Crop yield prediction method and system based on artificial intelligence

Abstract

The present invention provides a crop yield prediction method and system based on artificial intelligence, which relates to the field of artificial intelligence technology, including obtaining historical meteorological data, soil data, crop planting data and remote sensing image data of a specified area for many consecutive years, aligning them according to time series, and constructing a multidimensional spatiotemporal data cube; training a fusion prediction model, the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust the model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm; obtaining real-time meteorological data, soil data and remote sensing image data of a specified area, inputting them into a pre-constructed fusion prediction model, and introducing an agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, inferring the crop growth and estimated yield of each grid in the specified area, performing spatial interpolation on the estimated yield of each grid, and generating a yield distribution map of the specified area.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to a crop yield prediction method and system based on artificial intelligence. Background Art

[0002] Traditional crop yield prediction mainly relies on statistical methods and agronomic models. Statistical methods collect and analyze historical yield data, use mathematical tools such as regression analysis and time series, and establish empirical relationships between yield and various influencing factors to estimate future yields. Agronomic models are based on the physiological and ecological mechanisms of crop growth and development, taking into account environmental factors such as light, temperature, water, soil, as well as the influence of variety characteristics, cultivation management measures, and other factors, and predict yields through numerical simulation.

[0003] However, the above methods all have certain limitations in practical applications. On the one hand, yield formation is a complex biological process, which is affected by many uncertain factors and it is difficult to establish an accurate mathematical model. On the other hand, with the intensification of climate change and the frequent occurrence of extreme weather events, the representativeness of historical data and the applicability of agronomic models are constantly decreasing, and the reliability of prediction results is facing challenges. Summary of the invention

[0004] The embodiments of the present invention provide a crop yield prediction method and system based on artificial intelligence, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provides artificial intelligence-based crop yield prediction methods, including:

[0007] Acquire historical meteorological data, soil data, crop planting data and remote sensing image data of the designated area for many consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and hydraulic conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data in time series to construct a multidimensional spatiotemporal data cube;

[0008] The multidimensional spatiotemporal data cube is used as a training data set to train a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, and the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm;

[0009] Real-time meteorological data, soil data and remote sensing image data of the specified area are obtained, and input into a pre-built fusion prediction model. The agricultural knowledge graph is introduced into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model. The crop growth and estimated yield of each grid in the specified area are inferred, and the estimated yield of each grid is spatially interpolated to generate a yield distribution map of the specified area.

[0010] In an optional embodiment,

[0011] Aligning the historical meteorological data, soil data, crop planting data and remote sensing image data in time series to construct a multidimensional spatiotemporal data cube includes:

[0012] Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data, the daily meteorological grid data including temperature, precipitation and sunshine hours;

[0013] Using the Kriging interpolation method, in combination with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute grid layer consistent with the spatial resolution of the remote sensing image data, wherein the soil attribute grid layer includes soil fertility, soil moisture, temperature and hydraulic conductivity;

[0014] Spatially matching the row and column numbers of the crop planting data with the remote sensing image data, assigning crop variety, sowing date and planting density attributes to each pixel, and estimating missing data using a multiple interpolation method to obtain crop planting attribute data;

[0015] Perform sliding window segmentation on the remote sensing image data on the time axis to extract growth period images corresponding to the crop greening period, jointing period, heading period and maturity period;

[0016] The daily meteorological grid data, the soil attribute grid layer, the crop planting attribute data and the growth period image are constructed with pixels as basic units in the spatial dimension and days or growth periods as units in the temporal dimension to form a multidimensional spatiotemporal data cube integrating time and space.

[0017] In an optional embodiment,

[0018] Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data includes:

[0019] ;

[0020] in, Indicates the meteorological location to be interpolated S 0Estimated values ​​of meteorological elements at n represents the number of meteorological stations involved in the interpolation, Z(S i ) Indicates i Weather stations S i The meteorological element observation values ​​at d i0 Indicates the position to be interpolated S 0 With i Weather stations S i The Euclidean distance between p represents the distance weighting parameter;

[0021] Using the Kriging interpolation method, combined with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute raster layer consistent with the spatial resolution of the remote sensing image data, including:

[0022] ;

[0023] in, T(L 0 ) Indicates the soil position to be interpolated L 0 Estimated values ​​of meteorological elements at m Indicates the number of soil sampling points involved in interpolation, k j Indicates j The Kriging weight coefficient of each sampling point is determined by the semivariogram and the spatial position relationship of the sampling points. ST(L j ) Indicates j Soil sampling points L j The attribute measurement value at .

