with a valid data ratio > 92%; the sliding window mean of soil moisture data for each quarter is calculated to generate a quarterly moisture dataset; ground quadrat spectral data and UAV hyperspectral pixels are correlated through spatial buffer analysis, with each quadrat containing 18 UAV pixels.

[0026] In this embodiment, the method for obtaining fused features by coupling spectral texture and physiological features of the time-node data through three-dimensional convolution includes: A quantitative correlation between physiological characteristics and spectral / texture features was established by using a sequential chain mechanism of soil humus content, root water conductivity, and vegetation canopy structure; structural equation modeling was introduced to quantify the path coefficients of the chain. To address the temporal resolution differences between UAV spectral and soil sensor data, the DTW algorithm is employed to align high-frequency time-series data of physiological features with low-frequency data of spectral / texture features. The UAV hyperspectral data was used to eliminate atmospheric scattering by the dark target method, and atmospheric correction satellite data was used to select key bands from 125 bands based on the continuous projection algorithm to standardize physiological characteristics. The UAV hyperspectral imagery and satellite imagery were registered to the UTM coordinate system using a random sampling consistency algorithm. Soil sampling points were used to generate physiological feature raster maps through inverse distance weighted interpolation. Using the UAV flight date as the reference time node, the soil sensor data were aligned to this node using the DTW algorithm to form a quarterly physiological feature sequence. The spectral features, texture features, and physiological features are input into a convolutional long short-term memory network to extract coupled features. The weight vector of the coupled features is calculated using the sigmoid activation function to enhance the contribution of the red edge band and soil moisture. A 2D spatial mask is generated to assign 1.2 times the weight to the features of densely vegetated areas. Principal component analysis was used to reduce the dimensionality of the 128-dimensional coupled features to 32-dimensional features, and then the features were fused with a three-dimensional index system through feature concatenation to form fused features. In actual assessments, the UAV spectrum is at the quarterly level, the soil sensor at the hourly level, and the physiological characteristic raster map at a resolution of 2.5 cm. When NDVI > 0.5, it is considered a densely vegetated area. The three-dimensional index system includes humus, carbon sinks, and biodiversity. Quantitative correlation establishment: Through the sequential chain mechanism of soil humus content, root water conductivity, and vegetation canopy structure, structural equation modeling was introduced to quantify the path coefficients of the chain. The path coefficient of humus content to root water conductivity was 0.68, and the path coefficient of root water conductivity to vegetation canopy structure was 0.72. Data Alignment and Processing: To address the temporal resolution differences between UAV spectral and soil sensor data, the DTW algorithm was used to align high-frequency time-series data of physiological features with low-frequency data of spectral / texture features. Atmospheric scattering of UAV hyperspectral data was eliminated using the dark target method, and atmospheric correction was applied to satellite data. Based on a continuous projection algorithm, 28 key bands were selected from 125 bands to standardize physiological features. UAV hyperspectral and satellite imagery were registered to the UTM coordinate system using a random sampling consistency algorithm. Soil sampling points were used to generate a 2.5cm resolution physiological feature raster map through inverse distance weighted interpolation. Feature fusion generation: Spectral features, texture features, and physiological feature sequences are input into a convolutional long short-term memory network to extract coupled features. The weight vector of the coupled features is calculated using the Sigmoid activation function to enhance the contribution of the red edge band and soil moisture. A 2D spatial mask is generated, and features of dense vegetation areas with NDVI > 0.5 are assigned a weight of 1.2 times. Principal component analysis is used to reduce the dimensionality of the 128-dimensional coupled features to 32-dimensional features. The features are then fused with a three-dimensional index system through feature concatenation to form fused features.

