A method for dynamic monitoring and improvement of garden soil quality

Abstract

The application provides a kind of garden soil quality dynamic monitoring and improvement method, it is related to monitoring processing technical field, by receiving soil quality time series monitoring data, carry out multiscale sliding window segmentation, extract the comprehensive magnitude of change gradient and fluctuation amplitude, feature coding is carried out to historical evolution mode, obtain time series evolution characteristic vector;Calculate the pattern similarity between each monitoring point, causal transmission intensity and direction, construct double-layer conduction relationship diagram, execute stratified graph propagation diffusion and correct through physical constraint correction coefficient, obtain space-time conduction hidden state vector;Two vectors are combined, and the original and intervention deterioration risk curve is generated through survival risk analysis, the effect index is extracted by comparison, and the intervention effect evaluation table is generated;Read the evaluation table, combine the cost library and constraint rules, sort stratification according to the weighted comprehensive score of reduction amplitude and delay time, and output the differentiated improvement strategy set.The application realizes the dynamic monitoring and differentiated improvement decision of garden soil quality.

Description

Technical Field

[0001] This invention relates to the field of monitoring and processing technology, and more specifically, to a method for dynamic monitoring and improvement of garden soil quality. Background Technology

[0002] In the field of landscaping, soil quality has a crucial impact on plant growth, landscape maintenance, and ecological restoration. As human demands for green space quality increase, soil monitoring and remediation technologies are gradually shifting from traditional, experience-based methods to modern, automated technologies supported by data. Currently, some technologies have achieved preliminary dynamic soil monitoring, typically using conventional physical and chemical indicators (such as moisture content, nutrient content, and pH) as the monitoring basis. These technologies record data at multiple time points through distributed sensor networks and combine spatial modeling and statistical analysis to conduct regional soil quality assessments. While these technologies can achieve basic dynamic monitoring of regional soil quality, limitations remain. For example, most existing methods employ independent and unrelated monitoring point designs, analyzing data only for a single time period, lacking the ability to interpret global patterns of soil temporal changes. Furthermore, for the long-term dynamic evolution of the soil environment, existing monitoring methods fail to fully reveal its potential propagation mechanisms, spatiotemporal structural characteristics, and multi-scale correlation paths, making comprehensive, accurate, and causally driven monitoring and remediation difficult to implement.

[0003] However, soil quality, as a dynamic and complex system, is influenced by multiple environmental factors and long-term human activities, such as precipitation, irrigation, fertilization, and landscaping patterns. These factors continuously alter the physicochemical structure and ecological functions of the soil through complex spatiotemporal interactions. For example, trend analysis based on a single monitoring indicator lacks insight into the relationships between soil components, especially when significant spatiotemporal variability exists among multiple monitoring points, often making it impossible to accurately construct their intrinsic correlation patterns. Furthermore, the ability to detect soil degradation processes early is not highly sensitive, potentially missing the optimal time for improvement and intervention before problems become apparent. For instance, existing technologies tend to focus on direct operations, such as single-point fertilization and covering with soil amendment particles, whose effects often lack quantitative support, making it difficult to develop a differentiated strategy improvement system. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for dynamic monitoring and improvement of garden soil quality. This method can, to some extent, solve the problems that existing monitoring methods cannot comprehensively integrate the temporal evolution trajectory of a single monitoring point with the spatial correlation characteristics between adjacent monitoring points, resulting in the inability to identify soil quality deterioration trends in advance. It also addresses the issues that improvement measures may result in delayed responses or overuse of amendments due to fixed monitoring frequencies and neglect of historical evolution patterns.

[0005] According to one aspect of the present invention, a method for dynamic monitoring and improvement of garden soil quality is provided, comprising:

[0006] The soil quality time-series monitoring data from each monitoring point is received, a time-series data matrix is ​​constructed, multi-scale sliding window segmentation is performed, the comprehensive magnitude of the gradient and fluctuation amplitude of the index change within each window is extracted, the historical evolution pattern is feature-encoded, and the time-series evolution feature vector is obtained.

[0007] The pattern similarity, causal transmission strength, and directionality of the temporal evolution feature vectors among each monitoring point are calculated. A two-layer transmission relationship graph is constructed. The two-layer transmission relationship graph is subjected to layered graph propagation and diffusion and corrected by physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector.

[0008] The temporal evolution feature vector and the spatiotemporal transmission latent state vector are merged, and the original and post-intervention degradation risk curves are generated through survival risk analysis. The effect indicators are compared and extracted to generate an intervention effect evaluation table.

[0009] Read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraints, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a set of differentiated strategies.

[0010] Furthermore, the generation of the temporal evolution feature vector includes:

[0011] The comprehensive magnitude sequence of the same soil quality index is grouped by window scale, and the comprehensive magnitude sequences corresponding to short-period windows, medium-period windows and long-period windows are extracted. The statistical characteristics of the sequence at each scale are calculated respectively.

[0012] The statistical features include the sequence mean, sequence variance, sequence extreme values, and sequence trend slope;

[0013] Statistical features at different window scales are concatenated in the order of short period, medium period, and long period to form a multi-scale statistical feature vector.

[0014] The multi-scale statistical feature vectors of all soil quality indicators at the same monitoring point are jointly spliced ​​and z-score standardized.

[0015] Furthermore, the two-layer transmission relationship graph includes an undirected similarity relationship layer and a directed causal transmission layer;

[0016] The undirected similarity relation layer calculates the cosine similarity of the temporal evolution feature vector of the monitoring points and maps it to a non-negative value range through exponential transformation to obtain the pattern similarity intensity value. The monitoring point pairs with the pattern similarity intensity value exceeding the preset similarity threshold are used as undirected edges and the pattern similarity intensity value is used as the edge weight.

[0017] The directed causal transmission layer obtains the F-statistic and significance probability values ​​through the Granger causality test. The F-statistic is then transformed by logarithmic transformation and linear scaling to map to a non-negative value range to obtain the causal transmission strength value. Directional causal labels are established for monitoring point pairs that meet the significance condition, and the causal transmission strength value is used as the edge weight.

[0018] Furthermore, the formula for calculating the physical constraint correction coefficient is as follows:

[0019]

[0020] in, Source monitoring point Pointing to the target monitoring point The physical constraint correction coefficients corresponding to the directed edges. For monitoring points With monitoring points The Euclidean straight-line distance between them. Source monitoring point The set of target monitoring points connected by all outgoing edges in the directed transmission graph. The spatial decay scale parameter of the Gaussian kernel function. For monitoring points With monitoring points Soil type compatibility coefficient between them As a terrain orientation preference factor, For target monitoring points Vegetation buffer factor, This is a constant used to protect the denominator in the normalization process.

[0021] Furthermore, the survival risk analysis model includes a feature transformation module and a degradation risk intensity function;

[0022] The feature transformation module uses a multi-layer neural network structure to transform the comprehensive feature vector into a risk representation vector;

[0023] The formula for the degradation risk intensity function is:

[0024]

[0025] in, For the baseline risk function, For the regression coefficient vector, The risk representation vector output by the feature transformation module; The coefficient of variation of the comprehensive feature vector; The uncertainty adjustment index was selected through cross-validation. This is a time-dependent weight vector; is the numerical stability constant.

[0026] Furthermore, the survival risk analysis model is jointly trained using historical monitoring data, and the training loss function is constructed based on negative log-partial likelihood, expressed as follows:

[0027]

[0028] in, It is the partial likelihood function. The negative log-partial likelihood loss function is used. This represents the number of historical monitoring points. This is an event indicator variable; it takes a value of 1 when the event occurs and a value of 0 when the event is truncated. For monitoring points The observation time In time The risk set at any given time includes all observation times not less than [a certain value]. monitoring points Indicates the first risk in the risk set Each monitoring point at time The value of the degradation risk intensity function at any given time.