[0024] In an optional embodiment,

[0025] The multidimensional spatiotemporal data cube is used as a training data set to train a fusion prediction model, the fusion prediction model is built based on a deep learning model, the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and the model parameters of the fusion prediction model are dynamically adjusted by an adaptive learning rate optimization algorithm, including:

[0026] The multidimensional spatiotemporal data cube is used as a training data set; a deep learning model is constructed, and a temporal convolutional network structure is adopted to respectively extract the local spatiotemporal correlation features and the global spatiotemporal correlation features of the training data set through a causal convolutional layer and a dilated convolutional layer, and an attention mechanism is introduced at the end of the network of the deep learning model to adaptively adjust the weights of the local spatiotemporal correlation features and the global spatiotemporal correlation features;

[0027] Construct a crop growth model, taking into account meteorological factors, soil properties and crop variety parameters, simulate photosynthesis, respiration, material accumulation and organ formation during crop growth and development, and dynamically adjust the adaptability of crop growth to adverse environmental conditions;

[0028] In the training process of the fusion prediction model, the field is used as the basic unit, and the variable sequence of crop growth status for consecutive years is paired with the image features of the corresponding period to construct a multi-dimensional spatiotemporal feature map with consistent time steps;

[0029] An adaptive learning rate optimization algorithm is used to train the crop growth model and the deep learning model separately. By setting different initial learning rates, learning rate decay methods, and regularization strategies, the training speed and generalization performance of the two sub-models are balanced.

[0030] In each training iteration step, the crop growth model is used to simulate the crop growth and development process, and the state variables of biomass and leaf area index in each growth period are output. They are then spliced ​​with the image features of the same period to obtain a multi-dimensional spatiotemporal feature map; the multi-dimensional spatiotemporal feature map is input into the deep learning model, and end-to-end joint training and parameter optimization are achieved through gradient back propagation.

[0031] In an optional embodiment,

[0032] The agricultural knowledge graph is introduced into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, and the crop growth and estimated yield of each grid in the specified area are inferred. The estimated yield of each grid is spatially interpolated to generate the yield distribution map of the specified area, including:

[0033] Construct an agricultural knowledge graph, extract agricultural concepts and semantic relationships from unstructured texts through natural language processing technology, design an agricultural ontology model, and use ontology construction tools to create the basic framework of the agricultural knowledge graph;

[0034] Based on the constructed agricultural knowledge graph, a rule-based reasoning method is used to mine the constraint rules implicit in the agricultural knowledge graph. According to the crop growth cycle, phenological period division, and the relationship between agricultural operations and yield, a series of constraint rule sets covering different constraint dimensions are constructed.

[0035] Using the trained fusion prediction model, the yield prediction value of each grid in the specified area is obtained, the specified area is divided into several grid units of equal size, each grid is assigned a unique number, the entities in the agricultural knowledge graph are spatially matched with the corresponding grids, and a mapping table between entities and grids is established to realize the association between grid-level prediction results and constraint rules;

[0036] For each grid, retrieve the entities and their attribute values ​​related to the grid from the agricultural knowledge graph, match the attribute values ​​with the constraint rules, and when the attribute values ​​meet the condition part of the constraint rules, trigger the conclusion part of the rule to correct the prediction results of the grid. By applying the constraint rules one by one, a systematic correction of the prediction results of the entire specified area can be achieved.

[0037] The spatial interpolation method is used to spatially interpolate the corrected prediction results, and GIS tools are used to generate the yield distribution map corresponding to the interpolated corrected prediction results.

[0038] In an optional embodiment,

[0039] For each grid, the entities and their attribute values ​​related to the grid are retrieved from the agricultural knowledge graph, and the attribute values ​​are matched with the constraint rules. When the attribute values ​​meet the condition part of the constraint rules, the conclusion part of the rule is triggered to correct the prediction results of the grid. By applying the constraint rules one by one, the systematic correction of the prediction results of the entire specified area is achieved, including:

[0040] Preprocess the constraint rules corresponding to the agricultural knowledge graph, convert the condition part of the constraint rules into the form of attribute value range, and convert the conclusion part into the form of correction coefficient to achieve quantitative representation of the constraint rules;

[0041] For each grid, retrieve the entity corresponding to the grid from the agricultural knowledge graph, and obtain the attributes and attribute values ​​associated with the entity to form a grid-level attribute value vector;

[0042] Index the preprocessed constraint rules, use the condition part of the constraint rules as the index key, use the attribute type as the dimension of the rule tree, and use the attribute value range as the node of the rule tree to build a multi-dimensional rule tree;

[0043] Use attribute value vectors and rule trees for fast matching, adopt a priority-based matching strategy, assign priority weights according to the confidence and specificity of constraint rules, give priority to matching constraint rules with priority weights greater than the preset weight threshold, and find all constraint rules that meet the conditions;

[0044] For the successfully matched constraint rules, their conclusion part is triggered, and the yield prediction value of the grid is corrected according to the correction coefficient of the constraint rule. When multiple constraint rules are met in the same grid, the weighted average method is adopted to comprehensively consider the correction coefficient and priority weight of each constraint rule, calculate the final correction coefficient, and realize the coordinated correction of multiple constraint rules.