[0027] In this embodiment, a method for constructing a three-dimensional index system based on humus, carbon sink, and biodiversity to obtain the ecological grassland restoration index includes: The three-dimensional indicators are process, state, and response dynamic indicators. Process indicators include soil organic carbon mineralization rate, nitrogen conversion efficiency, and carbon input from root exudates. State indicators include interannual NDVI coefficient of variation, soil organic carbon density, and dominant species coverage ratio. Response dynamic indicators include vegetation recovery rate under drought stress, functional diversity index, and microbial biomass carbon / organic carbon ratio. The process indicators, state indicators, and response dynamic indicators are standardized. A judgment matrix is ​​constructed based on the importance scores of the indicators by ecological restoration experts, and the subjective weights of the process, state, and response indicators are calculated. The objective weights are calculated based on the information entropy of the indicators, and the subjective and objective weights are weighted and summed to obtain the combined weights. The ecological grassland restoration index is obtained by weighting and summing the three-dimensional indicators using combined weights. In actual assessments, the rate of soil organic carbon mineralization was quantified using an indoor culture method; nitrogen conversion efficiency = (total nitrogen content of plants × biomass) / (nitrogen application rate of organic fertilizer) × 100%; and the carbon input from root exudates was quantified using the root box method. NDVI interannual variation coefficient = NDVI standard deviation / NDVI mean) × 100%, the smaller the value, the higher the stability; soil organic carbon density = soil organic carbon content × bulk density × soil layer thickness; dominant species coverage ratio = native dominant species coverage / degraded indicator species coverage, a value > 1 indicates positive community succession. The vegetation recovery rate under drought stress was determined by identifying the drought period based on the meteorological drought index and calculating the NDVI recovery period. The functional diversity index was calculated using the Rao index, which was based on five functional traits. The microbial biomass carbon / organic carbon ratio was determined by measuring microbial biomass carbon using the chloroform fumigation method and calculating the ratio; the higher the value, the stronger the activity. The mean soil organic carbon mineralization rate was 0.85 mg / (g·d), the mean nitrogen conversion efficiency was 32.5%, and the mean carbon input from root exudates was 12.8 g / (m²). 2 •a)); The mean value of the interannual coefficient of variation of NDVI after correction was 12.3%, and the mean value of soil organic carbon density was 3.2 kg / m³. 2 The dominant species coverage ratios after restoration were 1.3 in area A, 1.2 in area B, and 1.1 in area C; the mean vegetation recovery rate under drought stress was 0.08 / month, the mean functional diversity index was 0.42, and the mean microbial biomass carbon / organic carbon ratio was 0.052. Weighting calculation: The indicators are standardized, and a judgment matrix is ​​constructed based on the scores of ecological restoration experts. Subjective weights are calculated, with subjective weights of 0.35, 0.35, and 0.30 for process, state, and response indicators, respectively. Objective weights are calculated based on the information entropy of the indicators, with weights of 0.32, 0.36, and 0.32, respectively. The combined weights are obtained by weighted summation, with weights of 0.335, 0.355, and 0.310, respectively. Restoration index calculation: The three-dimensional indicators were weighted and summed using combined weights to obtain the ecological grassland restoration index. In the second year after restoration, the index was 0.72 for area A, 0.68 for area B, 0.63 for area C, and 0.51 for the control area.

[0028] In this embodiment, the method for constructing a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on the fusion features and the ecological grassland restoration index includes: An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on extreme gradient boosting and random forest is constructed. Logistic regression is used as the meta-model, fusing the predicted ecological grassland restoration index output by extreme gradient boosting with the random forest correction coefficient. The expression is as follows:

[0029] in This is a preliminary prediction of the ecological grassland restoration index. These are the random forest correction coefficients. The integrated ecological grassland restoration index; When the integrated ecological grassland restoration index is between 0.8 and 1, the restoration level is excellent; when the integrated ecological grassland restoration index is between 0.6 and 0.8, the restoration level is good; when the integrated ecological grassland restoration index is between 0.4 and 0.6, the restoration level is medium; when the integrated ecological grassland restoration index is below 0.4, the restoration level is poor. The weighted sum of the fused ecological grassland restoration index and the loss function is used as the objective function. SMOTE oversampling is used to divide the dataset into training and test sets. The hyperparameters are adjusted by Bayesian optimization and iterated until the error between the ecological grassland restoration index of the test set and the predicted value of the ecological grassland restoration index is minimized. In the actual assessment, the fusion features were standardized, and the soil moisture data was downscaled to the quarterly level using spline interpolation to synchronize with the UAV hyperspectral data in time. Spatial alignment between UAV and satellite imagery was achieved through coordinate registration. Satellite and UAV spectral data were unified through band resampling, and extreme values ​​in soil pH and NDVI were removed using interquartile range. Extreme Gradient Boosting: Input 128-dimensional coupled features, output preliminary ecological grassland restoration index prediction; Random Forest: Input interference factor, output interference correction coefficient; Leave-one-out cross-validation was used to verify the accuracy of the ecological organic fertilizer grassland restoration effectiveness assessment model. Accuracy verification included the coefficient of determination and root mean square error. For accuracy verification, the coefficient of determination was required to be greater than or equal to 0.75, and the root mean square error of biomass inversion was required to be less than or equal to 15%. The grazing intensity index of high-disturbance areas is corrected; areas with a grazing intensity index greater than 0.5 are considered high-disturbance areas. Area A: It is 0.75. It is 0.04 after fusion. It is 0.72; Area B: It is 0.71. It is 0.042 after fusion. It is 0.68; Area C: It is 0.66. It is 0.045 after fusion. It is 0.63; The fused grassland restoration index and the weighted sum of the loss function were used as the objective function. SMOTE oversampling was employed to partition the dataset, with the training set comprising 70% and the test set 30%. Hyperparameters were adjusted using Bayesian optimization, iterating until the error on the test set was minimized. Leave-one-out cross-validation was used to verify accuracy, yielding a coefficient of determination of 0.78 and a root mean square error of 13.2% for biomass inversion, meeting the accuracy requirements.

[0030] In this embodiment, the method for integrating mobile phone signaling data with livestock point data to optimize the ecological organic fertilizer grassland restoration effectiveness evaluation model includes: By acquiring mobile phone signaling data and livestock point data, fusing livestock point density estimates, and calculating spatial distribution intensity, the following methods are employed:

[0031] in Spatial distribution intensity, Let i be the weight of the i-th livestock point. For kernel function bandwidth, Let i be the coordinates of the i-th point of interest. The number of livestock farms, Spatial distribution intensity; Based on the spatiotemporal distribution of active users in mobile signaling, a monthly interference intensity curve is constructed, expressed as:

[0032] in The time t disturbance index, This refers to the number of active users in the current month. This is the historical average for the same period. A spatiotemporal weighted average is used to fuse spatial distribution intensity and temporal interference index to obtain a comprehensive interference index, in which the entropy weight method is used to determine the weights; Pearson correlation coefficient was used to analyze the relationship between the comprehensive disturbance index and vegetation indicators, and restoration areas were divided according to the comprehensive disturbance index value. A comprehensive disturbance index is added to the input features of the ecological organic fertilizer grassland restoration effectiveness evaluation model to form a fusion feature set of ecological features and disturbance features; An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration was trained using a fusion feature set. The evaluation results before and after correction were compared using ground-measured data. The comprehensive interference index threshold was recalculated quarterly based on newly collected mobile phone signaling and point of interest data. In actual assessments, the proportion of high-quality forage grass was significantly negatively correlated with CI, meaning that the stronger the grazing disturbance, the lower the proportion of high-quality forage grass. Based on the comprehensive disturbance index value, the remediation area is divided into three categories: when the comprehensive disturbance index is less than or equal to 0.3, it is a low disturbance area, which does not require correction and the original assessment results are used directly; when the comprehensive disturbance index is less than or equal to 0.6 and greater than 0.3, it is a medium disturbance area, which requires linear correction of indicators such as biomass and NDVI; when the comprehensive disturbance index is greater than 0.6, it is a high disturbance area, which uses a machine learning model for nonlinear correction. Ecological characteristics include soil organic carbon mineralization rate, NDVI interannual variation coefficient, and functional diversity index; disturbance characteristics include comprehensive disturbance index value, livestock point density, and number of active mobile phone signaling users. Comprehensive Interference Index Calculation: Mobile signaling data and livestock points were acquired, and livestock point density estimation was fused to calculate the spatial distribution intensity, with a mean of 0.038. Based on the spatiotemporal distribution of active users in mobile signaling, a monthly interference intensity curve was constructed, with a mean of 0.194. A spatiotemporal weighted average was used to fuse spatial distribution intensity and temporal interference index to obtain the comprehensive interference index. The entropy weight method was used to determine the weights: spatial distribution intensity weight was 0.45, and temporal interference index weight was 0.55. The comprehensive interference index for area A was 0.32, for area B 0.35, and for area C 0.38. Model optimization and threshold update: A comprehensive interference index is added to the model input features to form a fusion feature set of ecological features and interference features. The fusion feature set is used to train the model. The evaluation results before and after correction are compared using ground-measured data. The comprehensive interference index threshold is recalculated quarterly based on newly collected mobile phone signaling and livestock point data. The current threshold is ≤0.3 for low interference area, 0.3-0.6 for medium interference area, and >0.6 for high interference area. The restoration levels were determined based on the integrated ecological grassland restoration index: 0.8-1 was excellent, 0.6-0.8 was good, 0.4-0.6 was moderate, and below 0.4 was poor. The final evaluation results were: Area A was rated as good, Area B as good, Area C as moderate, and the control area as moderate. These results are consistent with the actual grassland restoration status, validating the effectiveness of the method.