[0029] Furthermore, the generation of the post-intervention degradation risk curve includes:

[0030] The attention mechanism is used to synthesize the feature vector. Dynamically fuse the intervention feature vector with the intervention feature vector of the improvement measures to form an intervention fusion feature vector. ;

[0031] The intervention fusion feature vector is input into the feature transformation module to obtain the risk representation vector of the monitoring point under the intervention of the improved measures. ;

[0032] Substitute the risk representation vector and the intervention fusion feature vector into the degradation risk intensity function to calculate the degradation risk intensity value at each discrete time point in the future;

[0033] Cubic spline interpolation is used to generate the post-intervention degradation risk curve from the discrete risk intensity values.

[0034] According to another aspect of the present invention, a dynamic monitoring and improvement system for garden soil quality is provided, comprising:

[0035] The feature encoding module is used to receive soil quality time-series monitoring data from each monitoring point, construct a time-series data matrix, perform multi-scale sliding window segmentation, extract the comprehensive magnitude of the gradient and fluctuation amplitude of the index change within each window, perform feature encoding on the historical evolution pattern, and obtain the time-series evolution feature vector.

[0036] The relationship modeling module is used to calculate the pattern similarity, causal transmission strength and directionality of the temporal evolution feature vectors between monitoring points, construct a two-layer transmission relationship graph, perform layered graph propagation and diffusion on the two-layer transmission relationship graph and correct it through physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector.

[0037] The effect evaluation module is used to merge the temporal evolution feature vector with the spatiotemporal transmission latent state vector, generate the original and post-intervention degradation risk curves through survival risk analysis, compare and extract effect indicators, and generate an intervention effect evaluation table.

[0038] The strategy generation module is used to read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraint rules, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a differentiated strategy set.

[0039] Compared with existing technologies, this invention comprehensively extracts the gradient and fluctuation patterns of soil quality indicators by performing multi-scale sliding window segmentation and feature encoding on soil quality time-series monitoring data from multiple monitoring points. Utilizing the similarity and causal transmission relationships of temporal evolution characteristics among monitoring points, a two-layer transmission relationship diagram is constructed and stratified correction is performed, revealing the spatiotemporal evolution patterns and correlation mechanisms of soil quality. Combined with survival risk analysis and intervention effect evaluation, it achieves early identification and quantitative analysis of soil degradation risks. By comprehensively evaluating the intervention effect and the cost of improvement measures, differentiated stratified improvement strategies are generated, improving the accuracy of improvement measures and resource utilization efficiency. This effectively solves the problems of fixed monitoring, delayed response, and low improvement efficiency in existing technologies. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0041] Figure 1 This is a flowchart of a method for dynamic monitoring and improvement of garden soil quality according to an embodiment of the present invention;

[0042] Figure 2 This is a comparison chart showing the beneficial effects of the physical constraint correction coefficients according to embodiments of the present invention. Detailed Implementation

[0043] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0044] As mentioned in the background section above, existing technologies have two main problems: First, existing methods for monitoring garden soil rely on regular manual sampling and testing, with fixed monitoring frequencies and long intervals. This makes it impossible to capture sudden dynamic changes in soil quality caused by factors such as trampling by tourists, rain erosion, and seasonal maintenance operations, making it difficult to detect deterioration trends in a timely manner. Second, improvement decisions rely on threshold judgments based on single test results, ignoring the historical evolution trajectory of soil quality indicators and the spatial correlation between monitoring points, which leads to delayed improvement measures or excessive application of soil amendments.

[0045] Figure 1 This is a system block diagram of a method for dynamic monitoring and improvement of garden soil quality according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the methods for dynamic monitoring and improvement of garden soil quality include:

[0046] S1: Receive soil quality time-series monitoring data from each monitoring point, perform multi-scale sliding window segmentation, calculate the comprehensive magnitude of the gradient and fluctuation amplitude of the index changes within each window, perform feature encoding on the historical evolution pattern, and obtain the time-series evolution feature vector.

[0047] Soil quality time-series monitoring data were obtained from various monitoring points within the park area. This data includes soil quality index values ​​recorded at different timestamps from multiple monitoring points. Soil quality indexes cover parameters reflecting the physical and chemical state of the soil, such as soil pH, moisture content, organic matter content, nitrogen, phosphorus and potassium nutrient content, soil compaction, permeability, and soil hardness. Each monitoring point is assigned a unique monitoring point number, and each monitoring record carries the timestamp information of the collection time.

[0048] Meanwhile, each monitoring point is pre-labeled with the following attribute parameters: inherent attribute parameters such as soil degradation type label, degradation urgency level, vegetation density, visitor activity frequency, landscape aesthetic value rating, and ecosystem sensitivity rating, as well as management parameters such as landscape quality threshold, visitor complaint rate statistics, natural recovery cycle, soil microbial diversity index, carrying capacity, and soil hardness standard value. Among them, vegetation density is the number of vegetation cover plants or leaf area index per unit area around the monitoring point; visitor activity frequency is the average daily visitor flow statistics of the area where the monitoring point is located during the monitoring period; landscape aesthetic value rating and ecosystem sensitivity rating are assessed according to the grading standards preset by the garden management department; the statistical parameters in the management parameters are updated regularly by the garden maintenance management system.

[0049] Soil quality time-series monitoring data are grouped according to monitoring point number. The values ​​of various soil quality indicators collected at the same monitoring point at different timestamps are arranged in chronological order. An independent time-series data matrix is ​​constructed for each soil quality indicator. The time-series data matrix uses the monitoring point number as the row index, the timestamp as the column index, and the soil quality indicator value as the matrix element.

[0050] Perform multi-scale sliding window segmentation on the time series data of each monitoring point in each time series data matrix, including short-period windows, medium-period windows, and long-period windows:

[0051] Short-cycle windows correspond to time spans from hours to days, and are used to capture rapid fluctuations in soil compaction and moisture content caused by events such as tourist trampling and short-term heavy rainfall.

[0052] The medium-cycle window corresponds to a time span from week to month, and is used to identify periodic changes in soil nutrient content and organic matter content caused by seasonal maintenance operations and plant growth succession.

[0053] The long-term window corresponds to a time span from quarter to grade level, and is used to extract the annual evolution trends of soil organic matter content, permeability, etc. caused by changes in long-term fertilization plans and drainage conditions.

[0054] Slide windows at various scales on the time axis with a fixed step size, which is set to half the length of the corresponding window. Extract the soil quality index values ​​within the time period covered by each window to form multiple time window segments.

[0055] The gradient and fluctuation amplitude of the soil quality index values ​​within each time window segment are calculated respectively. The gradient is obtained by dividing the difference between the values ​​at the beginning and end of the calculation window by the time span of the window. The fluctuation amplitude is obtained by the standard deviation of all values ​​within the calculation window.

[0056] The calculated gradient values ​​and fluctuation amplitude values ​​within the same time window are weighted and summed. The weighting coefficients are determined based on the time span length corresponding to the window size.

[0057] When the window time span is less than one-tenth of the total monitoring time, the window is determined to be a short-cycle window. The weight value of the fluctuation amplitude is greater than the weight value of the change gradient, so that the comprehensive magnitude value can sensitively reflect rapid fluctuation events such as instantaneous compaction or sudden changes in moisture content.

[0058] When the window time span is longer than one-third of the total monitoring time, the window is determined to be a long-cycle window. The weight value of the change gradient is greater than the weight value of the fluctuation amplitude, so that the comprehensive magnitude value can stably reflect the long-term degradation trend such as the continuous depletion of organic matter or the gradual decline of permeability.

