Prediction model training method and system based on land measurement data

Abstract

The embodiment of the invention provides a prediction model training method and system based on land measurement data, and the method comprises the steps: firstly obtaining a land measurement data set containing historical land attribute distribution data and real-time land state recording data, and then carrying out the space grid division of the historical land attribute distribution data, performing timestamp alignment processing on real-time land state record data, performing cross-modal feature fusion to generate a land space feature map, and extracting abnormal fluctuation nodes in the land space feature map to determine a target training sample set; and finally, performing multi-stage iterative training on the initial prediction model based on the target training sample set to obtain a land state prediction model, thereby effectively improving the accuracy of land state prediction and providing a scientific basis for land governance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for training a prediction model based on land measurement data. Background Art

[0002] In the field of land management and planning, accurately predicting changes in land status is of great significance. With the intensification of environmental changes and human activities, the soil properties, vegetation cover, and other statuses of the land face many uncertainties. For example, unreasonable agricultural activities and the acceleration of urbanization may lead to soil degradation, affect vegetation growth, and thus change the balance of the entire ecosystem.

[0003] Traditional land assessment methods often rely on limited sample data and simple statistical analysis, making it difficult to accurately and dynamically predict land status. Modern measurement technologies can obtain a large amount of land measurement data, including historical land attribute distribution data and real-time land status record data. These data provide the possibility for constructing a more accurate land status prediction model. However, how to effectively utilize these multi-source and multi-modal data has become a key issue. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for training a prediction model based on land measurement data.

[0005] Combined with the first aspect of this application, a method for training a prediction model based on land measurement data is provided, which is applied to a prediction model training system based on land measurement data. The method includes:

[0006] Obtain a land measurement data set of a target area. The land measurement data set includes historical land attribute distribution data and real-time land status record data. The historical land attribute distribution data includes soil composition information and terrain feature parameters of multiple geographical blocks, and the real-time land status record data includes the surface temperature distribution and humidity change trajectory collected by current environmental monitoring devices;

[0007] Perform spatial grid division on the historical land attribute distribution data to generate attribute coding units with hierarchical association relationships, and perform timestamp alignment processing on the real-time land status record data to obtain a dynamic state sequence corresponding to the attribute coding units;

[0008] Perform cross-modal feature fusion on the attribute coding units and the dynamic state sequences to generate a land spatial feature map, where each node in the land spatial feature map represents the attribute evolution trend of the corresponding geographical block within a preset time range;

[0009] Extract the abnormal fluctuation nodes in the land spatial feature map, and determine the target training sample set according to the soil component information and surface temperature distribution corresponding to the abnormal fluctuation nodes;

[0010] Perform multi-stage iterative training on the initial prediction model based on the target training sample set to obtain a land state prediction model, where the land state prediction model is used to output the soil degradation risk assessment result and the vegetation coverage change probability within a future time period according to the input geographical block identifier.

[0011] In a possible implementation manner of the first aspect, the spatial grid division of the historical land attribute distribution data to generate attribute coding units with hierarchical association relationships includes:

[0012] Divide the target area into multiple geographical grid units with equal areas according to the longitude and latitude coordinate ranges in the historical land attribute distribution data;

[0013] For each geographical grid unit, extract the soil pH distribution, organic matter content gradient, and mineral concentration change curve contained therein, and generate a soil attribute vector corresponding to the geographical grid unit;

[0014] Based on the difference degree of soil attribute vectors between adjacent geographical grid units, construct a multi-level spatial association network, where the high-level network nodes represent the large-scale soil attribute means, and the low-level network nodes represent the local soil attribute fluctuation characteristics;

[0015] Map and associate the network nodes at each level in the spatial association network with the corresponding geographical grid units to generate the attribute coding units.

[0016] In a possible implementation manner of the first aspect, the cross-modal feature fusion of the attribute coding units and the dynamic state sequence to generate a land spatial feature map includes:

[0017] Perform frequency domain transformation on the surface temperature distribution data in the dynamic state sequence to extract the temperature fluctuation patterns at different time scales;

[0018] Align the temperature fluctuation patterns with the soil attribute vectors of the corresponding geographical grid units in the time domain to generate a hybrid feature vector integrating soil-temperature correlation features;

[0019] Determine the feature diffusion weights between nodes according to the transfer path of the hybrid feature vector in the multi-level spatial association network;

[0020] Perform spatial interpolation calculation on the mixed feature vectors based on the feature diffusion weights to generate a continuous feature distribution map covering the entire target area, and convert the continuous feature distribution map into graph structure data to form the land spatial feature map.

