Regionalized soil fertility space-time evolution prediction method and system

Abstract

The invention provides a regional soil fertility space-time evolution prediction method and system, and relates to the field of data processing prediction.The method comprises the steps that agricultural management events such as fertilization and irrigation are captured in real time through an edge computing node cluster, and an asymmetric time window is constructed based on a soil permeation rate model; extracting meteorological and satellite data in the window, querying the coupling rule table by using a triple index, and generating a coupling factor data set with a semantic mark; calculating a sliding time variance of a coupling factor in a grid unit, marking a high-entropy region and starting real-time stream processing when the sliding time variance exceeds a threshold value, and generating a soil state Hash fingerprint and performing dormancy collection when the sliding time variance is lower than the threshold value; constructing a regional event path directed graph, selecting a similar neighborhood event chain as a prediction framework, and registering an empty window region; the neighborhood topology is scanned in a buffer zone of a window area, after a compliance event sequence is captured, a window is removed, similar neighborhood historical data is migrated and injected into a local sequence, and parameters are finely adjusted through incremental learning to reconstruct an exclusive prediction model.

Description

Technical Field

[0001] This invention relates to the field of data processing and prediction, specifically to a method and system for predicting the spatiotemporal evolution of regional soil fertility. Background Technology

[0002] Dynamic prediction of soil fertility is a core aspect of precision agriculture management, directly impacting crop yield and resource utilization efficiency. With the development of satellite remote sensing and Internet of Things technologies, real-time acquisition of multi-source heterogeneous data provides a foundation for regional soil analysis. However, agricultural systems are subject to dual nonlinear interference from natural factors and human activities, making it difficult for traditional time-series modeling methods to capture the complex relationships between sudden events and delayed effects.

[0003] The mainstream approach adopts a fixed-period sampling strategy, aggregating data from meteorological stations, satellite sensors, and manual reporting on a monthly / ten-day basis; a soil parameter prediction framework is established through time series models or statistical regression methods; some systems introduce a weighted fusion mechanism to process multi-source data and trigger model retraining based on preset thresholds to cope with environmental changes.

[0004] The shortcomings of existing technologies are: fixed time slices cause the instantaneous impact of key human operations to be diluted by periodic averages; the data fusion process ignores the contextual relationship between event types and soil properties, resulting in a distorted expression of the causal contribution of natural elements; and the forced continuation of predictions in areas where historical patterns have failed lacks a logical legitimacy interruption mechanism, exacerbating the accumulation of prediction bias. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the spatiotemporal evolution of regional soil fertility. This addresses the problems mentioned in the background technology, such as the dilution of the instantaneous impact of key human-induced events by periodic averages due to fixed-time slicing; the neglect of the contextualized relationship between event types and soil properties during data fusion, leading to distorted expression of the causal contribution of natural elements; and the forced continuation of predictions in areas where historical patterns have failed, lacking a logical legitimacy interruption mechanism, which exacerbates the accumulation of prediction bias.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: a method and system for predicting the spatiotemporal evolution of regional soil fertility, comprising the following steps: S1: deploying an edge computing node cluster and embedding an agricultural rule knowledge base; capturing agricultural management event streams in real time through a distributed event monitoring module; when fertilization events, irrigation events, tillage events, or land use change events are identified, generating time anchors at the event trigger time; determining forward and backward offsets based on a soil infiltration rate model associated with event types; and constructing an asymmetric time window centered on the anchors.

[0009] S2: Extract time series data of natural elements from meteorological stations and satellite sensors within the asymmetric time window, combine event type codes, soil type codes and post-event delay periods into triplet indexes, query the pre-compiled coupling rule lookup table to obtain the weight coefficients of natural elements, and generate a coupling factor dataset with semantic tags.

[0010] S3: Calculate the sliding time variance of the coupling factors in the coupling factor dataset within each geographic grid cell. When the sliding time variance exceeds the stability threshold, it is marked as a high-entropy region and the real-time stream processing pipeline is started. When the variance is lower than the dormancy threshold, a soil state hash fingerprint is generated. Data acquisition is interrupted by fingerprint comparison until the change signal is triggered by the comparison of satellite image texture features.

[0011] S4: Construct a directed graph of regional event paths to generate event path topology, select the neighborhood event chain with the highest topological similarity and load it as a temporary prediction framework; when it is detected that the interval between adjacent events is less than the minimum allowed interval defined by the agricultural rule knowledge base or the event combination violates the mutual exclusion constraint of the agricultural rule knowledge base, freeze the prediction engine of the region and register it as a blank window area;

[0012] S5: Scan the event path topology within the buffer area formed by extending a set distance outside the geographical boundary of the empty window area, select the neighborhood event chain with the highest topological similarity and load it as a temporary prediction framework. When a compliant event sequence that has been captured for a set number of consecutive times is captured, the empty window state is lifted, the historical training dataset of similar neighborhoods is migrated to the empty window area, the local coupling factor sequence before the empty window area is lifted is injected, the model parameters are fine-tuned using an incremental learning algorithm, and the exclusive prediction model is reconstructed.

