Agricultural value prediction method based on dynamic space-time weighting

Abstract

The embodiment of the invention provides a dynamic space-time weighted agricultural value prediction method, and belongs to the field of agricultural big data and artificial intelligence, and the method comprises the steps: obtaining the data information of crops in a target farm; performing gating mechanism capture on the historical yield, the meteorological factors and the crop phenological period data through a three-layer bidirectional LSTM network to obtain growth period time sequence characteristics of crops in the target farm; performing semantic analysis on the market information to obtain crop price influence characteristics; constructing a double-layer LSTM-Attention network architecture according to the time sequence characteristics of the growth period and the crop price influence characteristics; determining a GIS three-level value grid architecture according to geographic factors, logistics influences, transportation conditions and supply-demand relationships; and constructing an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture. According to the method, cost reduction and efficiency improvement of agriculture are realized through the agricultural value prediction model, and an agricultural planting scheme is optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural big data and artificial intelligence, in particular to a dynamic spatio-temporal weighted agricultural value prediction method. BACKGROUND

[0002] With the penetration rate of smart agriculture increasing, multi-source heterogeneous data fusion has become a key direction for technological breakthrough. In current industry practice, cross-domain collaboration of agronomic data, economic data and policy data still faces three major technical bottlenecks: real-time economic data processing lag, lack of spatio-temporal dynamic correlation, and broken industry empowerment chain.

[0003] Real-time economic data processing lag: cloud centralized processing mode of high-frequency data such as market price and logistics cost leads to a delay of value prediction generally exceeding 24 hours, which cannot respond to the short-cycle transaction demand of fresh agricultural products; Lack of spatio-temporal dynamic correlation: existing models lack dynamic coupling mechanism for spatial dimension differences and time dimension holiday consumption cycle, and the error rate of plot-level value prediction exceeds 35%; Broken industry empowerment chain: lack of closed-loop support from "production data-value assessment-financial application".

[0004] The popularity of Internet of Things technology and the maturity of edge computing architecture provide a hardware foundation for breaking through the above bottlenecks, but how to build a cross-domain model compatible with agronomic mechanism and economic law is still a core technical problem that needs to be solved in the current industry. The existing market price model generally uses ARIMA time series analysis, which only captures the price trend item and ignores the cost prediction error caused by the periodic item and irregular item. For example, when the urea price rises by 15%, the traditional model cannot simultaneously calculate the cost offset effect brought by the improvement of fertilization technology; relying on manually setting quality premium coefficient does not establish a dynamic mapping relationship. Therefore, the industry empowerment logic of the existing technology stays at the level of prediction result display, and lacks quantitative support capability. The agricultural value determined by the existing technology has problems such as single dimension, poor spatio-temporal dynamic adaptability, and insufficient real-time data processing capability, which greatly affects the precise control and yield of agriculture. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a dynamic spatio-temporal weighted agricultural value prediction method, which realizes the cost reduction and efficiency increase of agriculture, and optimizes the agricultural planting scheme through an agricultural value prediction model.

[0006] In order to achieve the above purpose, the embodiments of the present application provide a method for constructing an agricultural value prediction model, which comprises: obtaining data information of crops in a target farm, including historical yield, meteorological factor and crop phenological period data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship; The growth period time sequence characteristics of crops in the target farm are captured by gating the historical yield, meteorological factors and crop phenological data through a three-layer bidirectional LSTM network. The market information was semantically analyzed using the Transformer time attention mechanism to obtain the characteristics of crop price impact; A two-layer LSTM-Attention network architecture is constructed based on the temporal characteristics of the reproductive period and the characteristics of crop price influence. The GIS three-level value grid architecture is determined based on the aforementioned geographical factors, logistical impacts, transportation conditions, and supply and demand relationships; An agricultural value prediction model is constructed based on the described two-layer LSTM-Attention network architecture and GIS three-level value grid architecture.

[0007] Optionally, the step of capturing the temporal characteristics of crop growth periods in the target farm by using a three-layer bidirectional LSTM network to gate the historical yield, meteorological factors, and crop phenological data includes: A time series model is obtained by using a three-layer bidirectional LSTM network deployed on edge computing nodes to perform time series modeling on historical yield, meteorological factors and crop phenological data; The reproductive period sensitivity coefficient is obtained by dynamically adjusting the time series model based on the gating mechanism. The influence weights of dynamically amplified meteorological factors are determined based on the sensitivity coefficient of the reproductive period and the phenological period indicator function. The reproductive period time sequence characteristics are determined based on the influence weights, including the start and end times of the critical reproductive period, temperature sensitivity coefficient, cumulative precipitation requirement, and growth rate threshold.

[0008] Optionally, the influence weight is: ; in, Here, f(t) is the sensitivity coefficient for the reproductive period, and f(t) is the phenological period indicator function. The influence weight at the current time t, The influence weight of the previous time step t-1.

[0009] Optionally, the step of obtaining crop price impact characteristics by performing semantic analysis on the market information using the Transformer temporal attention mechanism includes: The data information is divided into four windows according to the season using the Transformer time attention mechanism; The cross-year similarity patterns of the four windows are calculated using self-attention; The characteristics of crop price impacts are determined based on the cross-year similarity patterns.