[0021] In a possible implementation manner of the first aspect, extracting the abnormal fluctuation nodes in the land spatial feature map includes:

[0022] Calculate the feature change rate of each node in the land spatial feature map, and construct an anomaly detection threshold based on the difference in feature change rates between adjacent nodes;

[0023] Identify the nodes whose feature change rate exceeds the anomaly detection threshold, and trace their soil component mutation records in the historical land attribute distribution data;

[0024] Mark the nodes that simultaneously meet the feature rate anomaly and soil component mutation as the abnormal fluctuation nodes, and record their corresponding longitude and latitude coordinates and timestamp information.

[0025] In a possible implementation manner of the first aspect, the multi-stage iterative training of the initial prediction model based on the target training sample set includes:

[0026] Divide the target training sample set into a training subset and a validation subset, where the training subset contains geographical block data with clear soil degradation labels, and the validation subset contains geographical block data without labeled soil state changes;

[0027] In the first training stage, use the training subset to initialize the parameters of the initial prediction model, and generate initial prediction weights based on the correlation between soil component information and surface temperature distribution;

[0028] In the second training stage, input the validation subset into the initialized prediction model, perform error backpropagation according to the output vegetation cover change probability and real-time satellite image data, and adjust the dynamic parameters in the initial prediction weights;

[0029] In the third training stage, deploy the adjusted prediction model to the edge computing device, perform online calibration on the prediction results through the real-time received environmental sensor data, and update the final parameters of the land state prediction model.

[0030] In a possible implementation manner of the first aspect, the online calibration of the prediction results through the real-time received environmental sensor data includes:

[0031] Obtain the soil moisture sensor readings and light intensity monitoring values of the target geographical block in the current time period;

[0032] Compare the soil moisture sensor readings with the probability of vegetation cover change output by the land status prediction model to determine the humidity deviation coefficient;

[0033] Weight and correct the humidity deviation coefficient according to the monitored light intensity value to generate an environmental calibration factor;

[0034] Superimpose the environmental calibration factor on the output layer of the land status prediction model to dynamically adjust the soil degradation risk assessment results in the future time period.

[0035] In a possible implementation manner of the first aspect, the method further includes:

[0036] When it is detected that the soil degradation risk assessment result of the target geographical block exceeds the preset threshold, generate a recommended list of land treatment plans;

[0037] Extract the resource consumption data and expected effect data required for each plan in the recommended list of land treatment plans to construct a multi-objective optimization decision-making model;

[0038] According to the current available resource inventory data and environmental carrying capacity constraint conditions, screen the optimal land treatment plan from the multi-objective optimization decision-making model and output it to the user terminal.

[0039] In a possible implementation manner of the first aspect, the generation of the recommended list of land treatment plans includes:

[0040] Query the successful case data in the historical land treatment case library that matches the current soil degradation type;

[0041] Analyze the treatment technical path, material ratio parameters and implementation cycle information adopted in the successful case data;

[0042] Combine the terrain slope data and water source distribution data of the target geographical block to adaptively transform the treatment technical path to generate multiple candidate treatment plans with priority rankings;

[0043] Evaluate the matching degree of the candidate treatment plans with the real-time weather forecast data, and eliminate the plans that do not meet the climate conditions to form the recommended list of land treatment plans.

[0044] Combined with the second aspect of the present application, a prediction model training system based on land measurement data is provided. The prediction model training system based on land measurement data includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the prediction model training system based on land measurement data implements the foregoing prediction model training method based on land measurement data.

[0045] In combination with the third aspect of the present application, a computer-readable storage medium is provided. Computer-executable instructions are stored in the computer-readable storage medium. When the computer-executable instructions are executed, the foregoing method for training a prediction model based on land measurement data is implemented.

[0046] In combination with any of the above aspects, first, a land measurement data set including historical land attribute distribution data and real-time land status record data is obtained. Then, the historical land attribute distribution data is divided into spatial grids, and the real-time land status record data is subjected to timestamp alignment processing. Next, cross-modal feature fusion is performed to generate a land spatial feature map. Abnormal fluctuation nodes in the land spatial feature map are extracted to determine a target training sample set. Finally, the initial prediction model is subjected to multi-stage iterative training based on the target training sample set to obtain a land status prediction model, thereby effectively improving the accuracy of land status prediction and providing a scientific basis for land governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained in combination with these drawings without creative efforts.