[0013] Preferably, in S1, the edge computing node cluster is deployed in a grid-like distribution across a geographic area. Each node possesses independent data acquisition and processing capabilities as well as rule judgment capabilities. It integrates a local database and embedded computing chip via a microprocessor, and pre-loads an agricultural rule knowledge base to support localized judgment and response. After deployment, a distributed event monitoring module captures agricultural management event streams in real time. This module includes four independent event monitoring channels, corresponding to the agricultural management system application interface log channel, the agricultural machinery IoT operation signal channel, the multispectral satellite image interpretation event channel, and the manually labeled data channel. Each channel is parsed into standard event messages by the edge node's local access protocol converter and transmitted to the event recognition engine. Specifically, the application interface log channel identifies events by parsing the timestamp and event type encoding fields in the API logs; the agricultural machinery IoT operation signal channel converts the acquired signal stream into event messages using the standardized protocol stack within the edge computing node; and the multispectral satellite image channel outputs plot change information in the image preprocessing module using texture segmentation and change detection algorithms to identify and interpret events. The manually labeled data channel is connected to a handheld terminal to collect information filled in by users through electronic forms and convert it into structured event content. Once the event recognition engine confirms the event, it immediately generates a time anchor point. This time anchor point is registered with the event recognition time as a reference time point and serves as the starting point for subsequent data window construction. After the anchor point is generated, based on the soil infiltration rate model corresponding to the event type preset in the agricultural rule knowledge base, the forward and backward offsets are determined. The forward offset is dynamically calculated based on the soil moisture change rate before the event occurs, within the daily soil moisture data. When the rate of change exceeds a set value of 5%, the forward offset is set to 48 hours; if it is below this rate, it is set to 24 hours. The backward offset is calculated based on the event type and the persistent influence factor. For example, in irrigation events, the backward offset is determined by the soil saturated hydraulic conductivity. The backward offset for sandy soil is set to 72 hours, for loam to 96 hours, and for clay to 120 hours. Finally, an asymmetric time window centered on the anchor point and constrained by the event type characteristics is constructed for subsequent extraction and coupling analysis of time series data of natural elements. The asymmetric window supports the semantic enhancement analysis mechanism of event-driven time slices.

[0014] Preferably, in S2, after the asymmetric time window is generated, time-series data of natural elements needs to be extracted within the window. These natural elements include data from two sources: meteorological station data and satellite sensor data. Meteorological station data is acquired periodically through local meteorological station equipment connected to the edge computing node, and includes multi-dimensional elements such as temperature, rainfall, humidity, wind speed, air pressure, and evaporation. Satellite sensor data originates from high-resolution multispectral remote sensing image sequences. These remote sensing images are sliced ​​at a three-hour resolution within the asymmetric time window, and surface reflectance, vegetation index, texture variation values, and thermal infrared radiation characteristic values ​​are extracted through the image interpretation module to form a sequence representation with geographic coordinates. After acquiring the asymmetric time window data... After processing all natural element data within the symmetrical time window, a triplet index is constructed by combining the event type code, soil type code, and post-event delay period. The event type code, in integer form, corresponds to fertilization events, irrigation events, tillage events, or land use change events, numbered one to four respectively. The soil type code is expressed using two-letter codes, with sandy soil being SA, loam being LO, and clay being CL. The post-event delay period is discretely divided into hourly units. The triplet index is used to query a pre-compiled coupled rule lookup table, which is organized in a hierarchical tree structure. The root node index area carries the event type code, and the child node index areas store the corresponding soil type codes. The point data area forms a hash dictionary structure, using the post-event delay period as the key and the natural element weight coefficient value as the value for key-value pair compilation. The weight coefficient is stored in fixed-precision floating-point format, ranging from 0.1 to 1; the specific value is set by agricultural experts based on field experiment results during the design phase. After receiving the triplet index, the edge computing node performs a layer-by-layer matching operation locally. First, it reads the event type code from the root node to match the subtree, then retrieves the soil type code from the child nodes to enter the target leaf node, and finally retrieves the corresponding weight coefficient based on the post-event delay period. The entire query process is completed within 200ms. The query results are then compared with the natural element data for the corresponding time period. Weighted processing is performed to generate a sequence of coupling factor values. Semantic tags are attached to each coupling factor, which are generated according to the coupling rules. For example, an infiltration promotion label is attached to the coupling factor corresponding to sandy soil in the rainfall data, and a retention risk label is attached to the coupling factor corresponding to clay soil in the same rainfall data. A water loss risk label is attached to the wind speed index when the evaporation intensity is more than twice the average value. The resulting coupling factor dataset not only retains the numerical information of natural elements, but also injects event context and soil attribute dependence into the structure, forming a high-dimensional time series structure with contextual meaning. This provides semantically supported multi-source fusion basic data for subsequent time variance analysis and state assessment.

[0015] Preferably, in S3, based on the constructed coupling factor dataset, a sliding time variance calculation is performed on the coupling factor sequence within each geographic grid cell. This calculation uses a fixed-length time segment rolling method, dividing the segment length into units based on the agricultural operation cycle. The time segment length is preset to 72 hours for field crops and 48 hours for facility agriculture. Each geographic grid cell traverses all segments sequentially, calculating the mean and squared deviation of the coupling factors within each time segment, and outputting the variance value within the current window as an activity index. This index is used during the rolling process. The data is continuously updated online and stored in the time-series buffer of the edge computing node. When the sliding time variance exceeds a set stability threshold, the current geographic grid cell is determined to be in a high-entropy state. The stability threshold is triggered when the standard deviation of the coupling factor exceeds 0.3 for more than three consecutive time segments. Grid cells in a high-entropy state have their real-time stream processing channel activated on the edge node. This channel calls three types of input sources: satellite remote sensing data, ground sensor data, and manually reported data, and performs full-volume coupled computation and real-time analysis. During this process, the data sampling frequency is increased by three times compared to the basic sampling frequency to ensure... The system rapidly captures the diffusion trend of the impacts caused by drastic changes in natural and human interactions; conversely, when the sliding time variance is continuously below the dormancy threshold, the grid is marked as a low-entropy state. The dormancy threshold is triggered when the standard deviation of the coupling factor is below 0.1 and the duration of the segment is greater than 5. In the low-entropy state, real-time data stream processing is no longer performed; instead, a hash fingerprint of the current soil state is generated. This hash fingerprint is generated using a structured summary method and consists of three elements: the mean of soil organic matter content, the standard deviation of soil pH, and the temporal slope of water content. It serves as a unique identification segment representing the current soil state of the region. After the texture is generated, the data acquisition task is interrupted until a change signal is detected. The change signal is triggered by the satellite image texture feature comparison module. When the surface texture change index in the comparison result deviates from the expected threshold of the hash fingerprint by more than 20%, the data acquisition channel is reactivated. At the same time, the fingerprint comparison under the dormant threshold adopts an incremental update mechanism. That is, only when the crop growth index interpreted from the satellite image and the structural summary shown by the cached hash fingerprint continuously produce significant deviations, the meteorological station data acquisition module is activated to supplement the judgment basis, avoiding repeated processing of invalid data in low-change areas, thereby realizing hierarchical sampling and differentiated processing in geospatial space.