[0010] Optionally, the double-layer LSTM-Attention network architecture is constructed according to the growth period time sequence feature and the crop price influence feature, and the double-layer LSTM-Attention network architecture comprises: The time dimension of the double-layer LSTM-Attention network architecture is determined according to the growth period time sequence feature. The space dimension of the double-layer LSTM-Attention network architecture is determined according to the crop price influence feature.

[0011] Optionally, the GIS three-level value grid architecture comprises a plot level, a production area level and a market level. The plot level is a continuous weight surface generated by Kriging interpolation based on an inverse distance weighted algorithm and taking the organic matter content of soil sampling points at a threshold distance as samples. The production area level is a threshold buffer zone established with the market as the center, and the influence of traffic convenience on logistics cost is determined according to the threshold buffer zone. The market level is a regional supply and demand influence range divided by a Thiessen polygon, wherein each polygon vertex in the regional supply and demand influence range is set as the location of a wholesale market, and the edge distance is set as the supply and demand balance point.

[0012] Optionally, the soil fertility weight factor in the plot level is: The logistics cost weight factor in the production area level is: wherein, is the soil fertility weight factor, is the logistics cost weight factor, OM is the measured organic matter content of the plot, is the average value of the regional organic matter content, is the standard deviation of the regional organic matter content, the constant 0.15 is an organic matter premium coefficient, the constant 0.001 is a distance attenuation coefficient, and d is the Euclidean distance from the plot to the target wholesale market.

[0013] Optionally, the market weight factor in the market level is dynamically adjusted according to the market price. The correlation weight factor of the plot level, the production area level and the market level is dynamically adjusted according to the quality monitoring value, and the quality monitoring value comprises a crop sugar content value and a crop moisture content value.

[0014] Optionally, a time threshold and / or an event threshold are set to update at least one of the soil fertility weight factor, the logistics cost weight factor, the market weight factor and the correlation weight factor.

[0015] Optionally, the agricultural value prediction model is constructed according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture, and the agricultural value prediction model comprises: ​​According to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture, the predicted yield, market price, quality premium, agricultural input cost and logistics cost of the crops in the target farm are determined. According to the predicted yield, market price, quality premium, agricultural input cost and logistics cost of the crops in the target farm, an agricultural value prediction model is constructed.

[0016] Optionally, the agricultural value prediction model is:

[0017]

[0018] wherein Y is the predicted yield, P is the market price, is the quality premium, is the agricultural input cost, is the logistics cost.

[0019] In another aspect, the present application also provides a prediction method of agricultural value, which comprises predicting the value of the crops in the target farm according to the agricultural value prediction model obtained by the above-mentioned dynamic spatio-temporal weighted agricultural value prediction method.

[0020] In another aspect, the present application also provides a device for constructing an agricultural value prediction model, which comprises: an acquisition module, configured to acquire data information of crops in a target farm, including historical yield, meteorological factors and crop phenological data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship; a first processing module, configured to capture the growth period time sequence characteristics of the crops in the target farm by a three-layer bidirectional LSTM network through a gating mechanism on the historical yield, meteorological factors and crop phenological data; a second processing module, configured to obtain crop price influence characteristics by semantic analysis of the market information through a Transformer time attention mechanism; a third processing module, configured to construct a double-layer LSTM-Attention network architecture according to the growth period time sequence characteristics and the crop price influence characteristics; a fourth processing module, configured to determine a GIS three-level value grid architecture according to the geographical factors, logistics influence, transportation conditions and supply and demand relationship; a fifth processing module, configured to construct an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture.

[0021] In another aspect, the present application also provides a machine readable storage medium, which stores instructions, wherein the instructions, when executed by a processor, cause the processor to be configured to perform the dynamic spatio-temporal weighted agricultural value prediction method and / or the agricultural value prediction method.

[0022] The dynamic spatio-temporal weighted agricultural value prediction method comprises: obtaining data information of crops in a target farm, including historical yield, meteorological factors and crop phenological data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship; obtaining growth period time sequence characteristics of crops in the target farm through a three-layer bidirectional LSTM network for gate mechanism capturing of the historical yield, meteorological factors and crop phenological data; obtaining crop price influence characteristics through semantic analysis of the market information by a Transformer time attention mechanism; constructing a double-layer LSTM-Attention network architecture according to the growth period time sequence characteristics and the crop price influence characteristics; determining a GIS three-level value grid architecture according to the geographical factors, logistics influence, transportation conditions and supply and demand relationship; and constructing an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture. The four-dimensional feature space is formed by integrating agronomic data, quality data, cost data and market data through the dynamic spatio-temporal weighting mechanism. The double-layer LSTM-Attention network is adopted to capture the growth period law in the time dimension, and the GIS three-level value grid is adopted to analyze spatial heterogeneity, so as to reduce the error of value prediction and realize cost reduction and benefit increase of agriculture.

[0023] Other features and advantages of the present application will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0024] The accompanying drawings are included to provide a further understanding of embodiments of the application, and constitute a part of the specification, and are used to explain the embodiments of the application together with the specific embodiments below, but do not constitute a limitation on the embodiments of the application. In the drawings: Figure 1 is a flowchart of a dynamic spatio-temporal weighted agricultural value prediction method of the present application: Figure 2 is a schematic diagram of a device for constructing an agricultural value prediction model of the present application.