[0048] Figure 1 The flowchart of the method for training a prediction model based on land measurement data provided by the embodiments of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0050] The terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or terminal including a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or terminals.

[0051] References to "embodiments" in this specification mean that specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the invention. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0052] Figure 1 A flowchart showing the method for training a prediction model based on land measurement data provided by an embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps of the method for training a prediction model based on land measurement data in this embodiment can be shared with each other based on actual needs, or some of the steps can also be omitted or maintained. The details of the method for training a prediction model based on land measurement data include:

[0053] Step S110: Obtain a land measurement data set of a target area, where the land measurement data set includes historical land attribute distribution data and real-time land status record data. The historical land attribute distribution data includes soil composition information and topographic feature parameters of multiple geographical blocks, and the real-time land status record data includes the surface temperature distribution and humidity change trajectory collected by current environmental monitoring devices.

[0054] In an actual scenario, assume that the target area is a large agricultural land covering multiple different geographical blocks. For the historical land attribute distribution data, the soil composition information includes the content ratio of various minerals in the soil, acidity and alkalinity (pH value), organic matter content, etc. For example, in some geographical blocks, the contents of nitrogen, phosphorus, and potassium in the soil are 15 mg / kg, 10 mg / kg, and 20 mg / kg respectively, the pH value is 6.5, and the organic matter content is 3%. The topographic feature parameters include altitude, slope, aspect, etc. For example, in some areas, the altitude is 200 meters, the slope is 5 degrees, and the aspect is southeast. In terms of the real-time land status record data, current environmental monitoring devices are widely distributed in this agricultural land. The surface temperature distribution data can be collected by multiple temperature sensors, and these sensors record the temperature at regular intervals (such as every hour) to form a temperature distribution data at different geographical locations. For example, at 10 am, the temperature in the area near the farm irrigation water source is 22 °C, while the temperature in the area near the woods at the edge of the farmland is 20 °C. The humidity change trajectory is recorded by humidity sensors and can reflect the humidity change situation within a day or a period of time. For example, the humidity decreases from 80% in the morning to 40% in the afternoon as the sun rises, and then rises again as night approaches.

[0055] Step S120: Perform spatial grid division on the historical land attribute distribution data to generate attribute coding units with hierarchical association relationships, and perform timestamp alignment processing on the real-time land status record data to obtain a dynamic status sequence corresponding to the attribute coding units.

[0056] First, according to the longitude and latitude coordinate range in the historical land attribute distribution data, divide the target area into multiple geographical grid units with equal areas. Taking the previously mentioned agricultural land as an example, the longitude and latitude range of this agricultural land is determined. Suppose the longitude range is from 110°E to 115°E, and the latitude range is from 30°N to 35°N. According to certain division rules, divide it into square geographical grid units with a side length of 1 kilometer. In this way, the entire agricultural land is divided into many small grids with equal areas.

[0057] For each of the geographical grid units, extract the soil pH distribution, organic matter content gradient, and mineral concentration change curve contained therein to generate a soil attribute vector corresponding to the geographical grid unit. In a specific geographical grid unit, through detailed soil testing, it is found that the soil pH shows a distribution gradually decreasing from the northeast direction to the southwest direction within the grid, with the pH value gradually decreasing from 7.0 to 6.0; the organic matter content gradually decreases from 5% near the farmland entrance to 3% towards the center of the grid; the mineral concentration change curve shows that the concentration of potassium element gradually decreases from 25 mg / kg in the northwest corner of the grid to 15 mg / kg in the southeast corner. Based on these data, a soil attribute vector corresponding to this geographical grid unit can be generated.

[0058] Based on the difference degree of soil attribute vectors between adjacent geographical grid units, construct a multi-level spatial association network, where the high-level network nodes represent the large-scale soil attribute means, and the low-level network nodes represent the local soil attribute fluctuation characteristics. For example, there are certain differences in soil attribute vectors between adjacent geographical grid units. If we take a group of 10 adjacent geographical grid units, the overall soil pH mean, organic matter content mean, and mineral concentration mean, etc., can be used as the large-scale soil attribute means represented by the high-level network nodes. And the fluctuation conditions such as the deviation of each individual geographical grid unit relative to this group of means are represented by the low-level network nodes.

[0059] Map and associate the network nodes at each level in the spatial association network with the corresponding geographical grid units to generate the attribute coding units. In this way, each geographical grid unit has a corresponding attribute coding unit, and these attribute coding units not only contain their own soil attribute information but also establish connections with surrounding geographical grid units through the spatial association network.