[0016] Preferably, in S4, after the real-time processing channel is activated in the high-entropy state region, a directed graph of regional event paths is constructed according to the occurrence order of agricultural management events. The event path construction method is to extract the registered event type codes and corresponding timestamps within a specified geographic grid unit from the event database, arrange them in chronological order to form a node sequence, and each event node represents an agricultural behavior event. Nodes are connected in a directed manner according to the order in which the events occur to form a path. The constructed path is stored in a directed acyclic graph structure, and the weight value of the directed edge is assigned by the minimum allowable interval between events defined in the agricultural rule knowledge base. For example, the minimum allowable interval between fertilization and irrigation events is 48 hours, and irrigation events are prohibited within 72 hours after land tillage events. After generating the complete event path topology, consistency checks are performed on all connected edges. The mutual exclusion constraint detector traverses the time interval of each pair of adjacent nodes in the graph structure one by one and compares it with the minimum interval weight value of the edge. When the time interval between two event nodes is detected to be less than their corresponding minimum allowable interval, the first interrupt signal is immediately triggered to indicate a time sequence conflict. At the same time, it is further determined whether there is an event defined in the agricultural rule knowledge base. The system addresses the conflict relationship between components. When a mutually exclusive event combination connection edge is detected in the path, a second interrupt signal is triggered to indicate an agricultural logical structure conflict. These mutually exclusive event combinations include, but are not limited to, scenarios such as the coexistence of land use change events and planting events, the interval between fertilizer application events and organic fertilizer application events being less than 48 hours, and the same crop being planted continuously on the same plot more than three times. Once any of these interrupt signals occurs, the prediction engine corresponding to the current geographic grid unit is frozen, prohibiting further prediction based on the historical event sequence of that area. The freezing operation marks the area as a void area through edge node registration instructions, simultaneously suspending all data acquisition and prediction model execution processes in that area. Subsequent processing logic will rely on the void area recovery mechanism for logical migration judgment. In the void area state, prediction results are no longer pushed; instead, the system enters a structural observation and logic buffer mode, maintaining event monitoring but not participating in prediction decisions until the release conditions are met. The entire operation process ensures that in the event of a continuous interruption of agricultural behavior or a logical structure failure, invalid prediction behavior can be stopped, avoiding misleading decision outputs from historical models for abrupt changes, and providing a rigorous constraint basis for the integrity control of the prediction logical structure.

[0017] Preferably, in S5, when a region is registered as a blank area during the prediction process due to logical conflicts in event paths, to avoid the region remaining in a state of prediction interruption for a long time, a buffer zone needs to be formed by extending a certain distance outside its geographical boundary to find potential logical migration paths. The extension distance of the buffer zone is preset according to the land parcel type and crop type, generally within the range of 2 to 3 adjacent geographical grid units. Within the buffer zone, the directed graph structure of all event paths is scanned, and the neighborhood event chain with the highest topological similarity is selected as the candidate logical path. The topological similarity calculation adopts the dynamic time warping algorithm, and the input is the continuous event chain before the interruption of the blank area. The first input sequence, consisting of the timestamps and event type codes of the event sequences, and the second input sequence, consisting of similarly structured sequences of neighboring event chains within the buffer area, are used as inputs. The regularization algorithm aligns the two event path sequences along the time axis and event type dimension using a morphological distance calculation unit, identifies key peaks and valleys in each sequence, and outputs the minimum regularized path distance. The smaller this distance, the more similar the behavioral logic of the two event chains. Finally, the neighboring event chain with the lowest minimum regularized path distance is selected as the temporary prediction framework and loaded into the empty window area. After the temporary prediction framework is loaded, all real-time data inputs within the empty window area are redirected to this similar neighboring event chain. The data processing pipeline begins by labeling the original prediction engine's output as a migration calculation result to distinguish it from the in-situ model's generated result. Building upon this, it continues to monitor the event stream within the empty window region. When a compliant event sequence that meets the event interval threshold and coupling factor fluctuation range constraints is continuously captured, and this is repeated a set number of times, the empty window state is lifted. Specifically, the criteria are that the interval between at least three consecutive event groups is not less than the minimum allowable interval defined by agricultural rules, and the sliding time variance of their corresponding coupling factors is stable between 0.1 and 0.3. After lifting the empty window state, the entire historical training dataset from the selected neighborhood is migrated to the empty window region and injected... Before the window region is removed, the accumulated local coupling factor sequence is used to fine-tune the model parameters using an incremental learning algorithm. Under the premise of unchanged model structure, the parameters are optimized by using the transfer dataset as the base parameter and the local data as the incremental compensation, and a new dedicated prediction model is constructed. The reconstructed prediction model inherits the structural settings of the original similar neighborhood region model and adds local adaptation training for the window region. After startup, it is connected to the main prediction process as an independent prediction engine for the window region. This engine supports real-time input and has local response capabilities, realizing a complete closed-loop path from logical interruption to prediction recovery, ensuring that the regional model update behavior has interpretability and continuous stability.