[0059] When the window time span is between one-tenth and one-third of the total monitoring time, the window is determined to be a medium-cycle window. The change gradient and fluctuation amplitude are assigned equal weight values ​​so that the comprehensive magnitude value can take into account both the cyclical fluctuations of nutrients caused by seasonal fertilization and the potential gradual trend across multiple cycles.

[0060] It should be noted that the specific weight values ​​at each scale are determined by the actual degradation event records of the monitored parks and green spaces. By taking the confirmed trampling and compaction events or continuous drought trend events in the historical monitoring data as a reference, the weight values ​​are adjusted so that the peak or trend of the calculated comprehensive magnitude sequence most accurately matches the occurrence time and intensity of the recorded events, thus obtaining the weight coefficient combination applicable to the park area.

[0061] The result of the weighted summation operation is used as the comprehensive magnitude value of the time window segment. The comprehensive magnitude values ​​are calculated for all time window segments generated by each monitoring point under different window scales and different sliding positions, forming a comprehensive magnitude sequence set indexed by the monitoring point number and containing comprehensive magnitude values ​​of multiple scales and time periods.

[0062] Feature encoding of historical evolution patterns based on the comprehensive magnitude sequence set of each monitoring point for each soil quality index:

[0063] The comprehensive magnitude sequence set of the same soil quality index is grouped according to window scale. The comprehensive magnitude sequence corresponding to the short period window, the comprehensive magnitude sequence corresponding to the medium period window, and the comprehensive magnitude sequence corresponding to the long period window are extracted respectively. The statistical characteristics of the comprehensive magnitude sequence at each scale are calculated. The statistical characteristics include the sequence mean, sequence variance, sequence extreme value, and sequence trend slope.

[0064] The statistical features extracted from different window scales are spliced ​​together in the order of short period, medium period, and long period to form a multi-scale statistical feature vector of the soil quality index. The dimension of the multi-scale statistical feature vector is determined by the number of window scales and the number of statistical features at a single scale. The first part of the vector corresponds to the rapid fluctuation features of the short period, the middle part of the vector corresponds to the seasonal variation features of the medium period, and the last part of the vector corresponds to the trend evolution features of the long period.

[0065] The multi-scale statistical feature vectors of all soil quality indicators at the same monitoring point are jointly spliced ​​together. The multi-scale statistical feature vectors corresponding to each indicator such as soil pH, moisture content, organic matter content, nitrogen, phosphorus and potassium nutrient content, soil compaction and permeability are connected end to end to form the joint statistical feature vector of the monitoring point. This joint statistical feature vector integrates the historical evolution pattern information of all soil quality indicators at the monitoring point at multiple time scales.

[0066] The joint statistical feature vector is subjected to z-score standardization to eliminate the influence of dimensional differences between different soil quality indicators on feature encoding, so that the evolution characteristics of each indicator can be characterized on a unified numerical scale.

[0067] The standardized joint statistical feature vector is used as the time-series evolution feature vector of the monitoring point. Time-series evolution feature vectors are generated for all monitoring points in the garden area, forming a set of time-series evolution feature vectors indexed by the monitoring point number.

[0068] S2: Calculate the pattern similarity, causal transmission strength and directionality of the temporal evolution feature vectors between each monitoring point, construct a two-layer transmission relationship diagram, perform layered graph propagation and diffusion on the two-layer transmission relationship diagram and correct it through physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector.

[0069] Read the generated set of time-series evolution feature vectors, extract the time-series evolution feature vectors of each monitoring point, perform pairwise calculations on the time-series evolution feature vectors of any two monitoring points, and establish a two-layer correlation relationship.

[0070] The first layer is an undirected similarity relation layer. The cosine similarity between the temporal evolution feature vectors of two monitoring points is calculated and then mapped to a non-negative numerical range via an exponential transformation, serving as the pattern similarity strength value. Using monitoring points as nodes, pairs of monitoring points with pattern similarity strength values ​​exceeding a preset similarity threshold as undirected edges, and the pattern similarity strength values ​​as edge weights, an undirected similarity relation layer is constructed. The preset similarity threshold is determined based on the statistical distribution of pattern similarity strength values ​​for all monitoring point pairs within the garden area, with the upper quartile of this distribution used as the threshold value.

[0071] Specifically, for any pair of monitoring points, the cosine similarity calculation method is used to calculate the pattern similarity between the temporal evolution feature vectors of the two monitoring points. By treating the two temporal evolution feature vectors as vectors in a high-dimensional space, the cosine value of the angle between the two vectors is calculated.

[0072] The calculated cosine similarity values ​​are mapped to a range from zero to positive infinity using an exponential transformation. The base of the exponential transformation is the base of the natural logarithm, and the exponent is the product of the cosine similarity value and a scaling factor. The scaling factor is selected as the upper quartile as a reference point based on the empirical distribution of historical cosine similarity values ​​of all monitoring point pairs within the garden area. The target mapping intensity value of this reference point after exponential transformation is set and obtained by solving the exponential equation in reverse. The value after exponential transformation is used as the pattern similarity intensity value between two monitoring points. This pattern similarity intensity value is always non-negative, and the larger the value, the higher the instantaneous similarity between the two monitoring points in the soil quality evolution pattern.

[0073] For all monitoring point pairs, an undirected similarity relationship layer is constructed based on the pattern similarity strength values ​​of each monitoring point pair. The undirected similarity relationship layer uses monitoring points as nodes and the associations between monitoring point pairs whose pattern similarity strength values ​​exceed a preset similarity threshold as undirected edges. The weight of the undirected edge is the pattern similarity strength value between the corresponding monitoring point pairs.

[0074] The second layer is the directed causal transit layer. Granger causality tests are performed on the raw time-series data of any two monitoring points to obtain the F-statistic and significance probability values. Directional causal labels are established between monitoring point pairs that meet the significance condition. The F-statistic is then logarithmically transformed and linearly scaled to map to a non-negative range, serving as the causal transit strength value. The directed causal transit layer is constructed with monitoring points as nodes, causal labels as directed edges, and causal transit strength values ​​as edge weights.

[0075] Specifically, for any pair of monitoring points, the temporal causal transmission strength and directionality between the two monitoring points are calculated. Granger causality tests are performed on the original time series data of the two monitoring points. The Granger causality test determines whether the historical information of one time series has a predictive effect on the future value of another time series by constructing an autoregressive model. If the historical data of soil quality indicators of monitoring point A has a significant predictive effect on the future data of soil quality indicators of monitoring point B, then it is determined that monitoring point A has a causal influence on monitoring point B, and the directional marker from monitoring point A to monitoring point B is recorded.

[0076] The Granger causality test outputs an F-statistic and a significance probability. The F-statistic reflects the degree to which historical data of monitoring point A contributes to the prediction of future data of monitoring point B. The F-statistic is mapped to the range of zero to positive infinity through logarithmic transformation and linear scaling. The mapped value is used as the causal transmission strength from monitoring point A to monitoring point B.

[0077] Granger causality tests were performed on all monitoring point pairs to examine the causal relationship between monitoring point A and monitoring point B, and the causal relationship between monitoring point B and monitoring point A, respectively. Based on the test results, the causal directionality markers and causal transmission strength values ​​between the two monitoring points were determined.

[0078] If the probability value of the causal relationship between monitoring point A and monitoring point B is lower than the preset significance threshold, a one-way causal marker from monitoring point A to monitoring point B is established, and the corresponding causal transmission strength value is recorded.

[0079] If the probability value of the causal relationship between monitoring point B and monitoring point A is lower than the preset significance threshold, a one-way causal marker from monitoring point B to monitoring point A is established, and the corresponding causal transmission strength value is recorded.