[0060] Meanwhile, timestamp alignment processing is performed on the real-time land status record data. In the environmental monitoring of this agricultural land, there may be certain deviations in the data collection times of different sensors. For example, some temperature sensors may have inconsistent collection times with other sensors due to equipment failures or transmission delays. Through timestamp alignment processing, all data such as the surface temperature distribution and humidity change trajectories corresponding to the geographical grid cells are unified in time, so as to obtain the dynamic state sequence corresponding to the attribute coding unit.

[0061] Step S130: Perform cross-modal feature fusion on the attribute coding unit and the dynamic state sequence to generate a land spatial feature map, where each node in the land spatial feature map represents the attribute evolution trend of the corresponding geographical block within a preset time range.

[0062] Perform frequency domain transformation on the surface temperature distribution data in the dynamic state sequence to extract the temperature fluctuation patterns at different time scales. In this agricultural land, the surface temperature distribution data has different fluctuation patterns within a day. Through frequency domain transformation, the temperature data in the time domain can be converted to the frequency domain, so as to more clearly see the temperature fluctuation conditions under different frequency components. For example, when the solar radiation is strong during the day, the temperature fluctuations may mainly be at a shorter period (such as hourly fluctuations), while at night, the period of temperature fluctuations may become longer (such as fluctuations every 3 - 4 hours).

[0063] Align the temperature fluctuation pattern with the soil attribute vector of the corresponding geographical grid cell in the time domain to generate a hybrid feature vector integrating the soil-temperature correlation features. In each geographical grid cell, align the previously obtained soil attribute vector and the extracted temperature fluctuation pattern in chronological order. For example, when the soil acidity in the soil attribute vector remains stable during a certain period, what is the situation of the corresponding temperature fluctuation pattern during this period, and in this way, the two are integrated to generate a hybrid feature vector.

[0064] Determine the feature diffusion weights between each node according to the transfer path of the hybrid feature vector in the multi-level spatial correlation network. Since the transfer paths of the hybrid feature vector in the spatial correlation network are different, the influence of different nodes on the feature diffusion is also different. For example, in the high-level network nodes, a certain node has a larger feature diffusion weight to other nodes because it is connected to more low-level network nodes and the hybrid feature vectors of these nodes are similar; while in the low-level network nodes, if the hybrid feature vector of a certain node is quite different from the surrounding nodes, then its feature diffusion weight to other nodes will be smaller.

[0065] Perform spatial interpolation calculation on the mixed feature vectors based on the feature diffusion weights to generate a continuous feature distribution map covering the entire target area, and convert the continuous feature distribution map into graph structure data to form the land spatial feature map. Through spatial interpolation calculation, using the mixed feature vectors of each geographical grid cell and the feature diffusion weights between them, the features of discrete geographical grid cells are extended to the entire target area to form a continuous feature distribution map. Then, the continuous feature distribution map is converted into graph structure data, where the nodes of the graph correspond to geographical grid cells and the edges represent the relationships between nodes, thus obtaining the land spatial feature map. Each node in the map can characterize the evolution trend of attributes such as soil and temperature in the corresponding geographical block within a preset time range.

[0066] Step S140: Extract the abnormally fluctuating nodes from the land spatial feature map, and determine the target training sample set according to the soil component information and surface temperature distribution corresponding to the abnormally fluctuating nodes.

[0067] Calculate the feature change rate of each node in the land spatial feature map, and construct an anomaly detection threshold based on the difference in feature change rates between adjacent nodes. In the land spatial feature map, each node has its corresponding feature change rate. Taking the soil pH feature as an example, within a certain period of time, the soil pH of the geographical block represented by a certain node changes from 6.5 to 6.0, and the rate of this change is calculated. Then, observe the soil pH change rates of adjacent nodes. If the change rates of most adjacent nodes are within a small range, such as between -0.01 and 0.01, then an anomaly detection threshold can be constructed based on this range, for example, setting the anomaly detection threshold as the change rate being greater than 0.05 or less than -0.05.

[0068] Identify the nodes whose feature change rates exceed the anomaly detection threshold, and trace the soil component mutation records in the historical land attribute distribution data. When it is found that the feature change rate of a certain node exceeds the anomaly detection threshold, for example, the soil pH change rate of a certain node is 0.1, search for the corresponding geographical block in the historical land attribute distribution data. It may be found that there has been a large-scale application of chemical fertilizers in the past in this geographical block, resulting in mutations in some components of the soil, such as a sudden 50% increase in the nitrogen element content in the soil.