[0018] Preferably, the distributed event monitoring module deployed on the edge computing node in S1 has four parallel data channels: an agricultural management system application interface log channel, an agricultural machinery IoT operation signal channel, a multispectral satellite image interpretation event channel, and a manually labeled data channel. Each channel completes port binding and protocol adaptation configuration during the initialization phase of the event monitoring module. The agricultural management system log channel accesses the agricultural management platform log file through a standard HTTP interface, parses the event timestamp field and event type encoding field in the log records, and when a fertilization event, irrigation event, tillage event, or land use change event is detected, the event information is structured into a standard event message and pushed to the event identification. The system comprises several layers: an engine and an event recognition engine. The agricultural machinery IoT channel receives control commands or operational statuses from agricultural machinery, converts them into a unified format of equipment status change information via an industrial communication protocol converter, and injects this information into the event recognition engine. This ensures that agricultural machinery operations can be considered independent event sources for monitoring and judgment. A multispectral satellite image interpretation event channel retrieves remote sensing image sequences based on geographic coordinates. The image preprocessing module uses texture segmentation algorithms to identify plot boundaries and employs change detection algorithms to compare texture differences between time-series images. When the degree of change exceeds a set threshold of 15%, it is marked as a plot status change event. A manually labeled data channel receives operation behavior registration information filled out by farmers or agricultural administrators through a terminal data entry interface. The fields include operation type, plot number, start time, and end time. Received data undergoes local field validation and sequence conflict resolution before being incorporated into the event identification process. After event identification, the asymmetric time window construction phase begins. The forward offset of the asymmetric time window is calculated based on the soil moisture change rate before the event. If the soil moisture change rate exceeds 5%, the forward offset is set to 48 hours; if it is below 3%, it is shortened to 24 hours. The backward offset is determined by the event's persistence factor. For irrigation events, the backward offset is determined by the soil's saturated hydraulic conductivity, set to 72 hours for sandy soil, 96 hours for loam, and 120 hours for clay. These parameters are derived from the agricultural rule knowledge base. The process is determined by laboratory data. It dynamically adjusts the time window length using a soil infiltration rate model, adapting the time window spatially to soil heterogeneity and temporally to the event response process, thus generating an asymmetric window with event semantic attributes. Based on this, when entering the coupling factor generation stage, the triplet index is submitted to a hierarchical coupling rule lookup table. Nodes are matched layer by layer, and weight coefficients are extracted. Semantic tags are also attached, with the tags set according to the combination of event type and soil type. For example, the same rainfall is marked as infiltration promotion in sandy soil and retention risk in clay soil. This semantic information is written into the coupling factor dataset along with the coupling factor, ensuring that subsequent processing steps have context-sensitive capabilities.

[0019] Preferably, in the process of constructing the semantically labeled coupling factor dataset in S2, the structure design of the coupling rule lookup table adopts a hierarchical tree storage method to improve query efficiency and matching accuracy in the edge computing node environment. The lookup table is divided into three layers, corresponding to event type encoding, soil type encoding, and post-event delay time mapping, respectively. In the initialization phase, the root node index area is used to store the type encoding of all agricultural management events, including fertilization events, irrigation events, tillage events, and land use change events. Each event type encoding serves as the unique entry identifier for the subtree under the root node. Under each root node, the corresponding soil type is attached. Encoding nodes are designated SA for sandy soil, LO for loam, and CL for clay, with each soil type node serving as a second-level index layer for that event type under different soil conditions. Upon entering the leaf nodes, the data area stores key-value pairs using a hash dictionary. The delay period after the key-value event is discretely divided in hours; for example, an index segment is set with a step size of 10 hours between 1 hour and 120 hours. The value is the weight coefficient of the corresponding natural element within that delay period. The weight coefficient is a pre-compiled static floating-point value derived from agricultural experts' knowledge in areas such as infiltration rate response, fertilizer effect delay absorption period, and soil response time window length, established based on field experiments. As a result, when edge computing nodes process actual event data, they first receive a triplet index constructed by the event recognition module. This triplet contains the identified event type code, the plot soil type code, and the delay hours after the event occurred. Then, this triplet is used as a path instruction to match and look up the table layer by layer. First, the corresponding root node event type is located, then the child node layer matching the soil type is entered, and finally, the data area of ​​the leaf node is accessed based on the delay time to extract the matching weight coefficients. The entire matching query operation uses a memory-cached index mechanism, and the node structure is linked using array references. The query response latency is controlled within two hundred milliseconds, meeting the edge node's requirements for low power consumption and high speed. The processing requirements for the response are as follows: After the query is completed, the weight coefficient will be combined with the corresponding natural element values ​​to generate a coupling factor, and semantic tags will be dynamically added according to the event and soil combination type. These semantic tags are generated separately by the labeling rule module outside the rule lookup table structure. For example, when the event type is irrigation and the soil type is sandy, if rainfall is a participating natural element, an infiltration promotion label will be automatically added; if it is clay, it will be labeled as retention risk; and if the wind speed exceeds eight meters per second, a wind erosion risk warning will be added to the loam. This semantic information is written into the final dataset along with the coupling factor to support the semantic enhancement input for subsequent time-series entropy value calculation and event logic judgment process.