[0025] Explanation of reference signs 100 - device for constructing an agricultural value prediction model; 200 - acquisition module; 300 - first processing module; 400 - second processing module; 500 - third processing module; 600 - fourth processing module; 700 - fifth processing module. DETAILED DESCRIPTION

[0026] The specific embodiments of the embodiments of the application are described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the embodiments of the application, and are not used to limit the embodiments of the application.

[0027] It should be noted that the acquisition, transmission, storage, use, processing, etc. of data in the technical solutions of the present application comply with the relevant provisions of laws and regulations. In the embodiments of the present application, some industry existing solutions, components, models, etc. may be mentioned, which should be considered as exemplary, and the purpose is only to illustrate the feasibility of the implementation of the technical solutions of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0028] Figure 1 is a flowchart of a dynamic spatiotemporal weighted agricultural value prediction method of the present application, as Figure 1 The dynamic spatiotemporal weighted agricultural value prediction method of the present application includes: step S101, acquiring data information of crops in a target farm, including historical yield, meteorological factors and crop phenology data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship. The meteorological factors mainly include the following four aspects: wind speed, temperature, humidity and air pressure.

[0029] The crop phenology data mainly includes: growth stage data, environmental response parameters, physiological index threshold values and dynamic correlation factors. Among them, the growth stage data includes key growth period identification, growth interval accurate to day, sowing period, jointing period, heading period, filling period and maturity period; the environmental response parameters include temperature sensitivity, cumulative precipitation demand and light cycle response; the physiological index threshold values include growth rate threshold value, accumulated temperature demand and nutrient critical value; the dynamic correlation factors include quality correlation items, stress response parameters (high temperature sensitive period) and embodiment evidence.

[0030] According to a specific embodiment, the data information acquisition method can be to deploy an ARM64 architecture processor (such as Nvidia Jetson AGX Orin) and Redis real-time cache on the edge node, pre-process local high-frequency data (update frequency 1 hour / time) such as market price and logistics cost, reduce 60% of the cloud traffic through feature-level segmentation technology (such as uploading only abnormal data with price fluctuation of more than 5%), realize lightweight deployment of the model by using a K3s lightweight container framework, shorten the end-to-end delay of market data from collection to value prediction from 24 hours in the traditional architecture to within 500 ms, and support instant transaction decision-making of fresh agricultural products.

[0031] Specifically, data collection and edge preprocessing can include multi-source data synchronous collection, economic data collection, logistics cost collection, and policy data collection. Among them, the multi-source data synchronous collection can be agricultural data collection, and a LoRaWAN sensor network (coverage radius 5km) is deployed in the target farm for data collection, including: soil sensors (pH, EC value, organic matter content, sampling frequency 10 minutes); weather station (temperature, precipitation, sunshine hours, sampling frequency 1 hour); unmanned aerial vehicle multispectral camera (NDVI, EVI vegetation index, daily aerial photography 1 time). The economic data collection includes: collecting market prices, and the market data collection adopts an official data direct connection + authorized cooperation mode: connecting the standardized open API of the national large agricultural product trading center (Beijing Xinfadi, Shanghai Jiangqiao, etc.), obtaining the price and transaction volume data of vegetables, fruits, grains and oils, etc. through a compliance agreement, configuring 1-hour granularity timing pulling, combining the agricultural industry big data platform to supplement regional transaction information, and the data collection success rate reaches 98%. The logistics cost is obtained through the national logistics public information platform and the authorization of third-party service providers: relying on the 10km resolution transportation cost (Yuan / ton•km) obtained from the national road freight supervision platform API, and through Kriging interpolation to refine to 50m grid; At the same time, integrate the public quotation system of Shunfeng cold transport, Jingdong logistics, etc., covering cold chain, bulk transportation and other scenarios. Relying on official channels and knowledge graph analysis: subscribing to policy publishing system RSS and standardized API, combining BERT-Chinese pre-training model to build agricultural policy knowledge graph (including 2000+ clause entities), accurately extracting structured information such as subsidy standards and applicable conditions, realizing policy semantic analysis and dynamic update, and then optimizing the agricultural planting scheme.

[0032] Step S102 is to obtain the growth period time sequence characteristics of crops in the target farm through a three-layer bidirectional LSTM network to capture the historical yield, meteorological factors and crop phenology data.

[0033] The gating mechanism is a technology for dynamically controlling information flow, which is applied in deep learning models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM), and determines which information should be retained or filtered through the gate.

[0034] According to a specific embodiment, the growth period time sequence characteristics of the crops in the target farm are captured by a three-layer bidirectional LSTM network through a gating mechanism, comprising: a three-layer bidirectional LSTM network deployed on an edge computing node is used to perform time sequence modeling on the historical yield, meteorological factors and crop phenology data to obtain a time sequence model; the time sequence model is dynamically adjusted based on a gating mechanism to obtain a growth period sensitive coefficient; a dynamic amplification meteorological factor influence weight is determined according to the growth period sensitive coefficient and a phenology indicator function; and a growth period time sequence characteristic is determined according to the influence weight, wherein the growth period time sequence characteristic includes the start and end time of a key growth period, a temperature sensitivity coefficient, a cumulative precipitation requirement and a growth rate threshold.

[0035] Specifically, the influence weight is: ; wherein, is a growth period sensitive coefficient, and f(t) is a phenology indicator function, is the influence weight of the current time t, is the influence weight of the previous time t-1.