[0069] Nodes that simultaneously meet the criteria of abnormal characteristic rate and sudden change in soil composition are marked as the abnormal fluctuation nodes, and their corresponding longitude and latitude coordinates and timestamp information are recorded. If a node meets both the abnormal rate of change in soil pH and has a record of sudden change in soil composition, then this node is marked as an abnormal fluctuation node. Meanwhile, record the longitude and latitude coordinates of the geographical block corresponding to this node, such as 112° east longitude and 32° north latitude, as well as the timestamp when this abnormal fluctuation occurred, such as 10:00 am on May 10, 2023. Based on the soil composition information and surface temperature distribution data corresponding to these abnormal fluctuation nodes, the target training sample set can be determined, and this sample set will be used for subsequent prediction model training.

[0070] Step S150: Perform multi-stage iterative training on the initial prediction model based on the target training sample set to obtain a land status prediction model, where the land status prediction model is used to output the soil degradation risk assessment result and the vegetation coverage change probability within a future time period according to the input geographical block identifier.

[0071] Divide the target training sample set into a training subset and a validation subset, where the training subset contains geographical block data with clear soil degradation labels, and the validation subset contains geographical block data without labeled soil status changes. Among the previously determined target training sample set, assume that a part of the geographical blocks have undergone detailed soil tests and long-term observations, and it is clearly known that these geographical blocks have soil degradation, and the degree of soil degradation has been quantitatively labeled, such as being divided into labels of mild, moderate, and severe degradation according to the degree of soil fertility decline. The data of these geographical blocks constitute the training subset. While the data of another part of the geographical blocks do not have clear labels of soil status changes, but still contain relevant data such as soil composition information and surface temperature distribution. The data of these geographical blocks form the validation subset.

[0072] In the first training stage, use the training subset to initialize the parameters of the initial prediction model, and generate initial prediction weights based on the correlation between soil composition information and surface temperature distribution. The initial prediction model can be a neural network model, such as a multi-layer perceptron model. Input the data of the training subset into this initial prediction model, and initialize the parameters of the initial prediction model according to the relationship between soil composition information (such as the content of various minerals, organic matter content, etc.) and surface temperature distribution. For example, when the organic matter content in the soil is relatively high, the surface temperature is relatively stable, and this relationship can be reflected in the weight setting during model initialization. If the correlation between the organic matter content and the stability degree of the surface temperature is very strong, then a relatively large weight value will be given in the weight setting.

[0073] In the second training stage, input the validation subset into the initialized prediction model. According to the error backpropagation between the output vegetation coverage change probability and the real-time satellite image data, adjust the dynamic parameters in the initial prediction weights. Input the data of the validation subset into the already initialized prediction model, and the prediction model will output the vegetation coverage change probability of each geographical block. At the same time, obtain the actual vegetation coverage situation through satellite image data. For example, the satellite image shows that the vegetation coverage area of a certain geographical block has decreased from 80% to 70% within a month, while the vegetation coverage area corresponding to the vegetation coverage change probability predicted by the model has decreased to 75%, thus generating an error. According to this error, through the error backpropagation algorithm, adjust the dynamic parameters in the initial prediction weights in the model to make the prediction result of the model closer to the actual situation.

[0074] In the third training stage, deploy the adjusted prediction model to the edge computing device, perform online calibration on the prediction result through the real-time received environmental sensor data, and update the final parameters of the land status prediction model. Deploy the prediction model adjusted through the first two stages of training to the edge computing device located near this agricultural land. This edge computing device can receive data from environmental sensors in real time, such as data from soil moisture sensors and light intensity monitoring devices. For example, the current soil moisture sensor reading is 30%, and the vegetation coverage change probability predicted by the model is obtained based on previous training data and model parameters, but this humidity value may affect the vegetation coverage. Compare the soil moisture sensor reading with the vegetation coverage change probability output by the model to determine the humidity deviation coefficient. If the assumed soil moisture corresponding to the vegetation coverage change probability predicted by the model is 40%, then the humidity deviation coefficient can be calculated based on the difference between the two. Then, perform weighted correction on the humidity deviation coefficient according to the light intensity monitoring value to generate an environmental calibration factor. For example, when the light intensity is strong, the influence of humidity on vegetation coverage may be amplified or reduced, and the humidity deviation coefficient is weighted and corrected according to this relationship. Finally, superimpose the environmental calibration factor on the output layer of the land status prediction model to dynamically adjust the soil degradation risk assessment result in the future time period.