[0033] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the effectiveness of ecological organic fertilizer in grassland restoration based on multi-source data, characterized in that, Includes the following steps: Grassland monitoring data and environmental data are collected from the pre-defined restoration area, and the grassland monitoring data and environmental data are preprocessed. The grassland monitoring data includes soil physicochemical data and vegetation ecological data. The vegetation ecological data includes UAV hyperspectral data, satellite remote sensing data, biomass, spectral reflectance, and indicator species surveys. The environmental data includes meteorological data, topographic data, and human disturbance data. The soil physicochemical data includes humus content, pH value and ion concentration, and organic matter conversion rate. By combining the soil-vegetation-atmosphere coupling mechanism, ecological vegetation features are extracted from the grassland monitoring data and the environmental data to obtain grassland ecological features, and time node data are obtained by performing three-level spatiotemporal collaborative classification of the grassland ecological features. By coupling spectral texture and physiological features of the time node data through three-dimensional convolution, a fusion feature is obtained, and a three-dimensional index system based on humus, carbon sink, and biodiversity is constructed to obtain the ecological grassland restoration index. Based on the fusion features and the ecological grassland restoration index, a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration is constructed. The model is then optimized by integrating mobile phone signaling data and livestock point data. The data to be evaluated is input into the optimized model, and the evaluation results are output.

2. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, The method for obtaining the ecological characteristics of the grassland includes: Based on the soil, vegetation, and atmosphere continuum theory, core driving parameters were selected, including humus content, active aluminum content, soil moisture conductivity, red-edge band reflectivity, interannual NDVI coefficient of variation, and grazing intensity. Soil physicochemical characteristics, such as humus conversion efficiency, soil carbon pool activity index, and ion balance index, were obtained by feature extraction from soil physicochemical data. Physiological and structural characteristics of vegetation were extracted. Physiological characteristics included red edge position shift and photochemical reflectance index; structural characteristics included leaf area index and community height. A three-dimensional feature matrix of soil, vegetation, and atmosphere was constructed to obtain fused coupling features, which include soil-driven features, vegetation response features, and environmental regulation features. Soil-driven features include humus conversion efficiency, soil carbon pool activity index, and ion balance index. Vegetation response features include REP offset, PRI value, LAI, and community height. Environmental regulation features include slope, soil moisture, and grazing intensity. The quarterly change rate of NDVI was calculated, and the water-limiting period was identified by combining soil moisture data. The Mann-Kendall test was used to analyze the 5-year NDVI time series data and synchronously correlated the annual increase of humus to obtain the time dimension characteristics. The spatial variation coefficient of humic content and LAI was calculated using grids as units. Based on the coupled feature matrix, K-means clustering was used to divide functional zones and obtain spatial dimensional features. The functional zones include high-efficiency remediation zone, medium-efficiency remediation zone and low-efficiency remediation zone. The reflectance of the red-edge band, the normalized differential red-edge index, and the chlorophyll absorption characteristics were extracted as spectral features; the contrast and entropy values ​​were calculated based on the gray-level co-occurrence matrix of the RGB image to obtain texture features. The soil physicochemical characteristics, physiological characteristics, structural characteristics, fusion and coupling characteristics, temporal dimension characteristics, spatial dimension characteristics, spectral characteristics, and texture characteristics are output as grassland ecological characteristics.

3. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, A method for obtaining time-node data by performing three-level spatiotemporal collaborative classification of the grassland ecological characteristics includes: For macro satellite data: the Fmask algorithm is used to mark cloud pixels, the cloud-covered area is filled with linear interpolation, the maximum value of the effective images each month is calculated to synthesize NDVI, and a monthly macro vegetation cover map with a resolution of 30m is generated. The average NDVI of the 12 months before restoration is used as the benchmark to calculate the relative change rate of each month. For meso-level UAV data: using satellite imagery as a benchmark, SIFT feature matching is employed to control the UAV image registration error within one pixel; pseudo-invariant feature regions in the UAV imagery are selected, and reflectance is corrected to the satellite spectral response function through linear regression, ensuring that the reflectance deviation in the red-edge band is ≤3%; the UAV hyperspectral data is downsampled to 10m resolution and overlaid with the Sentinel-2 quarterly NDVI map to generate a composite layer of spectrum and vegetation index; For micro-level IoT data: the 3σ criterion is used to filter out sensor outliers, retaining time periods with a valid data ratio >90%; the moving window mean is calculated for each quarter's soil moisture data to generate a quarterly-scale moisture dataset; ground quadrat spectral data and UAV hyperspectral pixels are correlated through spatial buffer analysis; each quadrat contains at least 16 UAV pixels; Multiple hyperspectral indices were calculated, and texture parameters were extracted from UAV true-color images using gray-level co-occurrence matrix. Based on soil moisture data, water stress index was calculated and coupled with NDVI change rate to generate vegetation physiological state index. Hyperspectral indices, texture parameters and vegetation physiological state index were used as features to be fused. A convolutional long short-term memory network is used to perform three-dimensional feature fusion of the features to be fused, a three-level classification rule is constructed for dynamic classification decision-making, and the fused features of the four quarters of each year are clustered in time series to extract the cluster centers as key time node data. A Bayesian network is used to integrate the three-level classification results, conditional probabilities are defined, key time node data are matched with the three-level classification rules to generate a quarterly dynamic classification map, and the data corresponding to the time nodes in the quarterly dynamic classification map are output as time node data.

4. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, A method for obtaining fused features by coupling spectral texture and physiological features of the time-node data through three-dimensional convolution includes: A quantitative correlation between physiological characteristics and spectral / texture features was established by using a sequential chain mechanism of soil humus content, root water conductivity, and vegetation canopy structure; structural equation modeling was introduced to quantify the path coefficients of the chain. To address the temporal resolution differences between UAV spectral and soil sensor data, the DTW algorithm is employed to align high-frequency time-series data of physiological features with low-frequency data of spectral / texture features. The UAV hyperspectral data was used to eliminate atmospheric scattering by the dark target method, and atmospheric correction satellite data was used to select key bands from 125 bands based on the continuous projection algorithm to standardize physiological characteristics. The UAV hyperspectral imagery and satellite imagery were registered to the UTM coordinate system using a random sampling consistency algorithm. Soil sampling points were used to generate physiological feature raster maps through inverse distance weighted interpolation. Using the UAV flight date as the reference time node, the soil sensor data were aligned to this node using the DTW algorithm to form a quarterly physiological feature sequence. The spectral features, texture features, and physiological features are input into a convolutional long short-term memory network to extract coupled features. The weight vector of the coupled features is calculated using the sigmoid activation function to enhance the contribution of the red edge band and soil moisture. A 2D spatial mask is generated to assign 1.2 times the weight to the features of densely vegetated areas. Principal component analysis was used to reduce the dimensionality of the 128-dimensional coupled features to 32-dimensional features, and then the features were fused with the three-dimensional index system to form fused features.

5. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, Methods for constructing a three-dimensional index system based on humus, carbon sink, and biodiversity to obtain the ecological grassland restoration index include: The three-dimensional indicators are process, state, and response dynamic indicators. Process indicators include soil organic carbon mineralization rate, nitrogen conversion efficiency, and carbon input from root exudates. State indicators include interannual NDVI coefficient of variation, soil organic carbon density, and dominant species coverage ratio. Response dynamic indicators include vegetation recovery rate under drought stress, functional diversity index, and microbial biomass carbon / organic carbon ratio. The process indicators, state indicators, and response dynamic indicators are standardized. A judgment matrix is ​​constructed based on the importance scores of the indicators by ecological restoration experts, and the subjective weights of the process, state, and response indicators are calculated. The objective weights are calculated based on the information entropy of the indicators, and the subjective and objective weights are weighted and summed to obtain the combined weights. The ecological grassland restoration index is obtained by weighting and summing the three-dimensional indicators using combined weights.

6. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, A method for constructing a machine learning-based evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on the fusion features and the ecological grassland restoration index includes: An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration based on extreme gradient boosting and random forest is constructed. Logistic regression is used as the meta-model, fusing the predicted ecological grassland restoration index output by extreme gradient boosting with the random forest correction coefficient. The expression is as follows: ； in This is a preliminary prediction of the ecological grassland restoration index. These are the random forest correction coefficients. The integrated ecological grassland restoration index; When the integrated ecological grassland restoration index is between 0.8 and 1, the restoration level is excellent; when the integrated ecological grassland restoration index is between 0.6 and 0.8, the restoration level is good; when the integrated ecological grassland restoration index is between 0.4 and 0.6, the restoration level is medium; when the integrated ecological grassland restoration index is below 0.4, the restoration level is poor. The fused grassland restoration index and the weighted sum of the loss function are used as the objective function. SMOTE oversampling is used to divide the dataset into training and test sets. The hyperparameters are adjusted by Bayesian optimization and iterated until the error between the grassland restoration index in the test set and the predicted value of the grassland restoration index is minimized.

7. The method for evaluating the effectiveness of ecological organic fertilizer grassland restoration based on multi-source data according to claim 1, characterized in that, A method for integrating mobile phone signaling data with livestock point data to optimize the evaluation model for the effectiveness of ecological organic fertilizer grassland restoration includes: By acquiring mobile phone signaling data and livestock point data, fusing livestock point density estimates, and calculating spatial distribution intensity, the following methods are employed: ； in Spatial distribution intensity, Let i be the weight of the i-th livestock point. For kernel function bandwidth, Let i be the coordinates of the i-th point of interest. The number of livestock farms, Spatial distribution intensity; Based on the spatiotemporal distribution of active users in mobile signaling, a monthly interference intensity curve is constructed, expressed as: ； in The time t disturbance index, This refers to the number of active users in the current month. This is the historical average for the same period. A spatiotemporal weighted average is used to fuse spatial distribution intensity and temporal interference index to obtain a comprehensive interference index, in which the entropy weight method is used to determine the weights; Pearson correlation coefficient was used to analyze the relationship between the comprehensive disturbance index and vegetation indicators, and restoration areas were divided according to the comprehensive disturbance index value. A comprehensive disturbance index is added to the input features of the ecological organic fertilizer grassland restoration effectiveness evaluation model to form a fusion feature set of ecological features and disturbance features; An evaluation model for the effectiveness of ecological organic fertilizer grassland restoration was trained using a fusion feature set. The evaluation results before and after correction were compared using ground-measured data. The comprehensive interference index threshold was recalculated quarterly based on newly collected mobile phone signaling and point of interest data.

Images