[0036] For example, a three-layer bidirectional LSTM network (256 memory units per layer) is used, and a Layer Normalization layer is used to solve the gradient disappearance problem. The input includes 12 months of historical yield (weekly granularity), meteorological factors (accumulative temperature / precipitation / sunshine) and crop phenology data, and the time sequence characteristics of the key growth period are captured through a gating mechanism. For example, in the rice grain filling period (July-August), the model automatically increases the temperature factor weight by 20%-30%, and the corresponding formula is: ; wherein is a growth period sensitive coefficient (0.2-0.3), f t is a phenology indicator function. When it is detected that the market price fluctuation in the same period of history is more than 20%, a cross-year pattern matching algorithm is triggered, the similarity between the current sequence and the historical sequence is calculated by DTW (dynamic time warping), the weight coefficient of the year with a similarity of more than 80% is reused, and the response delay is controlled within 100 ms. This method can accurately obtain the growth period time sequence characteristics of the crops in the target farm.

[0037] Step S103 is to perform semantic analysis on the market information by a Transformer time attention mechanism to obtain crop price influence characteristics.

[0038] ​According to a specific embodiment, the semantic analysis of the market information by the Transformer time attention mechanism obtains crop price influence characteristics, including: dividing the data information into four windows according to seasons by the Transformer time attention mechanism; calculating the cross-year similar patterns of the four windows by self-attention; and determining crop price influence characteristics according to the cross-year similar patterns.

[0039] Specifically, the Transformer time attention mechanism is introduced, and the historical 10-year data is divided into four windows (spring, summer, autumn, and winter) according to seasons, and the cross-year similar patterns are calculated by self-attention. For example, for the rule of rising vegetable prices before the Spring Festival, the model learns that the price increases by 1.2% per day during the period from the 25th day of the 12th month to New Year's Eve, automatically adjusts the market factor weight, and the time series prediction error rate is ≤5%. The layer adopts a causal convolution structure to ensure that the prediction does not depend on future data, thereby improving the accuracy of time series prediction.

[0040] Step S104 is to construct a double-layer LSTM-Attention network architecture according to the growth period time series characteristics and the crop price influence characteristics. The LSTM-Attention architecture is a deep learning neural network architecture that combines a long short-term memory network (LSTM) and an attention mechanism (Attention).

[0041] According to a specific embodiment, the double-layer LSTM-Attention network architecture is constructed according to the growth period time series characteristics and the crop price influence characteristics, including: determining the time dimension of the double-layer LSTM-Attention network architecture according to the growth period time series characteristics; and determining the spatial dimension of the double-layer LSTM-Attention network architecture according to the crop price influence characteristics.

[0042] Specifically, the double-layer LSTM-Attention network includes: a first layer (agronomy layer): three layers of bidirectional LSTM (256 units per layer), which captures the temperature weight increase of 20%-30% in key growth periods such as rice grain filling period, where γ=0.25, and f(t) is a phenology indicator function; and a second layer (market layer): a Transformer network that analyzes 10 years of historical data and identifies a 1.2% daily increase in vegetable prices before the Spring Festival, with a response delay ≤100 ms.

[0043] This method focuses on different problems according to different key points, thereby improving the accuracy and robustness of the double-layer LSTM-Attention network architecture.

[0044] Step S105 is to determine a GIS three-level value grid architecture according to the geographical factors, the logistics influence, the transportation conditions, and the supply and demand relationship.

[0045] According to a specific embodiment, the GIS three-level value grid architecture includes a plot level, a production area level, and a market level: the plot level is a continuous weight surface generated by Kriging interpolation based on inverse distance weighted algorithm, with the organic matter content of soil sampling points at a threshold distance as samples; the production area level is a threshold buffer zone established around the market, and the influence of traffic convenience on logistics cost is determined according to the threshold buffer zone; and the market level is to divide the regional supply and demand influence range by using the Thiessen polygon, and set the vertex of each polygon in the regional supply and demand influence range as the location of the wholesale market, and the edge distance as the supply and demand balance point.

[0046] The soil fertility weight factor in the plot level is: The logistics cost weight factor in the production area level is: ; wherein, is the soil fertility weight factor, is the logistics cost weight factor, OM is the measured organic matter content of the plot, is the average value of the regional organic matter content, is the standard deviation of the regional organic matter content, the constant 0.15 is the organic matter premium coefficient, the constant 0.001 is the distance attenuation coefficient, and d is the Euclidean distance from the plot to the target wholesale market.

[0047] Specifically, the plot level (50m x 50m) is set, the continuous weight surface is generated by Kriging interpolation based on the inverse distance weighted (IDW) algorithm, with the pH organic matter content of soil sampling points at a distance of 200m as samples. For example, the fruit sugar degree of high organic matter plot (OM>2%) has a premium of 15%, and the corresponding weight calculation formula is: The production area level (county scale) is set, and the influence of traffic convenience on logistics cost is calculated by using GIS buffer analysis to establish a 100km buffer zone around the wholesale market: The market level (regional economic unit) is set, and the regional supply and demand influence range is divided by using the Thiessen polygon, and the vertex of each polygon is set as the location of the wholesale market, and the edge distance is set as the supply and demand balance point. When the regional supply and demand ratio is >1.2, the purchase price weight is reduced by 10%, and vice versa, the supply and demand ratio is <0.8, the weight is increased by 15%, and the spatial expression of market supply and demand relationship is realized.