[0080] If the significance probability values ​​of the causal relationship between monitoring point A and monitoring point B are both lower than the preset significance threshold, then a bidirectional causal marker is established between monitoring point A and monitoring point B, and the causal transmission strength values ​​in both directions are recorded respectively.

[0081] If neither monitoring point A nor monitoring point B meets the criteria for significant causal relationship, then no causal marker will be established.

[0082] Based on the calculated causal transmission strength values ​​and causal directionality labels between all monitoring points, a directed causal transmission layer is constructed. The directed causal transmission layer uses monitoring points as nodes and the time-series prediction paths between monitoring point pairs with causal directionality labels as directed edges. The weight of the directed edge is the causal transmission strength value between the corresponding monitoring point pairs.

[0083] The undirected similarity relationship layer and the directed causal transmission layer are fused into a graph structure, and a two-layer transmission relationship graph containing the undirected similarity edge set and the directed causal transmission edge set is constructed with the monitoring points as the shared node set.

[0084] A layered graph propagation operation is performed on the two-layer transmission relationship graph. The temporal evolution feature vector generated by each monitoring point in S1 is used as the initial node feature vector, and the features of each node are iteratively updated. In each iteration, adjacent node features are aggregated by weighted undirected edge weights in the undirected similarity relationship layer, and source node features are aggregated by weighted directed edge weights in the directed causal transmission layer. The aggregation results from the two layers are weighted and combined using an inter-layer fusion coefficient, and then linearly combined with the current node feature vector to update the graph. This process is repeated until convergence or the maximum number of iterations is reached, yielding the feature vector after the two-layer graph propagation. The inter-layer fusion coefficient is the ratio of the in-degree value of a monitoring point in the directed causal transmission layer to the sum of its in-degree and out-degree values.

[0085] After the two-layer graph propagation and diffusion, the feature vector integrates the temporal evolution pattern information of the monitoring point itself, the pattern common information transmitted by adjacent monitoring points through similarity relationships, and the temporal prediction information transmitted by the source node through causal transmission relationships. However, the graph propagation and diffusion operation only relies on the mathematical similarity and statistical causal relationship of the temporal evolution feature vectors, without considering the physical spatial distribution characteristics of garden soil and the physical constraint law of soil quality indicators in the spatial propagation process.

[0086] Furthermore, the soil physical property parameters of each monitoring point are read, including geographical coordinates, soil type, terrain slope, vegetation cover density, and drainage condition level.

[0087] Based on physical property parameters, physical constraint correction coefficients are constructed for each edge in the two-layer transmission relationship graph. Specifically:

[0088] For from monitoring points Pointing to monitoring point Directed edges, physical constraint correction coefficients The formula is expressed as:

[0089]

[0090] in, Source monitoring point Pointing to the target monitoring point The physical constraint correction coefficients corresponding to the directed edges. For monitoring points With monitoring points The Euclidean straight-line distance between them. Source monitoring point The set of target monitoring points connected by all outgoing edges in the directed transmission graph. The spatial decay scale parameter of the Gaussian kernel function. For monitoring points With monitoring points Soil type compatibility coefficient between them As a terrain orientation preference factor, For target monitoring points Vegetation buffer factor, This is a constant used to protect the denominator in the normalization process.

[0091] Among them, the terrain directionality preference factor is:

[0092]

[0093] in, This is a weighting adjustment coefficient for terrain directionality. For monitoring points The slope of the terrain minus the monitoring points The slope difference of the terrain slope, For monitoring points The drainage condition level value minus the monitoring point The difference in the level of drainage condition rating. monitoring points The terrain slope value, For monitoring points The terrain slope value.

[0094] The formula for the vegetation buffer factor is:

[0095]

[0096] in, This represents the intensity coefficient of the vegetation buffering effect. For target monitoring points The surrounding vegetation cover density value, This represents the maximum vegetation cover density value at all monitoring points within the park area.

[0097] For undirected edges in the undirected similarity relation layer, the physical constraint correction coefficient... The formula is expressed as:

[0098]

[0099] in, Connecting monitoring points in the undirected similarity relation layer With monitoring points The physical constraint correction coefficients corresponding to the undirected edges; The symmetric vegetation buffer factor for monitoring points at both ends of an undirected edge. and monitoring points With monitoring points vegetation cover density, This is the maximum reference value for vegetation cover density. For undirected similarity relationship layers and monitoring points There exists a set of all adjacent monitoring points connected by an undirected edge.

[0100] The pattern similarity strength values ​​of each undirected edge in the undirected similarity relation layer are compared with the corresponding... Perform element-wise multiplication to obtain the similarity edge weights after physical constraint correction; then, multiply the causal propagation strength values ​​of each directed edge in the directed causal propagation layer with their corresponding values. Perform element-wise multiplication to obtain the causal transitive edge weights after physical constraint correction.

[0101] Replace the original edge weights with the edge weights corrected by physical constraints, and re-execute the layered graph propagation and diffusion operation on the double-layered transmission graph with updated edge weights. The iteration rules are the same as described above, and the physical constraint corrected feature vectors are obtained.

[0102] After physical constraint correction, the feature vectors, based on the fusion of temporal evolution pattern information of adjacent monitoring points, further consider the constraint effect of soil physical property parameters on the spatial transmission process. This makes the update process of node feature vectors not only rely on mathematical similarity and statistical causal relationship, but also follow the transmission capacity and directional preference of soil quality indicators in the physical spatial propagation process, thus avoiding information propagation path distortion caused by ignoring physical constraints.

[0103] like Figure 2As shown in the diagram of the existing method on the left, the thickness of the boundary lines between the five monitoring points (M1 to M5) reflects the numerical values ​​of pure statistical similarity and causal strength. It can be seen that the weights of the long-distance heterogeneous soil path from M1 to M3 and the short-distance homogeneous downslope path from M1 to M4 are almost the same (0.79 and 0.81). This uniform distribution violates the objective law that soil quality indicators are constrained by distance attenuation, media compatibility, and topographic directionality during propagation in physical space. In the diagram of this invention on the right, after… After correction, the M1→M3 path was compressed to 0.13 due to the dual physical barriers of long distance and different soil types and marked with a red descending arrow, while the M1→M4 path was enhanced to 0.93 due to the dual gain of downhill conduction direction and low vegetation buffer and marked with a green ascending arrow, which intuitively reflects the graded differences of strong and weak conduction paths after correction.

[0104] Furthermore, the topological characteristics of each monitoring point in the two-layer transmission relationship graph are read, including the undirected similarity relationship layer degree, the directed causal transmission layer in-degree, the directed causal transmission layer out-degree, the betweenness centrality of the two-layer fusion graph, and the compact centrality of the two-layer fusion graph. After normalizing the topological characteristic values, they are concatenated with the physical constraint-corrected feature vector to form an extended feature vector. The first element of the extended feature vector is the physical constraint-corrected feature vector, and the second element is the normalized topological characteristic value. The extended feature vector simultaneously encodes the temporal evolution pattern information, similarity transmission relationship information, causal transmission relationship information, physical constraint correction information, and network topology information of the monitoring point.

[0105] The extended feature vector is used as the spatiotemporal conduction latent state vector of the monitoring point. Spatiotemporal conduction latent state vectors are generated for all monitoring points in the garden area, forming a set of spatiotemporal conduction latent state vectors indexed by the monitoring point number.

[0106] S3: Merge the temporal evolution feature vector with the spatiotemporal transmission latent state vector, generate the original and post-intervention degradation risk curves through survival risk analysis, compare and extract effect indicators, and generate an intervention effect evaluation table.

[0107] Read the set of temporal evolution feature vectors generated by S1 and the set of spatiotemporal conduction latent state vectors generated by S2. For each monitoring point in the garden area, extract the temporal evolution feature vector and the spatiotemporal conduction latent state vector corresponding to that monitoring point.