[0075] When it is detected that the soil degradation risk assessment result of the target geographical block exceeds the preset threshold, generate a recommended list of land treatment plans. Assume that the preset soil degradation risk threshold is 50%. When the soil degradation risk assessment result of a certain target geographical block reaches 60%, it is necessary to generate a recommended list of land treatment plans.

[0076] First, query the successful case data in the historical land treatment case library that matches the current soil degradation type. If the current soil degradation type is caused by a decline in soil fertility, search for all cases in the historical case library that have been successfully treated due to a decline in soil fertility. These cases contain information such as the treatment technology paths, material ratio parameters, and implementation cycle information used at that time. For example, there is a successful case where soil fertility was improved by applying organic fertilizers. The material ratio of the organic fertilizers is a mixture of nitrogen, phosphorus, and potassium in a ratio of 3:2:1, and the implementation cycle is 6 months.

[0077] Analyze the treatment technology paths, material ratio parameters, and implementation cycle information used in the successful case data. For the above case of organic fertilizers, analyze its technical path in detail, that is, how the organic fertilizers are evenly applied to the soil, whether it is through mechanical sowing or manual sowing; how the material ratio parameter 3:2:1 is determined and whether it is related to the local soil type and crop requirements; the changes in soil fertility at different stages within the 6-month implementation cycle, etc.

[0078] Combine the terrain slope data and water source distribution data of the target geographical block to adaptively transform the treatment technology path and generate multiple candidate treatment plans with priority rankings. If the terrain slope of the target geographical block is relatively large, then when applying organic fertilizers, it may be necessary to consider preventing fertilizer loss, and the original technology path can be transformed by using stratified fertilization or fertilization along the contour line. At the same time, if the water source is far away, then irrigation needs to be considered after fertilizer application, and some water-saving irrigation measures may need to be added. Generate multiple candidate treatment plans based on these adaptive transformations and rank them according to factors such as feasibility and cost.

[0079] Evaluate the matching degree of the candidate treatment plans with the real-time weather forecast data and eliminate the plans that do not meet the climate conditions to form the recommended list of land treatment plans. If there will be a large amount of rainfall in the future weather forecast for a period of time, then some plans that require a dry environment for fertilization operations are not suitable. After eliminating these plans, the remaining plans form the recommended list of land treatment plans.

[0080] Extract the resource consumption data and expected effect data required for each plan in the recommended list of land treatment plans to construct a multi-objective optimization decision model. For each candidate treatment plan, determine the resources it needs to consume, such as labor, materials (fertilizers, irrigation equipment, etc.), and finances (costs for purchasing fertilizers and equipment, etc.), as well as the expected effects that can be achieved, such as the degree of soil fertility improvement and the proportion of increased vegetation cover. Construct a multi-objective optimization decision model based on these data, for example, taking the minimization of resource consumption and the maximization of expected effects as the objective functions.

[0081] Based on the current available resource inventory data and the constraints of the environmental carrying capacity, the optimal land treatment plan is screened from the multi-objective optimization decision model and output to the user terminal. Assume that the current available inventory of organic fertilizers is limited, and the local environmental carrying capacity is limited and cannot bear excessive fertilizer application to avoid environmental pollution. According to these constraints, the optimal land treatment plan is screened in the multi-objective optimization decision model, and then the land treatment plan is output to the user terminal responsible for the management of this agricultural land for them to perform land treatment operations.

[0082] In the above embodiments, the prediction model training system based on land measurement data for implementing the above method embodiments includes at least one processor, a control module (chipset) coupled to at least one of the (at least one) processors, a memory coupled to the control module, a non-volatile memory (NVM) / storage device coupled to the control module, at least one loading to / output device coupled to the control module, and a network interface coupled to the control module.

[0083] The processor may include at least one single-core or multi-core processor, and the processor may include any combination of general-purpose processors or dedicated processors (such as graphics processors, application processors, baseband processors, etc.). For some alternative embodiments, the prediction model training system based on land measurement data can be an electronic device such as the gateway described in the embodiments of the present application.

[0084] For some alternative embodiments, the prediction model training system based on land measurement data may include at least one computer-readable medium having instructions (e.g., a memory or an NVM / storage device) and at least one processor integrated with the at least one computer-readable medium and configured to execute the instructions to implement the modules to perform the actions described in the present disclosure.

[0085] For one embodiment, the control module may include any suitable interface controller to provide any suitable interface to at least one of the (at least one) processors and / or any suitable device or component communicating with the control module.