[0048] Based on the GIS inverse distance weighted algorithm, a 50m x 50m plot level value grid is generated, combining soil fertility (such as organic matter content), traffic distance (Euclidean distance to wholesale market), market supply and demand (price influence range of Thiessen polygon division) and other parameters, and a input-output ratio (ROI) calculation model is constructed: yield price quality premium (agricultural material cost logistics cost) total input. Through the visualization of the heat map, when the ROI of a certain plot is lower than the industry average by 15%, the system automatically recommends crop replacement schemes (such as changing wheat to high-value-added vegetables). In the application of a certain agricultural park in the Yangtze River Delta, the system guides farmers to adjust the planting structure, and the plot-level ROI is increased by 22%, the use of chemical fertilizers is reduced by 18%, and the annual per mu income is increased by 370.

[0049] The plot level of the present application calculates the soil fertility premium by the inverse distance weighted algorithm, and the high organic matter plot (OM>2%) has a sugar premium of 15%; the logistics cost weight is determined by the buffer zone analysis in the production area, and the circulation cost increases by 8% for every 100km away from the wholesale market; the supply and demand influence range is divided by the Thiessen polygon in the market level, and the purchase price weight decreases by 10% when the supply-demand ratio is >1.2.

[0050] The method further comprises: dynamically adjusting the market weight factor in the market level according to the market price; and dynamically adjusting the correlation weight factor of the plot level, the production area level and the market level according to a quality monitoring value, the quality monitoring value including a crop sugar value and a crop moisture content value. A time threshold and / or an event threshold are set to update at least one of the soil fertility weight factor, the logistics cost weight factor, the market weight factor and the correlation weight factor.

[0051] Specifically, the market weight factor adjustment: when the market price minute-level data daily fluctuation is >5% (the purchase price increases by 25% in one day), the market level weight factor is updated dynamically. Taking the Thiessen polygon supply-demand ratio as the benchmark: when the regional supply-demand ratio is >1.2, the weight decreases by 10%, and when the supply-demand ratio is <0.8, the weight increases by 15%. The edge node calculates the price gradient in real time (gradient=(current price-historical average price) / standard deviation), and the weight adjustment is completed within 3 hours, and the actual measurement error is reduced to 9.1%.

[0052] Quality correlation weight adjustment: based on the unmanned aerial vehicle multispectral inversion sugar value (accuracy ±0.5°Brix) or soil moisture monitoring (error ≤3%), when the measured value deviates from the prediction by >10% (sugar drops by 12%), the multi-level linkage update is triggered.

[0053] Specifically, the application can set a market price fluctuation threshold: when the daily fluctuation of minute-level price data is more than 5%, trigger the LSTM network parameter fine-tuning, and use the SGD-Momentum algorithm (learning rate 0.001) to update the market factor weight; set the quality index mutation threshold: through spectral remote sensing monitoring of indicators such as sugar content, when the measured value is more than 10% lower than the predicted value, start the quality model retraining, and update the correlation weight of soil-quality-price. This mechanism enables the model to complete weight reconstruction within 30 minutes in extreme weather (such as high temperature leading to sugar content reduction), and the prediction error rate is reduced by 52% compared with the static model.

[0054] For example, when the daily fluctuation of minute-level market price is more than 5%, trigger incremental fine-tuning: freeze non-market layer parameters to lock the weights of agronomy layer and quality layer (full connection). When the deviation between the measured value and the predicted value of spectral inversion sugar content is more than 10%, obtain 50m grid soil organic matter (OM), real-time sugar content (S), and historical premium records to build a dynamic mapping of soil OM, sugar content, and premium.

[0055] The application also includes three-level data cleaning of the GIS three-level value grid architecture, including physical layer filtering, statistical layer cleaning, and semantic layer standardization processing. The physical layer filtering includes: establishing an agricultural economic data anomaly rule library, and filtering the physical layer data according to the anomaly rule library, such as price rules: vegetable price > 50 yuan / kg (regular category) is considered abnormal; cost rules: logistics cost > 3 times the regional average is considered abnormal; real-time filtering is performed using a rule engine, and the abnormal data identification rate is 99.2%.

[0056] The statistical layer cleaning is for continuous data such as logistics cost, which uses the 3σ principle combined with the Isolation Forest algorithm: calculate the 3σ range, and remove data outside the mean ± 3 times the standard deviation; build an isolation forest for the remaining data, and samples with a tree depth of more than 20 are considered outliers; this combination strategy improves the accuracy of outlier identification to 98.3%.

[0057] The semantic layer standardization is to convert policy text into value factors according to the developed agricultural economic ontology mapping tool, such as green certification awarding product premium of 12%; agricultural machinery purchase subsidy cost reduction of 8%; semantic consistency verification is realized through knowledge graph reasoning, and the feature dimension of the standardized data is reduced from 120 to 45, improving the model training efficiency by 30%.

[0058] Step S106 is to build an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture.

[0059] According to a specific embodiment, the agricultural value prediction model is constructed according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture, including: determining the predicted yield, market price, quality premium, agricultural input cost and logistics cost of crops in the target farm according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture; and constructing an agricultural value prediction model according to the predicted yield, market price, quality premium, agricultural input cost and logistics cost of crops in the target farm.