[0045] According to a second aspect of the embodiments of the present invention,

[0046] The AI-based crop yield prediction system includes:

[0047] The first unit is used to obtain historical meteorological data, soil data, crop planting data and remote sensing image data of a designated area for many consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and water conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data according to time series to construct a multidimensional spatiotemporal data cube;

[0048] The second unit is used to use the multidimensional spatiotemporal data cube as a training data set to train a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, and the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm;

[0049] The third unit is used to obtain real-time meteorological data, soil data and remote sensing image data of the specified area, input them into a pre-built fusion prediction model, and introduce the agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, infer the crop growth and estimated yield of each grid in the specified area, perform spatial interpolation on the estimated yield of each grid, and generate a yield distribution map of the specified area.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] The present invention obtains data from multiple dimensions such as meteorology, soil, crop planting and remote sensing, and aligns them in time series to construct a multidimensional spatiotemporal data cube, fully exploring the complementary gain relationship between different data sources, thereby significantly improving the accuracy and reliability of yield prediction. The fusion prediction model is constructed based on the deep learning model, which can automatically learn and extract key features in the crop growth process, overcome the dependence of traditional machine learning models on prior knowledge, and introduce an adaptive learning rate optimization algorithm to dynamically adjust the model parameters, so that the model can adapt to the growth characteristics of different regions and different crops, and has stronger robustness and generalization ability.

[0058] Introducing agricultural knowledge graphs into the fusion prediction model constrains and corrects the prediction results. On the one hand, the prediction results are optimized by using the agronomic rules and expert experience in the knowledge graph to improve the accuracy of the prediction; on the other hand, by associating the prediction results with the concepts and entities in the knowledge graph, the prediction results are given clear semantic information, which enhances the interpretability of the prediction process and results, and helps agronomic experts to analyze and verify the model prediction. The crop growth and yield of each grid in the specified area are estimated, and the yield distribution map is generated through spatial interpolation, realizing the yield prediction from the regional scale to the farmland scale. It provides agricultural producers with more refined yield prediction information, which can directly guide the precise management of farmland and the optimal allocation of resources, and has important practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a process of predicting crop yield based on artificial intelligence according to an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of the structure of a crop yield prediction system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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 part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a crop yield prediction method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S101. Acquire historical meteorological data, soil data, crop planting data and remote sensing image data of a designated area for consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and water conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data in time series to construct a multidimensional spatiotemporal data cube;

[0065] S102. Using the multidimensional spatiotemporal data cube as a training data set, training a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm;

[0066] S103. Obtain real-time meteorological data, soil data and remote sensing image data of the specified area, input them into a pre-built fusion prediction model, and introduce the agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, infer the crop growth and estimated yield of each grid in the specified area, perform spatial interpolation on the estimated yield of each grid, and generate a yield distribution map of the specified area.

[0067] The present invention obtains data from multiple dimensions such as meteorology, soil, crop planting and remote sensing, and aligns them in time series to construct a multidimensional spatiotemporal data cube, fully exploring the complementary gain relationship between different data sources, thereby significantly improving the accuracy and reliability of yield prediction. The fusion prediction model is constructed based on the deep learning model, which can automatically learn and extract key features in the crop growth process, overcome the dependence of traditional machine learning models on prior knowledge, and introduce an adaptive learning rate optimization algorithm to dynamically adjust the model parameters, so that the model can adapt to the growth characteristics of different regions and different crops, and has stronger robustness and generalization ability.

[0068] Introducing agricultural knowledge graphs into the fusion prediction model constrains and corrects the prediction results. On the one hand, the prediction results are optimized by using the agronomic rules and expert experience in the knowledge graph to improve the accuracy of the prediction; on the other hand, by associating the prediction results with the concepts and entities in the knowledge graph, the prediction results are given clear semantic information, which enhances the interpretability of the prediction process and results, and helps agronomic experts to analyze and verify the model prediction. The crop growth and yield of each grid in the specified area are estimated, and the yield distribution map is generated through spatial interpolation, realizing the yield prediction from the regional scale to the farmland scale. It provides agricultural producers with more refined yield prediction information, which can directly guide the precise management of farmland and the optimal allocation of resources, and has important practical value.