[0020] Preferably, in S3, after the coupling factor dataset is generated and stored in the edge node buffer, a rolling calculation of the sliding time variance needs to be performed on the time series data according to a set method. This process defines fixed-length time segments based on the agricultural operation cycle. In dryland planting, the length of each segment is set to 72 hours, and in paddy field areas, it is set to 48 hours. Each geographic grid unit traverses the time series with this segment length as the step size, forming an interleaved rolling window. Within each window segment, the system extracts the coupling factor values ​​in chronological order, calculates their mean, and sums the squared deviations between the mean and the mean. The average value is taken as the mean for that time segment. The sliding variance value within the window is used as an indicator of the temporal activity of the coupling factor. If the variance value exceeds a preset stability threshold of 0.3 within multiple consecutive segments, the region is determined to be in a high-entropy state. This activates the dedicated real-time stream processing channel for the edge node, which immediately initiates a full-scale coupled data processing flow. It calls three types of data sources as input: satellite remote sensing data, ground sensor data, and manually reported data. These are then uniformly integrated into the processing flow to complete the full index fusion calculation, ensuring that high-frequency event responses can promptly capture the structural disturbances they cause to the soil fertility evolution trend. To ensure processing accuracy, in a high-entropy state, the edge node automatically processes all sampled frequencies... The efficiency has been increased to three times the base frequency, with the remote sensing data slicing cycle reduced from 6 hours to 2 hours, the ground sensor sampling interval adjusted from 15 minutes to 5 minutes, and the frequency of manual data inspection increased to once every 48 hours. Conversely, when the sliding variance remains below the dormant threshold of 0.1 for more than 5 time segments, the area is marked as low-entropy. In the low-entropy state, the real-time processing channel is shut down, and a structure summary monitoring mechanism is adopted, replacing real-time coupled processing by calculating the soil state hash fingerprint. The soil state hash fingerprint consists of three core indicators: the mean of soil organic matter content, the standard deviation of soil pH, and the water content over the past period. The slope of change within the week, these three values ​​are combined and structured to form fingerprint data. Each fingerprint segment represents a unique structural identifier of the current soil state in the area. This fingerprint is stored in the edge node and compared with the texture features of subsequent remote sensing images. When the similarity between the texture structure features extracted from the satellite image and the fingerprint information is lower than a set threshold of 80%, a change signal is triggered, and the real-time data acquisition process for that plot is restarted. This mechanism achieves optimized resource allocation through a state monitoring strategy based on structural summaries. While maintaining the predictive response, it effectively reduces the burden of redundant data processing in low-change areas, and improves the overall efficiency of computing resource allocation and dynamic adaptability.

[0021] Preferably, in S4, to further enhance the constraints of soil fertility evolution prediction behavior in terms of temporal continuity and event logic, a directed graph of regional event paths is constructed based on the identified and registered agricultural event sequences. The path structure uses event type codes and their timestamps as input within each geographic grid cell, extracting event records from the event database in chronological order. Each event type constitutes a node in the graph, and nodes are connected in chronological order to form directed edges, ultimately generating an event path topology expressed as a directed acyclic graph. During path construction, each pair of event nodes is assigned a directed edge weight value, which represents agricultural rule knowledge. The database sets minimum allowable time intervals for these two types of events. For example, the minimum allowable interval between fertilizer application and organic fertilizer application is set to 48 hours, and between land tillage and irrigation is set to 72 hours. Planting behavior and land use change are mutually exclusive event types and cannot coexist in adjacent path segments. The directed edge weights are loaded by consulting the agricultural rule knowledge base in real-time during graph construction, embedding the interval limits for each pair of event types from the preset rules into the weight field of the corresponding edge in the path structure. Subsequently, a mutual exclusion constraint detector traverses the constructed graph structure, judging each pair of adjacent event nodes one by one, starting with the current... The actual interval between events is calculated using the timestamps of the two nodes connected by an edge. This interval is then compared to the minimum allowable interval weight value on that edge. If the event interval is less than the weight value, a first interruption signal is triggered, indicating a time logic conflict. Next, it is determined whether the current event path contains a combination of mutually exclusive events defined by the rules. For example, planting the same crop three times consecutively is considered a violation of crop rotation requirements. If three consecutive nodes in the graph have the same code, a crop rotation conflict is triggered, generating a second interruption signal. Another type of mutually exclusive event combination is fertilization or planting immediately after a land use change. This type of behavior is defined as a structurally illegal operation by agricultural rules, also triggering a second interruption signal. Once either type of interruption signal is generated, The prediction engine deployed in the current area is immediately frozen. The area is marked as an empty window through a registration command, and the interruption timestamp and event path sequence are recorded and archived in the log structure for reference in the subsequent recovery mechanism. In the frozen state, the area no longer receives new prediction requests and no longer participates in the main prediction process. It enters the structure observation state, retaining only the functions of event listening and texture structure feature comparison, to prepare the basic judgment conditions for logical migration and model reconstruction. The entire path structure judgment and freezing mechanism is based on graph structure, restricted by rule base, and judged by timestamp interval, realizing real-time control of the logical integrity of agricultural operation sequence and ensuring the rationality of causal reasoning in the prediction process.