[0060] Specifically, the agricultural value prediction model is:

[0061]

[0062] wherein Y is the predicted yield, P is the market price, is the quality premium, is the agricultural input cost, is the logistics cost.

[0063] For example, the system generates a 50m x 50m ROI heat map, when the ROI of a certain plot is < 15% of the industry average, the crop replacement recommendation algorithm is started, for example: wheat (ROI = 1.2) is replaced by strawberry (ROI = 3.5), which needs to meet: soil pH 5.5-6.5; distance from market ≤ 50 km; irrigation conditions meet. This recommendation mechanism makes the per mu income of farmers in a certain production area in North China increase by 2200 yuan.

[0064] The application also includes a design of a three-level evaluation system, including expected value calculation, such as generating a value distribution using Monte Carlo simulation, for example, the expected value of a certain orchard = 3250 kg / mu x 2.8 yuan / kg x 1.12 (quality premium) = 10192 yuan / mu, with a 95% confidence interval [9850, 10530]; risk quantification indicators, such as calculating the price volatility rate (60-day historical standard deviation) and the probability of natural disasters (based on a meteorological model) to generate a risk level (low, medium, high). A pilot showed that this indicator reduced the loan delinquency rate from 4.2% to 2.8%; blockchain notarization valuation, using Hyperledger Fabric to notarize yield prediction process data, and generating a mortgage valuation report through a smart contract. The notarized data includes: original sensor data hash model training parameters, value prediction results, and bank-verified valuation credibility through on-chain verification. This method shortens the loan approval period, reducing the loan approval period from 15 days to 3 days.

[0065] The application also constructs a blockchain-enabled authentication process: planting data is chained, such as real-time writing of organic fertilizer application records (2000 kg / acre), pest control logs, etc. into the blockchain, and the hash value is stored in block height n; quality evidence is generated, such as spectral remote sensing data (NDVI = 0.85) and laboratory test reports (sugar content 8.3°) associated through a smart contract to automatically generate a quality evidence package; authentication materials are automated, such as generating 16 materials required for green food certification based on a template engine, including: planting process blockchain storage report; quality prediction and actual measurement comparison analysis; input-output ratio compliance proof; the application of a certain organic vegetable base shows that the authentication period is shortened from 90 days to 30 days, the premium rate is increased from 8% to 18%, and the annual sales are also increased.

[0066] The application stores planting process data (such as organic fertilizer application records) through a consortium chain, and a smart contract automatically generates value proof materials required for green food certification, shortening the authentication period from the traditional 90 days to 30 days, and promoting the product premium rate from less than 5% to 12%-18%. At the same time, a standardized financial report containing expected value, risk assessment (price volatility rate, natural disaster probability), and blockchain collateral value is generated, and federal learning is used to protect data privacy, so that the loan approval period of farmers is compressed from 15 days to 3 days, and the approval rate is increased from 38% to 75%. After a certain pilot county accesses the system, the non-performing rate of agricultural loans is reduced from 4.2% to 2.8%, the loan size is increased by 60%, and the industrial value cycle of precise valuation, low-risk lending, and high-premium sales is realized.

[0067] The application forms a technical closed loop through a dynamic spatio-temporal weighting mechanism, and the spatio-temporal weight is dynamically adjusted depending on real-time data supported by edge computing. The land-level analysis provides accurate data foundation for blockchain storage, and finally realizes the whole-chain value improvement from technical innovation to industrial empowerment. Compared with the prior art, the application realizes three major breakthroughs of agricultural value prediction from post-event statistics to real-time decision-making, from regional estimation to land-level calculation, and from single yield to multi-dimensional value.

[0068] According to another embodiment, the application can integrate yield and market average price (such as historical average price x predicted yield) with fixed weight, and the weight value is set by experience (such as market factor weight 0.3 fixed). A linear regression model is used without considering spatio-temporal dynamic factors; this method is suitable for wheat, rice and other stable price bulk agricultural products. This method is suitable for scenarios with low real-time requirements and small market price fluctuations. The application can also use an ARIMA model to analyze price time series, ignoring crop quality differences. This method is suitable for low-value-added and small-quality-difference agricultural products (such as potatoes and Chinese cabbage). The application can also use Hadoop cluster batch processing to generate value reports daily, which is suitable for annual value planning and low-time-efficiency agricultural policy simulation and long-term planting planning.

[0069] In another aspect, the present application also provides a method for predicting agricultural value, which comprises predicting the value of crops in a target farm according to the agricultural value prediction model obtained by the dynamic spatio-temporal weighted agricultural value prediction method.

[0070] In another aspect, the present application also provides a device for constructing an agricultural value prediction model, as shown in the accompanying drawings, which comprises an acquisition module 200 for acquiring data information of crops in a target farm, including historical yield, meteorological factors and crop phenology data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship; a first processing module 300 for capturing the growth period time sequence characteristics of crops in the target farm by a three-layer bidirectional LSTM network through a gating mechanism on the historical yield, meteorological factors and crop phenology data; a second processing module 400 for obtaining crop price influence characteristics by semantic analysis of the market information through a Transformer time attention mechanism; a third processing module 500 for constructing a double-layer LSTM-Attention network architecture according to the growth period time sequence characteristics and crop price influence characteristics; a fourth processing module 600 for determining a GIS three-level value grid architecture according to the geographical factors, logistics influence, transportation conditions and supply and demand relationship; and a fifth processing module 700 for constructing an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture. Figure 2

[0071] In another aspect, the present application also provides a machine-readable storage medium having instructions stored thereon, wherein the instructions, when executed by a processor, cause the processor to be configured to perform the dynamic spatio-temporal weighted agricultural value prediction method and / or the method for predicting agricultural value.