[0086] The control module may include a memory controller module to provide an interface to the memory. The memory controller module may be a hardware module, a software module, and / or a firmware module.

[0087] The memory may be used, for example, to load and store data and / or instructions for the prediction model training system based on land measurement data. For one embodiment, the memory may include any suitable volatile memory, such as a suitable DRAM.

[0088] For one embodiment, the control module may include at least one load-to / output controller to provide an interface to the NVM / storage device and the (at least one) load-to / output device.

[0089] For example, the NVM / storage device may be used to store data and / or instructions. The NVM / storage device may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (at least one) non-volatile storage device (e.g., at least one hard disk drive (HDD), at least one compact disc (CD) drive, and / or at least one digital versatile disc (DVD) drive).

[0090] The NVM / storage device may include storage resources that are physically part of the device on which the prediction model training system based on land survey data is installed, or it may be accessible to the device without being part of the device. For example, the NVM / storage device may be accessed via the (at least one) load-to / output device based on a network.

[0091] The (at least one) load-to / output device may provide an interface for the prediction model training system based on land survey data to communicate with any other suitable device. The load-to / output device may include communication components, spelling components, sensor components, etc. The network interface may provide an interface for the prediction model training system based on land survey data to communicate based on at least one network. The prediction model training system based on land survey data may wirelessly communicate with at least one component of the wireless network based on any prior and / or protocol in at least one wireless network prior and / or protocol, e.g., accessing a wireless network based on a communication prior.

[0092] For one embodiment, at least one of the (at least one) processors may be logically loaded with at least one controller of the control module (e.g., the memory controller module). For one embodiment, at least one of the (at least one) processors may be logically loaded with at least one controller of the control module to form a system-level load. For one embodiment, at least one of the (at least one) processors may be logically integrated with at least one controller of the control module on the same die. For one embodiment, at least one of the (at least one) processors may be logically integrated with at least one controller of the control module on the same die to form a system-on-chip (SoC).

[0093] The above has introduced the embodiments of the present application in detail. Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, based on the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

[0094] An embodiment of the present invention discloses a computer-readable storage medium that stores a computer program for electronic data exchange. Among them, the computer program causes the computer to execute the steps in the prediction model training method based on land measurement data described in the foregoing embodiments.

[0095] An embodiment of the present invention discloses a computer program product. The computer program product includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause the computer to execute the steps in the prediction model training method based on land measurement data described in the foregoing embodiments.

[0096] The device embodiments described above are only illustrative. The modules described as separate components may or may not be physically separated. The components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed to multiple modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0097] Through the above specific descriptions of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solutions, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, and the storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disk memories, tape memories, or any other medium that can be used for a computer to read or store data.

[0098] Finally, it should be noted that: the above-disclosed are only the preferred embodiments of the present invention, which are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a prediction model based on land survey data, characterized in that The method includes: Obtain a land measurement data set of a target area, where the land measurement data set includes historical land attribute distribution data and real-time land status record data. The historical land attribute distribution data contains soil composition information and topographic feature parameters of multiple geographical blocks, and the real-time land status record data contains the surface temperature distribution and humidity change trajectory collected by current environmental monitoring devices; Perform spatial grid division on the historical land attribute distribution data to generate attribute coding units with hierarchical association relationships, and perform timestamp alignment processing on the real-time land status record data to obtain a dynamic state sequence corresponding to the attribute coding units; Perform cross-modal feature fusion on the attribute coding units and the dynamic state sequence to generate a land spatial feature map, where each node in the land spatial feature map represents the attribute evolution trend of the corresponding geographical block within a preset time range; Extract the abnormally fluctuating nodes in the land spatial feature map, and determine a target training sample set according to the soil composition information and surface temperature distribution corresponding to the abnormally fluctuating nodes; Perform multi-stage iterative training on an initial prediction model based on the target training sample set to obtain a land status prediction model, where the land status prediction model is used to output the soil degradation risk assessment result and the vegetation coverage change probability within a future time period according to the input geographical block identifier.

2. The prediction model training method based on land survey data according to claim 1, wherein The performing spatial grid division on the historical land attribute distribution data to generate attribute coding units with hierarchical association relationships includes: Divide the target area into multiple geographical grid units with equal areas according to the longitude and latitude coordinate ranges in the historical land attribute distribution data; For each geographical grid unit, extract the soil pH distribution, organic matter content gradient, and mineral concentration change curve contained therein to generate a soil attribute vector corresponding to the geographical grid unit; Based on the difference degree of soil attribute vectors between adjacent geographical grid units, construct a multi-level spatial association network, where the high-level network nodes represent the large-scale soil attribute mean values, and the low-level network nodes represent the local soil attribute fluctuation characteristics; Map and associate the network nodes at each level in the spatial association network with the corresponding geographical grid units to generate the attribute coding units.