[0022] Preferably, after the empty window area is registered and frozen in S5, in order to achieve a complete migration process from the logical interruption state to the recovery of predictive capabilities, it is necessary to perform topological similarity calculation based on the event sequence before the interruption to find the closest neighboring event chain to construct a temporary predictive framework. The topological similarity matching operation uses the dynamic time warping algorithm as its core mechanism. This algorithm receives two input sequences as comparison objects. The first input is the event path segment consisting of the type encoding sequence of continuous events before the interruption of the empty window area and the corresponding timestamp. The second input is the set of historical event path segments of all neighboring plots in the buffer area. The algorithm uses the morphological distance calculation unit to compare each pair of events. The path is aligned along two dimensions: time and event type. During alignment, key peaks and valleys are identified and path nodes are matched sequentially. The minimum regularized path distance is output as a similarity metric. The path with the smallest distance value is selected from all neighborhood event chains, and its complete path structure is loaded into a window region as a temporary prediction framework. Data input within this window region is no longer processed by the local model but redirected to the processing pipeline of this similar neighborhood. The prediction results are uniformly labeled with migration tags to distinguish their origin. Subsequently, edge nodes monitor the event stream generated in this window region in subsequent times. When a preset number of consecutive compliant event sequences are captured, the window region is de-emptied. In the window state, the compliant event sequence must simultaneously meet two conditions: first, the time interval between events is not less than the minimum allowable interval set in the agricultural rule knowledge base; second, the sliding time variance of the corresponding coupling factor is stable between 0.1 and 0.3, and this condition is met continuously for at least three times. After confirming that these conditions are met, the dedicated model reconstruction process is initiated. The operation procedure is as follows: first, the entire historical training dataset of similar neighborhoods on which the temporary prediction framework relies is migrated to the current empty window area as basic training data; then, the local coupling factor sequence collected before the empty window area was frozen is injected into the training process as incremental input. An incremental learning algorithm is used to adjust the parameters without changing the model structure. Fine-tuning enables the model to simultaneously maintain the logical foundation of neighborhood event paths and integrate the actual characteristics of local soil conditions. This fine-tuning operation uses global parameters as initial weights and local incremental data to correct weight shifts. During the adjustment process, the existing rule boundaries remain unchanged, and only the mapping relationship of internal coupling factors and the weights of predicted output responses are optimized to ensure that the model has the ability to accurately respond to local inputs after migration. After training, the reconstructed model is used as the dedicated prediction model for the empty window area and is connected to the main prediction engine. This model will accept local data streams and support subsequent online adaptive training, and participate in the prediction result output process. Thus, the empty window area completes the complete logical migration loop from freezing to prediction reconstruction.

[0023] (III) Beneficial Effects

[0024] This invention provides a method and system for predicting the spatiotemporal evolution of regionalized soil fertility. It has the following beneficial effects:

[0025] 1. This invention dynamically captures key events such as fertilization and irrigation and generates time anchors through a distributed event monitoring module. Combined with a soil infiltration rate model, it constructs an asymmetric time window to accurately pinpoint the sensitive period of soil state changes associated with the events. Based on a triplet index, it uses a real-time query coupled with a rule lookup table to assign differentiated weight coefficients to natural elements, significantly improving the semantic parsing accuracy of meteorological and satellite data. It employs a sliding time variance detection mechanism to intelligently switch between high and low power consumption modes. When data fluctuations exceed a threshold, a real-time stream processing pipeline is activated to ensure accuracy in the high-entropy region. When the data fluctuations fall below a dormancy threshold, a soil state hash fingerprint is generated and data acquisition is interrupted until satellite imagery triggers a change signal, significantly reducing the ineffective energy consumption of edge computing nodes. By predicting the sequence of illegal operations through a directed graph of event paths and freezing the prediction engine, it effectively avoids the risk of model distortion caused by agricultural conflicts.

[0026] 2. This invention scans the neighborhood event chain with the highest topological similarity within the buffer zone of the empty window area and loads it as a temporary prediction framework to ensure that reference output can still be provided during the interruption of regional prediction. After capturing a set number of compliant event sequences, the empty window state is lifted, the historical training dataset of similar neighborhoods is migrated and the local coupling factor sequence is injected to realize cross-domain knowledge transfer. An incremental learning algorithm is used to fine-tune the model with the migrated dataset as the base parameter and the local sequence as the incremental data, which not only inherits the characteristics of the mature neighborhood model, but also integrates the actual environmental parameters before the empty window area is lifted. The reconstructed dedicated prediction model is quickly adapted to the local real-time data stream through online adaptive training and finally seamlessly connected to the main prediction engine to ensure the continuity and self-healing capability of the regional prediction system. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention provides a method and system for predicting the spatiotemporal evolution of regional soil fertility. The method includes deploying a cluster of edge computing nodes within a plot of land on an agricultural service platform before the start of a regional farmland fertilization operation. Each node has a built-in agricultural rule knowledge base and is connected to four types of event listening channels, namely, the agricultural management system application interface log channel, the agricultural machinery Internet of Things operation signal channel, the multispectral satellite image interpretation event channel, and the manually labeled data channel. When farmers submit fertilization plans through the agricultural management platform, the log channel receives data records containing event type codes and timestamp fields, and immediately pushes them to the event recognition engine of the edge node to complete event confirmation and generate anchor time records. At the same time, the edge node accesses local soil data, identifies the soil type of the plot as loam, and collects the humidity change rate over the past 24 hours through the local humidity sensor. The calculated humidity change rate reaches 6%, which meets the set forward offset condition. Therefore, the forward offset of the asymmetric time window is set to 48 hours. The backward offset is set to 96 hours based on the continuous impact of loam type on fertilization events in the agricultural rule knowledge base. Finally, an asymmetric time window with a total duration of 144 hours is constructed with this fertilization event as the center.

[0029] Within the asymmetric time window, edge nodes synchronously collect natural element data from meteorological stations surrounding the plot through a scheduling module, including rainfall, temperature, wind speed, humidity, and evaporation. Combined with 2-hour resolution multispectral imagery acquired from a satellite remote sensing platform, the vegetation index and texture change index of the plot are extracted. All data are aligned with timestamps to construct a time-series dataset. Subsequently, a triplet index is constructed based on the fertilization event type code (set to 1), the soil type code (set to LO), and the currently extracted delay time period. This index is submitted to the coupling rule lookup table stored locally on the edge nodes. The node paths are matched sequentially in the event type layer index area, the soil type layer index area, and the delay time period hash dictionary layer, and the corresponding natural element weight coefficients are returned. For example, the rainfall factor weight is 0.5, and the wind speed factor weight is 0.7. Then, each natural element data and weight is weighted to form a coupling factor, and semantic tags are added to the generated coupling factor. For example, if the wind speed at the current node is 8 m / s, the system identifies its enhanced impact on soil evapotranspiration, so a water loss risk label is added and written into the final coupling factor dataset.