[0072] ​The application discloses a dynamic space-time weighted agricultural value prediction method, which comprises the following steps: obtaining data information of crops in a target farm, including historical yield, meteorological factors and crop phenological period data, market information, geographical factors, logistics influence, transportation conditions and supply and demand relationship; obtaining the growth period time sequence characteristics of the crops in the target farm by a three-layer bidirectional LSTM network through a gating mechanism; obtaining the crop price influence characteristics by performing semantic analysis on the market information through a Transformer time attention mechanism; constructing a double-layer LSTM-Attention network architecture according to the growth period time sequence characteristics and the crop price influence characteristics; determining a GIS three-level value grid architecture according to the geographical factors, the logistics influence, the transportation conditions and the supply and demand relationship; and constructing an agricultural value prediction model according to the double-layer LSTM-Attention network architecture and the GIS three-level value grid architecture. The application integrates agronomic data, quality data, cost data and market data to form a four-dimensional feature space through a dynamic space-time weighting mechanism. The double-layer LSTM-Attention network is adopted to capture the growth period law in the time dimension, and the GIS three-level value grid is adopted to analyze spatial heterogeneity, so that the error of value prediction is reduced, and the cost reduction and benefit increase of agriculture are realized.

[0073] The application reduces data noise and eliminates prediction deviation caused by invalid data and space-time misplacement by acquiring multi-source data such as historical yield, meteorological factors, crop phenological period, market information and geographical factors, and performing three-level cleaning (99.2% of physical layer abnormal recognition rate, 98.3% of statistical layer outlier recognition accuracy) and space-time alignment (space error is less than or equal to 1 pixel), thereby providing a high-purity data basis for modeling and improving the reliability and consistency of input data. The application adopts a three-layer bidirectional LSTM network to capture growth period time sequence characteristics, dynamically adjusts the weight of meteorological factors in the key growth period through a gating mechanism, improves the capture accuracy of key growth period characteristics, reduces yield prediction error, reduces cross-year model matching response delay, realizes rapid reuse of historical similar year weight, and enhances the real-time performance of time dimension modeling. The application adopts a Transformer time attention mechanism to analyze market information, identifies market price cycle characteristics through seasonal window division and self-attention calculation, reduces market price prediction error, for example, the capture accuracy of sudden price increase (vegetable price increases by 20%) in the Mid-Autumn Festival and other periods is improved from 30% of a traditional model to 85%, effectively distinguishes price trend items, cycle items and burst items, and reduces deviation caused by market law misjudgment.

[0074] The double-layer LSTM-Attention network architecture of the present application integrates the fertility period time sequence characteristics and market price influence characteristics, dynamically adjusts the weight, and reduces the comprehensive value prediction error rate; improves the robustness and stability of the model in extreme weather or market fluctuations. The GIS three-level value grid architecture of the present application improves the spatial resolution and plot-level evaluation through plot-level, production area-level, and market-level spatial weight calculation (soil fertility weight), reduces the prediction error of the quality premium caused by soil organic matter, reduces the production area-level logistics cost calculation error, and improves the market-level supply and demand ratio adjustment accuracy. The present application reduces the analysis error of the plot-level ROI based on the ROI model, improves the per mu income after the farmer adjusts the planting structure, reduces the amount of chemical fertilizer used, reduces the logistics cost, shortens the financial loan approval cycle and the high-quality agricultural product certification cycle.

[0075] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer-usable program code embodied in the medium.

[0076] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0077] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0078] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 Figure 1

[0079] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0080] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the processor can execute instructions. The memory can also include non-volatile memory, such as read only memory (ROM), electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), or flash memory, for example. Memory is an example of computer readable media.

[0081] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0082] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0083] ​​The above merely provides an example of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall fall into the scope of claims of the present application.

Claims

1. A dynamic spatiotemporal weighted method for predicting agricultural value, characterized in that, The method includes: Obtain data on crops from the target farm, including historical yields, meteorological factors and crop phenological data, market information, geographical factors, logistical impacts, transportation conditions and supply and demand relationships; The growth period time sequence characteristics of crops in the target farm are captured by gating the historical yield, meteorological factors and crop phenological data through a three-layer bidirectional LSTM network. The market information was semantically analyzed using the Transformer time attention mechanism to obtain the characteristics of crop price impact; A two-layer LSTM-Attention network architecture is constructed based on the temporal characteristics of the reproductive period and the characteristics of crop price influence. The GIS three-level value grid architecture is determined based on the aforementioned geographical factors, logistical impacts, transportation conditions, and supply and demand relationships; An agricultural value prediction model is constructed based on the described two-layer LSTM-Attention network architecture and GIS three-level value grid architecture.