[0108] For the temporal evolution feature vector and the spatiotemporal propagation latent state vector of the same monitoring point, feature merging is performed through vector concatenation. The concatenation operation arranges all elements of the two vectors sequentially according to their vector dimensions to form a comprehensive feature vector, denoted as . .

[0109] Comprehensive feature vector The first part consists of all elements of the temporal evolution feature vector, and the second part consists of all elements of the spatiotemporal transmission latent state vector. The total dimension is... It equals the sum of the dimension of the temporal evolution feature vector and the dimension of the spatiotemporal transmission latent state vector, and simultaneously encodes the temporal evolution pattern and spatial causal transmission state of the monitoring point itself.

[0110] The comprehensive feature vector of each monitoring point The survival risk analysis model, built upon the theoretical framework of survival analysis, is used to predict the temporal distribution and risk changes of a monitoring point from its current state to the occurrence of a soil quality deterioration event. A soil quality deterioration event is defined as the point in time when the comprehensive soil quality score of a monitoring point first significantly decreases, triggering a quality warning state. The quality warning state is calibrated using the statistical distribution characteristics of historical deterioration event records. The survival risk analysis model consists of two parts: a feature transformation module and a deterioration risk intensity function.

[0111] The feature transformation module adopts a multi-layer neural network structure, including an input layer, There are one hidden layer and one output layer. The input layer receives data in one dimension. Comprehensive feature vector The hidden layer consists of alternating stacks of fully connected layers and hyperbolic tangent activation functions. Hidden layers contain Each hidden layer has 12 neurons that receive the output from the previous layer and perform a linear transformation. This linear transformation is achieved using a weight matrix and a bias vector. The transformation result is then mapped to the negative-to-positive interval using a hyperbolic tangent activation function, introducing non-linear expressive power and preventing unbounded growth of the neuron's output. The output of the last hidden layer is the risk representation vector, denoted as [missing information]. Its dimension is lower than the input dimension. It encodes deep feature patterns related to degradation risk in the comprehensive feature vector, and the dimensions are adaptively selected on the training dataset through cross-validation.

[0112] The degradation risk intensity function is used to calculate the monitoring points. At any time in the future The instantaneous degradation risk intensity is expressed by the formula:

[0113]

[0114] in, For the baseline risk function, For the regression coefficient vector, The risk representation vector output by the feature transformation module; The coefficient of variation of the comprehensive feature vector; The uncertainty adjustment index was selected through cross-validation. This is a time-dependent weight vector; is the numerical stability constant.

[0115] It should be noted that the weight matrix and bias vector in the feature transformation module, as well as the regression coefficient vector in the degradation risk intensity function, are related to this. Uncertainty Adjustment Index Time-dependent weight vector The learning parameters are obtained through a joint training process using historical monitoring data. Specifically:

[0116] The training dataset contains Each historical monitoring point sample records the comprehensive feature vector of that monitoring point and the observed occurrence time or truncation time of the degradation event. and event indicator variables If monitoring point If a degradation event occurs during the observation period, the event indicator variable takes the value of 1; otherwise, if the data is truncated, the event indicator variable takes the value of 0.

[0117] The training process employs an end-to-end optimization strategy. The training loss function is constructed based on negative log-partial likelihood. Specifically, the comprehensive feature vectors of samples from each monitoring point are sequentially input into the feature transformation module, and after being transformed layer by layer by a multi-layer neural network, a risk representation vector is output. Then represent the risk vector With integrated feature vector Substitute the degradation risk intensity function into the data to calculate the degradation risk intensity of each monitoring point at its event time. Deterioration risk intensity Define the risk set. In time The partial likelihood function formula for the set of all monitoring points that are still in a state of risk at any given time is as follows:

[0118]

[0119] Taking the negative logarithm of the partial likelihood function yields the training loss function, expressed as:

[0120]

[0121] in, It is the partial likelihood function. The negative log-partial likelihood loss function is used. This is an event indicator variable; it takes a value of 1 when the event occurs and a value of 0 when the event is truncated. For monitoring points The observation time In time The risk set at any given time includes all observation times not less than [a certain value]. The monitoring points.

[0122] The gradient of the training loss function with respect to the neural network weight matrix, bias vector, and degradation risk intensity function parameter set is calculated using the backpropagation algorithm, and all parameters to be learned are updated using stochastic gradient descent or its adaptive variant optimizer.

[0123] The joint training strategy enables the risk representation vector extracted by the feature transformation module to be directly optimized for the degradation risk prediction task. When the coefficient of variation of the comprehensive feature vector of a certain monitoring point is large, the gradient of the training loss function for that monitoring point is suppressed by the uncertainty adjustment term, reducing the fitting weight to unreliable data. When the risk representation vector of a certain monitoring point has a high matching degree with the time-dependent weight vector, the gradient of the training loss function for that monitoring point is enhanced by the time-dependent similarity term, strengthening the extraction capability of time-aware features. The trained neural network automatically learns to distinguish the importance weights of reliable features and unreliable features, and extracts feature representations that match the time-dynamic risk pattern.

[0124] For a preset future time discretization interval, the risk representation vector of each monitoring point is... and comprehensive feature vector Substituting the degradation risk intensity function, the degradation risk intensity values ​​corresponding to each discrete time point are calculated, resulting in a discrete risk score sequence for the monitoring point. Cubic spline interpolation is then performed on the discrete risk score sequence. This cubic spline interpolation, by constructing a piecewise cubic polynomial function, ensures that each discrete point passes through the curve accurately while maintaining its smoothness and continuity. The resulting continuous curve serves as the original degradation risk curve for the monitoring point. The original degradation risk curve, with future time as the horizontal axis and degradation risk intensity as the vertical axis, describes the natural evolution scenario under which no remedial measures are taken at the monitoring point. The trajectory of degradation risk over time was analyzed. Original degradation risk curves were generated for all monitoring points within the garden area, forming a set of original degradation risk curves indexed by the monitoring point number. This set of original degradation risk curves recorded the baseline of degradation risk evolution for each monitoring point under no-intervention conditions.

[0125] For practical applications of garden soil quality improvement, multiple intervention schemes for improvement measures are pre-defined. These measures include common soil improvement techniques such as increasing organic fertilizer application, adjusting irrigation frequency, improving drainage systems, supplementing trace elements, and loosening and aerating the soil. Each improvement measure corresponds to a specific intervention feature vector, encoding key parameters such as the intervention type identifier, intervention intensity level, intervention duration, and intervention implementation time. For any monitoring point... With arbitrary improvement measures The combination of these features is achieved through an attention mechanism, which integrates the feature vectors. and improvement measures The intervention feature vectors are dynamically fused, and the attention mechanism integrates the feature vectors. For the query vector, the attention weight coefficients of each parameter dimension in the intervention feature vector are calculated. These weight coefficients reflect the relative importance of each parameter of the improvement measure to the change in the risk of degradation at the monitoring point. The intervention feature vector is then weighted according to these attention weights and then... The features are concatenated to form an intervention-fused feature vector, denoted as... It also includes the spatiotemporal characteristics of the monitoring point itself and the intervention parameters of the improvement measures.

[0126] Intervention fusion feature vector The feature transformation module of the input survival risk analysis model uses a multi-layer neural network structure to perform intervention fusion feature vector and synthesize the feature vector. The same layer-by-layer linear transformation and nonlinear activation transformation processes are used to generate a risk representation vector for the monitoring point under the intervention of the improved measures in the output layer, denoted as... .