[0030] After generating the coupling factor dataset, the edge node starts the sliding time variance calculation module, which uses a window length of 72 hours to traverse the coupling factor data sequence. It observes that the variance value exceeds 0.3 in three consecutive windows, confirming that the current plot is in a high-entropy state. The system immediately activates the real-time streaming channel, compresses the satellite image sampling frequency from 6 hours to 2 hours, adjusts the ground sensor sampling interval from 15 minutes to 5 minutes, and retrieves supplementary observation data from the manual channel to perform real-time fusion analysis.

[0031] Meanwhile, on the third day of operation in the high-entropy state, the event database records that the same crop has been planted continuously on the plot. When the node is building a new event path topology, it identifies that the same crop has been planted three times in a row according to the agricultural rule knowledge base, triggering a crop rotation warning and forming a second interruption signal. This signal is issued by the mutual exclusion constraint detector and drives the current prediction engine to freeze, register the plot as a void area, and record the last three event path sequences before the interruption for subsequent logical migration.

[0032] After the window area remains frozen for 5 days, the edge node searches for the event chain of adjacent plots within the buffer area and matches a neighborhood event chain with the highest path structure similarity using a dynamic time warping algorithm. The minimum path distance is 0.24. The system loads this neighborhood path as a temporary prediction framework and redirects the real-time data input of the window area to the data processing module corresponding to this neighborhood to continue generating prediction results, but adds migration tags to identify non-local sources.

[0033] Simultaneously, the system continued to monitor event changes within the window area. It observed three consecutive fertilization and irrigation operations, with the intervals between each event exceeding the 48-hour limit set by the agricultural rules. The corresponding coupling factor sliding time variance remained stable between 0.15 and 0.2. Once the release condition was met, a dedicated prediction model reconstruction operation was executed. First, the training data from the matching neighborhood was copied to the window area. The model initialization parameters inherited the original neighborhood path structure. Then, the coupling factor sequence generated before the window area was frozen was injected as incremental data. An incremental learning algorithm was used for fine-tuning, correcting the mapping relationship between the coupling factors and the output values, forming a new model. The model structure remained unchanged, but the parameters were locally adapted and optimized, and then integrated into the main prediction process. From then on, the window area possessed independent prediction capabilities, and the prediction process was fully restored to its pre-freezing state.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the spatiotemporal evolution of regionalized soil fertility, characterized in that, Includes the following steps: S1: Deploy an edge computing node cluster and build an agricultural rule knowledge base; capture agricultural management event streams in real time through a distributed event listening module; when fertilization events, irrigation events, tillage events or land use change events are identified, generate time anchors at the event trigger time; determine forward and backward offsets based on the soil infiltration rate model associated with the event type; and construct an asymmetric time window centered on the anchor. S2: Extract time series data of natural elements from meteorological stations and satellite sensors within the asymmetric time window, combine event type codes, soil type codes and post-event delay periods into triplet indexes, query the pre-compiled coupling rule lookup table to obtain the weight coefficients of natural elements, and generate a coupling factor dataset with semantic tags. S3: Calculate the sliding time variance of the coupling factors in the coupling factor dataset within each geographic grid cell. When the sliding time variance exceeds the stability threshold, it is marked as a high-entropy region and the real-time stream processing pipeline is started. When the variance is lower than the dormancy threshold, a soil state hash fingerprint is generated. Data acquisition is interrupted by fingerprint comparison until the change signal is triggered by the comparison of satellite image texture features. S4: Construct a directed graph of regional event paths to generate event path topology, select the neighborhood event chain with the highest topological similarity and load it as a temporary prediction framework; when it is detected that the interval between adjacent events is less than the minimum allowed interval defined by the agricultural rule knowledge base or the event combination violates the mutual exclusion constraint of the agricultural rule knowledge base, freeze the prediction engine of the region and register it as a blank window area; S5: Scan the event path topology within the buffer area formed by extending a set distance outside the geographical boundary of the empty window area, select the neighborhood event chain with the highest topological similarity and load it as a temporary prediction framework. When a compliant event sequence that has been captured for a set number of consecutive times is captured, the empty window state is lifted, the historical training dataset of similar neighborhoods is migrated to the empty window area, the local coupling factor sequence before the empty window area is lifted is injected, the model parameters are fine-tuned using an incremental learning algorithm, and the exclusive prediction model is reconstructed.

2. The method for predicting the spatiotemporal evolution of regionalized soil fertility according to claim 1, characterized in that: In S1, the distributed event monitoring module processes agricultural management system application interface logs, agricultural machinery IoT operation signals, multispectral satellite image interpretation events, and manually labeled data in parallel. The forward offset of the asymmetric time window is dynamically calculated based on the soil moisture change rate before the event occurs, and the backward offset incorporates the event's persistence impact factor. Irrigation events extend the backward offset according to the soil's saturated hydraulic conductivity. The soil infiltration rate model dynamically configures the window length based on the laboratory saturated hydraulic conductivity parameters of sandy soil, loam soil, and clay soil. The agricultural management system application interface log channel parses event timestamps and event type codes. The agricultural machinery IoT operation signal channel generates standardized event messages through a protocol converter. The multispectral satellite image channel performs texture segmentation and change detection to output image interpretation events. Manually labeled data is accessed through the terminal form.

3. The method for predicting the spatiotemporal evolution of regionalized soil fertility according to claim 1, characterized in that: The coupling rule lookup table described in S2 is stored in a hierarchical tree structure, with the root node index area storing the event type code; The child node index area carries the soil type code; the leaf node data area stores a pre-compiled hash dictionary with delay time period as the key and weight coefficient as the value. The query operation is executed in real time on the edge computing node. After receiving the triple index, the node path is matched layer by layer and the weight coefficient is returned. The semantic tag is dynamically attached to the coupling factor dataset. Rainfall data is marked with infiltration promotion markers in sandy soil and retention risk markers in clay soil.