3. The prediction model training method based on land survey data according to claim 2, wherein The performing cross-modal feature fusion on the attribute coding units and the dynamic state sequence to generate a land spatial feature map includes: Perform frequency domain transformation on the surface temperature distribution data in the dynamic state sequence to extract temperature fluctuation patterns at different time scales; Align the temperature fluctuation patterns with the soil attribute vectors of the corresponding geographical grid units in the time domain to generate a mixed feature vector integrating soil-temperature association features; Determine the feature diffusion weights between nodes according to the transmission path of the mixed feature vector in the multi-level spatial association network; Perform spatial interpolation calculation on the mixed feature vector based on the feature diffusion weights to generate a continuous feature distribution map covering the entire target area, and convert the continuous feature distribution map into graph structure data to form the land spatial feature map.

4. The prediction model training method based on land survey data according to claim 3, characterized in that Extracting the abnormal fluctuation nodes in the land spatial feature map includes: Calculating the feature change rate of each node in the land spatial feature map, and constructing an anomaly detection threshold based on the difference in feature change rates between adjacent nodes; Identifying the nodes whose feature change rate exceeds the anomaly detection threshold, and tracing the soil composition mutation records in the historical land attribute distribution data; Marking the nodes that simultaneously meet the feature rate anomaly and soil composition mutation as the abnormal fluctuation nodes, and recording their corresponding longitude and latitude coordinates and timestamp information.

5. The prediction model training method based on land survey data according to claim 1, characterized in that The multi-stage iterative training of the initial prediction model based on the target training sample set includes: Dividing the target training sample set into a training subset and a validation subset, where the training subset contains geographical block data with clear soil degradation labels, and the validation subset contains geographical block data without labeled soil state changes; In the first training stage, using the training subset to initialize the parameters of the initial prediction model, and generating initial prediction weights based on the correlation between soil composition information and surface temperature distribution; In the second training stage, inputting the validation subset into the initialized prediction model, performing error backpropagation according to the vegetation cover change probability output and real-time satellite image data, and adjusting the dynamic parameters in the initial prediction weights; In the third training stage, deploying the adjusted prediction model to edge computing devices, online calibrating the prediction results through the real-time received environmental sensor data, and updating the final parameters of the land state prediction model.

6. The method for training a prediction model based on land survey data according to claim 5, characterized in that The online calibration of the prediction results through the real-time received environmental sensor data includes: Obtaining the soil moisture sensor readings and light intensity monitoring values of the target geographical block in the current time period; Comparing the soil moisture sensor readings with the vegetation cover change probability output by the land state prediction model to determine the humidity deviation coefficient; Weightedly correcting the humidity deviation coefficient according to the light intensity monitoring value to generate an environmental calibration factor; Superimposing the environmental calibration factor on the output layer of the land state prediction model to dynamically adjust the soil degradation risk assessment results in the future time period.

7. The prediction model training method based on land survey data according to claim 1, characterized in that The method further includes: When detecting that the soil degradation risk assessment result of the target geographical block exceeds the preset threshold, generating a recommended list of land treatment plans; Extracting the resource consumption data and expected effect data required for each plan in the recommended list of land treatment plans, and constructing a multi-objective optimization decision model; According to the current available resource inventory data and environmental carrying capacity constraint conditions, screening the optimal land treatment plan from the multi-objective optimization decision model and outputting it to the user terminal.

8. The prediction model training method based on land survey data according to claim 7, wherein The generation of the recommended list of land treatment plans includes: Querying the successful case data in the historical land treatment case library that matches the current soil degradation type; Analyzing the governance technical paths, material ratio parameters, and implementation cycle information adopted in the successful case data; Combining the terrain slope data and water source distribution data of the target geographical block, adaptively transforming the governance technical path, and generating multiple candidate governance plans with priority rankings; Evaluate the matching degree of the candidate governance solutions with real-time weather forecast data, and eliminate the solutions that do not meet the climate conditions to form the recommended list of land governance solutions.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a computer, the method for training a prediction model based on land measurement data according to any one of claims 1-8 is implemented.

10. A prediction model training system based on land survey data, characterized in that, It includes a processor and a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a computer, the method for training a prediction model based on land measurement data according to any one of claims 1-8 is implemented.

Images