[0127] Risk representation vector Feature vector fusion with intervention Substituting the values ​​into the degradation risk intensity function, we calculate the degradation risk intensity values ​​of the monitoring point at each discrete time point in the future under the intervention of the improvement measures. The baseline risk function of the degradation risk intensity function is then calculated. Regression coefficient vector remains unchanged Remaining unchanged, the risk representation vector is replaced with The comprehensive feature vector is replaced with the intervention fusion feature vector. Used to calculate the coefficient of variation term, time-dependent weight vector It remains unchanged.

[0128] Using the same cubic spline interpolation method as the original degradation risk curve, the discrete risk intensity values ​​under the intervention scenario are interpolated to generate a continuous degradation risk curve for the monitoring point after the intervention of the improvement measures. The degradation risk curve after the intervention reflects the time evolution trajectory of the degradation risk of the monitoring point after the implementation of the improvement measures.

[0129] For each monitoring point, the original degradation risk curve and the post-intervention degradation risk curve are acquired. Within a pre-defined future time interval, both curves are sampled synchronously at fixed time intervals to obtain the risk intensity numerical sequences of the two curves at the same time point. The risk intensity difference between the two curves at each time point is calculated. The risk intensity difference is the difference between the risk intensity value of the original degradation risk curve and the risk intensity value of the post-intervention degradation risk curve. A positive difference indicates that the improvement measures reduced the degradation risk intensity at that time point. The proportion of time points with positive differences within the time interval to the total number of sampling points is used as an indicator of effective coverage. The arithmetic mean of the differences corresponding to all time points with positive differences is calculated, and this arithmetic mean is used as an indicator of average reduction.

[0130] Identify the time point when the degradation risk intensity reaches its peak on the original degradation risk curve, and identify the time point when the degradation risk intensity reaches its peak on the post-intervention degradation risk curve. Calculate the time difference between the post-intervention peak time point and the original peak time point. The time difference serves as an indicator of peak delay time. A positive time difference indicates that the improvement measures have delayed the arrival time of the degradation risk peak.

[0131] On the original degradation risk curve, select the starting and peak times of the degradation risk intensity increase phase, and calculate the average rate of increase of degradation risk intensity during this period. The average rate of increase is defined as the original average rate of increase, calculated by subtracting the risk intensity value at the starting time from the peak time and dividing by the time span. On the post-intervention degradation risk curve, select the starting and peak times of the corresponding degradation risk intensity increase phase, and obtain the average rate of increase using the same calculation method. The relative rate of change between the original average rate of increase and the post-intervention average rate of increase is calculated. The relative rate of change is the original average rate of increase minus the post-intervention average rate of increase, divided by the original average rate of increase. This serves as an indicator of the rate of change. A positive relative rate of change indicates that the improvement measures have slowed the rate of increase of degradation risk.

[0132] The intervention effect of each monitoring point's degradation risk curve under each improvement measure is evaluated by comprehensively considering four evaluation indicators: effective coverage area, average reduction rate, peak delay time, and rate of increase change. If the effective coverage ratio corresponding to a certain improvement measure is higher than the preset lower limit, the average reduction rate is higher than the preset lower limit, and the peak delay time difference is positive or the relative rate of increase change is positive, then the improvement measure is determined to have a significant degradation risk suppression effect on the monitoring point.

[0133] For each monitoring point, all preset improvement measures and intervention plans are traversed, and four evaluation indicators are calculated for each improvement measure: effective coverage ratio, average reduction magnitude, peak delay time difference, and relative rate of increase change. An intervention effect evaluation table for the monitoring point under all improvement measures is generated. The intervention effect evaluation table uses the improvement measure identifier as the row index and the four evaluation indicators as the column index, recording the quantitative intervention effect of each improvement measure at the monitoring point.

[0134] An intervention effect simulation and quantitative evaluation process was conducted at all monitoring points within the park area, resulting in a set of post-intervention degradation risk curves indexed by both the monitoring point number and the improvement measure identifier. Simultaneously, a set of intervention effect evaluation tables indexed by the monitoring point number was generated. The intervention effect evaluation table set also includes curve-derived parameters for each monitoring point, such as the current degradation radius, the peak risk intensity value of the original degradation risk curve, and the average slope of the rising phase of the original degradation risk curve.

[0135] S4: Read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraint rules, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a set of differentiated improvement strategies.

[0136] For all monitoring points within the park area, the intervention effect evaluation table is read sequentially. All records of improvement measures in the intervention effect evaluation table are traversed, and the values ​​of the four evaluation indicators corresponding to each improvement measure are extracted.

[0137] Meanwhile, the current deterioration radius, the risk intensity value at the peak of the original deterioration risk curve, and the average slope of the rising stage of the original deterioration risk curve of the monitoring point are read from the intervention effect evaluation table, and the soil deterioration type label of the monitoring point is read from the monitoring point attributes.

[0138] The specific rules for determining whether the improvement measure meets the criteria for significant inhibitory effect are as follows:

[0139] Read the soil degradation type label of the monitoring point. If the soil degradation type label of the monitoring point is acidification degradation, the effective coverage area of ​​the improvement measures must be greater than 1.5 times the current degradation influence radius of the monitoring point, the average reduction rate must be greater than 0.3 times the risk intensity value of the original degradation risk curve at the peak time, and the peak delay time must be greater than zero.

[0140] If the soil degradation type label of the monitoring point is salinization degradation, then the effective coverage area index of the improvement measures must be greater than twice the current degradation influence radius of the monitoring point, the average reduction index must be greater than 0.4 times the risk intensity value of the original degradation risk curve at the peak moment, and the absolute value of the rate of increase index must be greater than 0.2 times the average slope of the rising stage of the original degradation risk curve.

[0141] If the soil degradation type label for the monitoring point is compaction degradation, then the effective coverage area of ​​the improvement measures must be greater than the current degradation impact radius of the monitoring point, the average reduction rate must be greater than 0.25 times the risk intensity value of the original degradation risk curve at the peak moment, and the peak delay time must be greater than zero or the absolute value of the rate of increase change must be greater than 0.15 times the average slope of the rising stage of the original degradation risk curve.

[0142] If a certain improvement measure meets the judgment conditions of the corresponding degradation type, then the identification of the improvement measure, the corresponding values ​​of the four evaluation indicators, the current degradation influence radius of the monitoring point, the risk intensity value at the peak of the original degradation risk curve, and the average slope of the rising stage of the original degradation risk curve are recorded in the candidate solution set.

[0143] If a certain improvement measure does not meet the judgment criteria for the corresponding degradation type, then the improvement measure will not be included in the candidate solution set. After traversal, the candidate solution set corresponding to the monitoring point is obtained.

[0144] For each monitoring point, the improvement measures in the candidate solution set are retrieved from the pre-set improvement measure cost database based on the improvement measure identifier. The improvement measure cost database, indexed by the improvement measure identifier, stores the unit area implementation cost parameter, manpower requirement parameter, material consumption parameter, and implementation cycle parameter for each improvement measure.

[0145] Multiply the retrieved unit area implementation cost parameter by the deterioration impact area of ​​the monitoring point to obtain the total implementation cost of the improvement measure at the monitoring point. The deterioration impact area is calculated by the current deterioration impact radius of the monitoring point according to the circular area formula.

[0146] The retrieved implementation cycle parameter is compared with the degradation urgency level of the monitoring point. If the implementation cycle parameter is greater than the maximum allowable implementation cycle corresponding to the degradation urgency level, the improvement measure is marked as a cycle-exceeding scheme.

[0147] For each monitoring point, the improvement measures in the candidate scheme set are retrieved from the preset implementation constraint rule base according to the improvement measure identifier.

[0148] The implementation constraint rule base uses the improvement measure identifier as an index and stores the applicable soil pH range, applicable soil moisture range, applicable temperature range, prohibited construction month set, and mutually exclusive improvement measure set for each improvement measure.