2. The method according to claim 1, characterized in that, The process involves using a three-layer bidirectional LSTM network to perform gating on historical yield, meteorological factors, and crop phenological data to capture the temporal characteristics of crop growth periods in the target farm, including: A time series model is obtained by using a three-layer bidirectional LSTM network deployed on edge computing nodes to perform time series modeling on historical yield, meteorological factors and crop phenological data; The reproductive period sensitivity coefficient is obtained by dynamically adjusting the time series model based on the gating mechanism. The influence weights of dynamically amplified meteorological factors are determined based on the sensitivity coefficient of the reproductive period and the phenological period indicator function. The reproductive period time sequence characteristics are determined based on the influence weights, including the start and end times of the critical reproductive period, temperature sensitivity coefficient, cumulative precipitation requirement, and growth rate threshold.

3. The method according to claim 2, characterized in that, The influence weight is: ; in, Here, f(t) is the sensitivity coefficient for the reproductive period, and f(t) is the phenological period indicator function. The influence weight at the current time t, The influence weight of the previous time step t-1.

4. The method according to claim 1, characterized in that, The process of obtaining crop price impact characteristics through semantic analysis of market information using the Transformer time attention mechanism includes: The data information is divided into four windows according to the season using the Transformer time attention mechanism; The cross-year similarity patterns of the four windows are calculated using self-attention; The characteristics of crop price impacts are determined based on the cross-year similarity patterns.

5. The method according to claim 1, characterized in that, The construction of a two-layer LSTM-Attention network architecture based on the temporal characteristics of the reproductive period and the influence characteristics of crop prices includes: The temporal dimension of the two-layer LSTM-Attention network architecture is determined based on the temporal characteristics of the reproductive period. The spatial dimension of the two-layer LSTM-Attention network architecture is determined based on the characteristics of crop price impact.

6. The method according to claim 1, characterized in that, The GIS three-tier value grid architecture includes land parcel level, production area level, and market level: The plot-level data is a continuous weighted surface generated by Kriging interpolation based on the inverse distance weighting algorithm, using the organic matter content of soil sampling points with threshold spacing as samples. The production area level is to establish a threshold buffer zone centered on the market, and the impact of transportation convenience on logistics costs is determined based on the threshold buffer zone. The market level is defined by using Thiessen polygons to divide the regional supply and demand influence range. Each vertex of the polygon in the regional supply and demand influence range is set as the location of the wholesale market, and the edge distance is set as the supply and demand equilibrium point.

7. The method according to claim 6, characterized in that, The soil fertility weighting factor at the plot level is: ; The weighting factor for logistics costs at the production area level is: ; in, As a weighting factor for soil fertility, The logistics cost weighting factor is OM, which represents the measured organic matter content of the land parcel. This represents the average organic matter content of the region. denoted as the standard deviation of regional organic matter content, constant 0.15 as the organic matter premium coefficient, constant 0.001 as the distance decay coefficient, and d as the Euclidean distance from the plot to the target wholesale market.

8. The method according to claim 6, characterized in that, The market weight factor in the market level is dynamically adjusted according to market prices. The correlation weighting factors of the plot-level, production area-level, and market-level are dynamically adjusted based on the quality monitoring values, which include crop sugar content and crop moisture content.

9. The method according to claim 7 or 8, characterized in that, Set time thresholds and / or event thresholds to update at least one of the soil fertility weighting factor, logistics cost weighting factor, market weighting factor, and related weighting factor.

10. The method according to claim 1, characterized in that, The construction of the agricultural value prediction model based on the two-layer LSTM-Attention network architecture and the GIS three-level value grid architecture includes: Based on the aforementioned two-layer LSTM-Attention network architecture and GIS three-level value grid architecture, the predicted yield, market price, quality premium, agricultural input cost, and logistics cost of crops in the target farm are determined. An agricultural value prediction model is constructed based on the predicted yield, market price, quality premium, agricultural input cost, and logistics cost of crops in the target farm.

11. The method according to claim 10, characterized in that, The agricultural value prediction model for: Where Y represents the projected output and P represents the market price. Paying premium for quality, For agricultural input costs, For logistics costs.

12. A method for predicting agricultural value, characterized in that, The prediction method includes using an agricultural value prediction model obtained from the dynamic spatiotemporal weighted agricultural value prediction method according to any one of claims 1-11 to predict the value of crops in the target farm.

13. An apparatus for constructing an agricultural value prediction model, characterized in that, The device includes: The acquisition module is used to acquire data on crops in the target farm, including historical yields, meteorological factors and crop phenological data, market information, geographical factors, logistical impacts, transportation conditions and supply and demand relationships; The first processing module is used to capture the temporal characteristics of the growth period of crops in the target farm by using a three-layer bidirectional LSTM network to perform gating mechanism on the historical yield, meteorological factors and crop phenological data; The second processing module is used to perform semantic analysis on the market information through the Transformer time attention mechanism to obtain the characteristics of crop price impact. The third processing module is used to construct a two-layer LSTM-Attention network architecture based on the reproductive period time characteristics and crop price impact characteristics. The fourth processing module is used to determine the GIS three-level value grid architecture based on the geographical factors, logistics impact, transportation conditions, and supply and demand relationship. The fifth processing module is used to construct an agricultural value prediction model based on the dual-layer LSTM-Attention network architecture and the GIS three-level value grid architecture.

14. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, this instruction causes the processor to be configured to perform the dynamic spatiotemporal weighted agricultural value prediction method according to any one of claims 1-11, and / or the agricultural value prediction method according to claim 12.

Images