4. The method for predicting the spatiotemporal evolution of regionalized soil fertility according to claim 1, characterized in that: In S3, the sliding time variance calculation is performed using fixed-length time segments. Fixed-length time segments are set based on the agricultural operation cycle, and the dataset is traversed sequentially by segment to output variance values. When the real-time streaming pipeline is activated, it calls satellite remote sensing data, ground sensor data, and manually reported data to perform full-volume coupled calculations. The soil state hash fingerprint generates a structured summary using the mean soil organic matter content, the standard deviation of soil pH, and the slope of soil moisture content. The change signal is triggered by the comparison results of satellite image texture features. The high-entropy region exclusively uses the real-time processing channel of the edge computing node, increasing the data sampling frequency to a set multiple of the basic sampling frequency. The fingerprint comparison under the dormant threshold uses an incremental update mechanism; only when the crop growth indicators interpreted from the satellite image continuously deviate from the threshold of the cached soil state hash fingerprint is the meteorological station data acquisition module in the empty window area locally activated.

5. The method for predicting the spatiotemporal evolution of regionalized soil fertility according to claim 1, characterized in that: In S4, the directed graph of regional event paths is stored in a directed acyclic graph structure. Event type encoding sequences are extracted from the event database in chronological order to generate the event path topology. The minimum allowable interval between events is loaded as the weight value of the directed edge according to the agricultural rule knowledge base. The mutual exclusion constraint detector traverses the directed graph structure. When it detects that the time interval between adjacent nodes is less than the corresponding weight value of the directed edge, it triggers a first interrupt signal, or when there is a connecting edge of a combination of mutually exclusive events defined in the agricultural rule knowledge base, it triggers a second interrupt signal. Both the first and second interrupt signals drive the regional prediction engine to freeze and generate a registration instruction for the empty window area. The mutual exclusion constraint rule is that continuous land use change events and planting events cannot coexist. Specifically, it includes: irrigation events are prohibited within 72 hours after a land tillage event; the interval between fertilizer application events and organic fertilizer application events is less than 48 hours, which is considered a conflict; and crop rotation warnings are triggered when the same crop is planted continuously on the same plot more than 3 times.

6. The method for predicting the spatiotemporal evolution of regionalized soil fertility according to claim 1, characterized in that: In S5, the topological similarity calculation adopts a dynamic time warping algorithm. The timestamps and type codes of the event sequences before the interruption of the window area are read as the first input; the corresponding data of the neighborhood event chain is obtained as the second input; the minimum warping path distance is output by aligning the feature peaks and valleys through the morphological distance calculation unit; the neighborhood event chain with the lowest minimum warping path distance value is loaded as a temporary prediction framework; when the temporary prediction framework is running, the data input of the window area is redirected to the data processing pipeline of the similar neighborhood, and the output results are marked with data source tags; the compliant event sequences must simultaneously meet the event interval threshold and coupling factor fluctuation range constraints; the reconstruction process of the dedicated prediction model includes: migrating the historical training dataset of the similar neighborhood to the window area, injecting the local coupling factor sequence before the window area is released, and using an incremental learning algorithm to fine-tune the model parameters with the migrated dataset as the base parameters and the local sequence as the incremental data; the reconstructed dedicated prediction model inherits the prediction model parameters of the original similar neighborhood region and starts online adaptive training based on local real-time data. The trained model is then connected to the main prediction engine.

7. The system of the regionalized soil fertility spatiotemporal evolution prediction method according to claim 1, characterized in that: Deploy an edge computing node cluster in the target agricultural area and incorporate an agricultural rule knowledge base; The distributed event monitoring module connects to the edge computing node cluster and is configured to capture agricultural management event streams in real time. When fertilization, irrigation, tillage, or land use change events are identified, a time anchor point is generated at the event trigger time. The forward and backward offsets are determined based on the soil permeability rate model associated with the event type, and an asymmetric time window centered on the anchor point is constructed. The coupling factor generation module is configured to extract time-series data of natural elements from meteorological stations and satellite sensors within the asymmetric time window, combine event type codes, soil type codes, and post-event delay periods into triplet indexes, query a pre-compiled coupling rule lookup table to obtain the weight coefficients of natural elements, and output a coupling factor dataset with semantic tags. The dynamic monitoring and processing module includes a variance calculation unit, a high-entropy processing channel, and a fingerprint management unit. The variance calculation unit calculates the sliding time variance of the coupling factor dataset within each geographic grid unit. The high-entropy processing channel activates the real-time streaming pipeline when the variance exceeds a stability threshold. The fingerprint management unit generates a soil state hash fingerprint when the variance is below a dormancy threshold and interrupts data acquisition through fingerprint comparison until a change signal is triggered by satellite image texture feature comparison. The event topology analysis engine constructs a directed graph of regional event paths to generate event path topology. When it detects that the interval between adjacent events is less than the minimum allowed interval defined by the agricultural rule knowledge base or that the event combination violates the mutual exclusion constraint, the regional prediction engine is frozen and registered as an empty window area. The neighboring event path topology is scanned within the empty window area buffer region, and the neighboring event chain with the highest topological similarity is selected and loaded as a temporary prediction framework. The window recovery module is configured to remove the window state after capturing a set number of compliant event sequences consecutively; migrate historical training datasets from similar neighborhoods to the window area; inject the local coupling factor sequence before the window is removed; fine-tune the model parameters using an incremental learning algorithm; reconstruct the dedicated prediction model and connect it to the main prediction engine.