[0149] If the soil pH, soil moisture content, and air temperature are all within their respective applicable ranges, and the current month is not a prohibited construction month, and there are no mutually exclusive improvement measures in the list of planned improvement measures, then the improvement measure is retained in the candidate solution set. If any of the above conditions are not met, or the improvement measure is marked as a period-exceeding measure, then the improvement measure is removed from the candidate solution set. After filtering by the implementation constraint rules, the set of valid solutions corresponding to the monitoring point is obtained. For each monitoring point, the vegetation configuration density, visitor activity frequency, landscape viewing value rating, and ecosystem sensitivity rating data already labeled in S1 are read. The ratio of vegetation configuration density to visitor activity frequency is calculated to obtain the ecological load ratio, and the ratio of landscape viewing value rating to ecosystem sensitivity rating is calculated to obtain the function orientation ratio.

[0150] The garden functional area type labels are determined in order of priority:

[0151] If the ecological load ratio is greater than 1 and the ecosystem sensitivity rating is greater than the landscape appreciation value rating, then the garden functional area type label for this monitoring point is assigned as ecological conservation area.

[0152] If the above conditions are not met and the functional orientation ratio is greater than 1 and the frequency of visitor activities is lower than the vegetation configuration density, then the garden functional area type label for this monitoring point is assigned as a landscape display area.

[0153] If neither of the above two conditions is met, the monitoring point will be assigned the garden functional area type label as an activity and leisure area.

[0154] For each monitoring point, the average reduction rate weighting coefficient and peak delay time weighting coefficient are determined based on the garden functional area type label of the improvement measures in the effective scheme set:

[0155] If the garden functional area type label is landscape display area, then the difference between the risk intensity value at the peak moment of the original deterioration risk curve of the monitoring point and the landscape quality threshold is calculated and divided by the landscape quality threshold to obtain the landscape deviation degree. The cumulative number of tourist complaints is calculated and divided by the total number of tourists in the monitoring period to obtain the complaint rate. The landscape deviation degree and the complaint rate are normalized and then added together. The normalized landscape deviation degree is calculated and divided by the sum of the two to obtain the landscape deviation degree ratio and assigned to the average reduction magnitude weight coefficient. The average reduction magnitude weight coefficient is subtracted from 1 to obtain the peak delay time weight coefficient.

[0156] If the garden functional area type label is ecological conservation area, then the recovery urgency coefficient is obtained by dividing the time span from the current moment to the peak moment of the original deterioration risk curve of the monitoring point by the natural recovery cycle. The ecological health degree is obtained by dividing the soil microbial diversity index by the healthy soil microbial diversity benchmark index. The recovery margin is obtained by subtracting the recovery urgency coefficient from the value 1. The recovery margin and the ecological health degree are normalized and then added together. The ecological health degree after normalization is calculated and divided by the sum of the two to obtain the ecological health degree ratio and assigned to the peak delay time weight coefficient. The average reduction magnitude weight coefficient is obtained by subtracting the peak delay time weight coefficient from 1.

[0157] If the garden functional area type label is activity and leisure area, the load rate is obtained by dividing the frequency of visitor activities by the carrying capacity. The latest soil hardness value is extracted from the soil quality time series monitoring data of this monitoring point as the measured soil hardness data. The absolute value of the difference between the measured soil hardness data and the standard value of soil hardness is calculated and divided by the standard value of soil hardness to obtain the hardness deviation. The load rate and the hardness deviation are normalized and added together. The normalized load rate is calculated and divided by the sum of the two to obtain the load rate ratio and assigned to the average reduction magnitude weighting coefficient. The average reduction magnitude weighting coefficient is subtracted from 1 to obtain the peak delay time weighting coefficient.

[0158] For each improvement measure in the effective solution set, the average reduction rate index value and the peak delay time index value of that improvement measure are read. The sum of the average reduction rate index value multiplied by the average reduction rate weighting coefficient and the peak delay time index value multiplied by the peak delay time weighting coefficient is calculated to obtain the weighted comprehensive score value of that improvement measure. All improvement measures in the effective solution set are sorted in descending order of their weighted comprehensive scores, and each improvement measure is assigned a ranking number starting from 1 and increasing sequentially, forming a ranking list of improvement measures for that monitoring point.

[0159] Based on the order of the improvement measures in the ranking list, the improvement measures are divided into layers, specifically:

[0160] The total number of improvement measures in the effective scheme set for this monitoring point is set to be... Obtain the stratification ratio parameters of the garden area where the monitoring point is located. The stratification ratio parameters include the front section ratio. and the proportion of the latter part ,in The value ranges from 0.2 to 0.4. The value ranges from 0.2 to 0.4, and + ≤1, the proportion of the middle section is 1- - Calculate the position value of the front boundary segment. Multiply Round down to the nearest integer and calculate the position value of the middle boundary. Multiply by (1- Then round down to the nearest integer.

[0161] Iterate through each improvement measure in the sorted list and read its sort order number:

[0162] If the ranking number is less than or equal to the previous boundary number, the strategy type label of the improvement measure is assigned as a priority improvement strategy, and the improvement measure identifier, evaluation index value, total implementation cost, and ranking number are recorded in the priority improvement strategy list.

[0163] If the ranking number is greater than the value of the previous boundary position and less than or equal to the value of the middle boundary position, then the strategy type label of the improvement measure is assigned as a preventive improvement strategy, and the improvement measure identifier, evaluation index value, total implementation cost, and ranking number are recorded in the preventive improvement strategy list.

[0164] If the ranking number is greater than the middle boundary position value, the strategy type label of the improvement measure is assigned as maintenance enhancement strategy, and the improvement measure identifier, evaluation index value, total implementation cost, and ranking number are recorded in the maintenance enhancement strategy list.

[0165] The above-mentioned candidate scheme screening, constraint filtering, weighted comprehensive scoring and ranking, and hierarchical division process were performed on all monitoring points within the garden area to generate a priority improvement strategy list, a prevention improvement strategy list, and a maintenance enhancement strategy list for each monitoring point.

[0166] A differentiated improvement strategy set data structure is constructed, with monitoring point number as the first-level index, strategy type label as the second-level index, and improvement measure identifier as the third-level index. Stored fields include evaluation indicator numerical field, total implementation cost field, sorting order field, soil degradation type label field, and garden functional area type label field. Records from the three strategy lists for each monitoring point are sequentially written into the corresponding index positions of the differentiated improvement strategy set data structure. The differentiated improvement strategy set data structure is output, allowing garden management departments to query the specific improvement measures to be taken at each monitoring point under different priority levels, along with their technical and cost parameters.

[0167] Here, those skilled in the art will understand that the specific operations of each step in the above-described method for dynamic monitoring and improvement of garden soil quality have been referenced above. Figures 1 to 2 The description of the dynamic monitoring and improvement methods for garden soil quality is detailed here, and therefore, its repeated description will be omitted.

[0168] According to another aspect of the present invention, a dynamic monitoring and improvement system for garden soil quality is provided, comprising:

[0169] The feature encoding module is used to receive soil quality time-series monitoring data from each monitoring point, construct a time-series data matrix, perform multi-scale sliding window segmentation, extract the comprehensive magnitude of the gradient and fluctuation amplitude of the index change within each window, perform feature encoding on the historical evolution pattern, and obtain the time-series evolution feature vector.

[0170] The relationship modeling module is used to calculate the pattern similarity, causal transmission strength and directionality of the temporal evolution feature vectors between monitoring points, construct a two-layer transmission relationship graph, perform layered graph propagation and diffusion on the two-layer transmission relationship graph and correct it through physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector.

[0171] The effect evaluation module is used to merge the temporal evolution feature vector with the spatiotemporal transmission latent state vector, generate the original and post-intervention degradation risk curves through survival risk analysis, compare and extract effect indicators, and generate an intervention effect evaluation table.

[0172] The strategy generation module is used to read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraint rules, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a differentiated strategy set.

[0173] In summary, the method for dynamic monitoring and improvement of garden soil quality based on the embodiments of the present invention has been clarified. By performing multi-scale sliding window segmentation and feature encoding on the time-series monitoring data of soil quality from multiple monitoring points, it can comprehensively extract the changing gradient and fluctuation patterns of soil quality indicators. Utilizing the similarity and causal transmission relationship of the time-series evolution characteristics among monitoring points, a two-layer transmission relationship diagram is constructed and stratified correction is performed, revealing the spatiotemporal evolution law and correlation mechanism of soil quality. Combined with survival risk analysis and intervention effect evaluation, it achieves early identification and quantitative analysis of soil degradation risk. By comprehensively evaluating the intervention effect and the cost of improvement measures, differentiated stratified improvement strategies are generated, improving the accuracy of improvement measures and resource utilization efficiency, thereby effectively solving the problems of fixed monitoring, delayed response, and low improvement efficiency in existing technologies.

Claims

1. A method for dynamic monitoring and improvement of garden soil quality, characterized in that, include: The soil quality time-series monitoring data from each monitoring point is received, a time-series data matrix is ​​constructed, multi-scale sliding window segmentation is performed, the comprehensive magnitude of the gradient and fluctuation amplitude of the index change within each window is extracted, the historical evolution pattern is feature-encoded, and the time-series evolution feature vector is obtained. The pattern similarity, causal transmission strength, and directionality of the temporal evolution feature vectors among each monitoring point are calculated. A two-layer transmission relationship diagram is constructed. The two-layer transmission relationship diagram is subjected to layered graph propagation and diffusion and corrected by physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector. The temporal evolution feature vector and the spatiotemporal transmission latent state vector are merged, and the original and post-intervention degradation risk curves are generated through survival risk analysis. The effect indicators are compared and extracted to generate an intervention effect evaluation table. Read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraints, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a set of differentiated strategies.

2. The method for dynamic monitoring and improvement of garden soil quality according to claim 1, characterized in that, The generation of the temporal evolution feature vector includes: The comprehensive magnitude sequence of the same soil quality index is grouped by window scale, and the comprehensive magnitude sequences corresponding to short-period windows, medium-period windows and long-period windows are extracted. The statistical characteristics of the sequence at each scale are calculated respectively. The statistical features include the sequence mean, sequence variance, sequence extreme values, and sequence trend slope; Statistical features at different window scales are concatenated in the order of short period, medium period, and long period to form a multi-scale statistical feature vector. The multi-scale statistical feature vectors of all soil quality indicators at the same monitoring point are jointly spliced ​​and z-score standardized.

3. The method for dynamic monitoring and improvement of garden soil quality according to claim 2, characterized in that, The two-layer transmission relationship diagram includes an undirected similarity relationship layer and a directed causal transmission layer; The undirected similarity relation layer calculates the cosine similarity of the temporal evolution feature vector of the monitoring points and maps it to a non-negative value range through exponential transformation to obtain the pattern similarity intensity value. The monitoring point pairs with the pattern similarity intensity value exceeding the preset similarity threshold are used as undirected edges and the pattern similarity intensity value is used as the edge weight. The directed causal transmission layer obtains the F-statistic and significance probability values ​​through the Granger causality test. The F-statistic is then transformed by logarithmic transformation and linear scaling to map to a non-negative value range to obtain the causal transmission strength value. Directional causal labels are established for monitoring point pairs that meet the significance condition, and the causal transmission strength value is used as the edge weight.

4. The method for dynamic monitoring and improvement of garden soil quality according to claim 3, characterized in that, The formula for calculating the physical constraint correction coefficient is as follows: in, Source monitoring point Pointing to the target monitoring point The physical constraint correction coefficients corresponding to the directed edges. For monitoring points With monitoring points The Euclidean straight-line distance between them. Source monitoring point The set of target monitoring points connected by all outgoing edges in the directed transmission graph. The spatial decay scale parameter of the Gaussian kernel function. For monitoring points With monitoring points Soil type compatibility coefficient between them As a terrain orientation preference factor, For target monitoring points Vegetation buffer factor, This is a constant used to protect the denominator in the normalization process.

5. The method for dynamic monitoring and improvement of garden soil quality according to claim 4, characterized in that, The survival risk analysis model includes a feature transformation module and a degradation risk intensity function; The feature transformation module uses a multi-layer neural network structure to transform the comprehensive feature vector into a risk representation vector; The formula for the degradation risk intensity function is: in, For the baseline risk function, For the regression coefficient vector, The risk representation vector output by the feature transformation module; The coefficient of variation of the comprehensive feature vector; The uncertainty adjustment index was selected through cross-validation. This is a time-dependent weight vector; is the numerical stability constant.

6. The method for dynamic monitoring and improvement of garden soil quality according to claim 5, characterized in that, The survival risk analysis model is jointly trained using historical monitoring data. The training loss function is constructed based on negative log-partial likelihood, and the formula is as follows: in, It is the partial likelihood function. The negative log-partial likelihood loss function is used. This represents the number of historical monitoring points. This is an event indicator variable; it takes a value of 1 when the event occurs and a value of 0 when the event is truncated. For monitoring points The observation time In time The risk set at any given time includes all observation times not less than [a certain value]. monitoring points Indicates the first risk in the risk set Each monitoring point at time The value of the degradation risk intensity function at any given time.

7. The method for dynamic monitoring and improvement of garden soil quality according to claim 6, characterized in that, The generation of the post-intervention degradation risk curve includes: The attention mechanism is used to synthesize the feature vector. Dynamically fuse the intervention feature vector with the intervention feature vector of the improvement measures to form an intervention fusion feature vector. ; The intervention fusion feature vector is input into the feature transformation module to obtain the risk representation vector of the monitoring point under the intervention of the improved measures. ; Substitute the risk representation vector and the intervention fusion feature vector into the degradation risk intensity function to calculate the degradation risk intensity value at each discrete time point in the future; Cubic spline interpolation is used to generate the post-intervention degradation risk curve from the discrete risk intensity values.

8. A dynamic monitoring and improvement system for garden soil quality, characterized in that, include: The feature encoding module is used to receive soil quality time-series monitoring data from each monitoring point, construct a time-series data matrix, perform multi-scale sliding window segmentation, extract the comprehensive magnitude of the gradient and fluctuation amplitude of the index change within each window, perform feature encoding on the historical evolution pattern, and obtain the time-series evolution feature vector. The relationship modeling module is used to calculate the pattern similarity, causal transmission strength and directionality of the temporal evolution feature vectors between monitoring points, construct a two-layer transmission relationship graph, perform layered graph propagation and diffusion on the two-layer transmission relationship graph and correct it through physical constraint correction coefficients to obtain the spatiotemporal transmission hidden state vector. The effect evaluation module is used to merge the temporal evolution feature vector with the spatiotemporal transmission latent state vector, generate the original and post-intervention degradation risk curves through survival risk analysis, compare and extract effect indicators, and generate an intervention effect evaluation table. The strategy generation module is used to read the intervention effect evaluation form, combine it with the cost library of improvement measures and implementation constraint rules, sort and stratify it according to the weighted comprehensive score of reduction magnitude and delay time, and output a differentiated strategy set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for dynamic monitoring and improvement of garden soil quality as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for dynamic monitoring and improvement of garden soil quality as described in any one of claims 1 to 7.

Images