[0069] In an optional embodiment,

[0070] Aligning the historical meteorological data, soil data, crop planting data and remote sensing image data in time series to construct a multidimensional spatiotemporal data cube includes:

[0071] Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data, the daily meteorological grid data including temperature, precipitation and sunshine hours;

[0072] Using the Kriging interpolation method, in combination with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute grid layer consistent with the spatial resolution of the remote sensing image data, wherein the soil attribute grid layer includes soil fertility, soil moisture, temperature and hydraulic conductivity;

[0073] Spatially matching the row and column numbers of the crop planting data with the remote sensing image data, assigning crop variety, sowing date and planting density attributes to each pixel, and estimating missing data using a multiple interpolation method to obtain crop planting attribute data;

[0074] Perform sliding window segmentation on the remote sensing image data on the time axis to extract growth period images corresponding to the crop greening period, jointing period, heading period and maturity period;

[0075] The daily meteorological grid data, the soil attribute grid layer, the crop planting attribute data and the growth period image are constructed with pixels as basic units in the spatial dimension and days or growth periods as units in the temporal dimension to form a multidimensional spatiotemporal data cube integrating time and space.

[0076] In an optional embodiment,

[0077] Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data includes:

[0078] ;

[0079] in, Indicates the meteorological location to be interpolated S 0 Estimated values ​​of meteorological elements at n represents the number of meteorological stations involved in the interpolation, Z(S i ) Indicates i Weather stations S i The meteorological element observation values ​​at d i0 Indicates the position to be interpolated S 0 With i Weather stations S i The Euclidean distance between p represents the distance weighting parameter;

[0080] Using the Kriging interpolation method, combined with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute raster layer consistent with the spatial resolution of the remote sensing image data, including:

[0081] ;

[0082] in, T(L 0 ) Indicates the soil position to be interpolated L 0 Estimated values ​​of meteorological elements at m Indicates the number of soil sampling points involved in interpolation, k j Indicates j The Kriging weight coefficient of each sampling point is determined by the semivariogram and the spatial position relationship of the sampling points. ST(L j ) Indicates j Soil sampling points L j The attribute measurement value at .

[0083] Exemplarily, inverse distance weighted interpolation is performed on historical meteorological data to generate daily meteorological raster data that is consistent with the spatial resolution of the remote sensing image data.

[0084] Obtain daily observation data from ground meteorological stations in and around the study area, including temperature, precipitation and sunshine hours; use the spatial resolution of remote sensing image data (such as 30m) as the target resolution, rasterize the study area, and generate a regular grid network; for each grid, take the center point of the grid as the predicted position, and calculate the weight coefficient of each meteorological station according to its spatial distance from the surrounding meteorological stations, and the weight coefficient is inversely proportional to the distance; use the inverse distance weighted interpolation formula to calculate the temperature, precipitation and sunshine hours of each grid on each day;

[0085] Repeat the above steps until the daily meteorological element values ​​of all rasters in the entire study period are calculated, and a spatiotemporal raster dataset of temperature, precipitation, and sunshine hours that is consistent with the spatial resolution of the remote sensing image data is obtained.

[0086] The Kriging interpolation method is used to spatially interpolate soil data in combination with digital elevation models and land use data to generate a soil attribute raster layer that is consistent with the spatial resolution of remote sensing image data.

[0087] Obtain soil property data from several soil sampling points in the study area, including soil fertility, moisture content, temperature, and hydraulic conductivity; obtain digital elevation model (DEM) and land use type data of the study area, and perform spatial registration with remote sensing image data;

[0088] Using Kriging interpolation theory, a soil property spatial interpolation model considering terrain and land use factors is established. In the interpolation process, DEM and land use data are used as independent variables, and the relationship between soil properties and terrain and land use is established through methods such as co-kriging or regression kriging to improve interpolation accuracy; using the established spatial interpolation model, soil properties are predicted for each 30m×30m grid to obtain soil fertility, soil moisture, temperature and hydraulic conductivity grid layers covering the study area.

[0089] The crop planting data were spatially matched with the row and column numbers of the remote sensing image data, and the crop variety, sowing date and planting density attributes were assigned to each pixel. The missing data were estimated using the multiple interpolation method to obtain the crop planting attribute data.

[0090] Obtain crop planting statistics in the study area, including crop varieties, sowing dates, planting density, etc., and combine them with farmland boundary data to generate crop planting distribution vector maps; perform spatial overlay analysis on crop planting distribution vector maps and remote sensing image data, convert vector maps into raster data with the same row and column numbers as the remote sensing image data, and each pixel corresponds to a crop planting attribute value; for pixels with missing planting information, use multiple interpolation methods to estimate missing values ​​based on the crop planting attributes of surrounding pixels, and commonly used methods include K nearest neighbor method, random forest method, etc.;

[0091] Repeat the above steps until all pixels are assigned crop variety, sowing date and planting density attributes, and obtain a crop planting attribute raster dataset that is spatially aligned with the remote sensing image data.

[0092] The remote sensing image data is segmented by sliding windows on the time axis to extract growth period images corresponding to the greening period, jointing period, heading period and maturity period of crops.

[0093] According to the knowledge of crop phenology in the study area, the time range of crop greening, jointing, heading and maturity is determined; the time range of a key growth period is used as a window to perform sliding segmentation on the time series of remote sensing image data, and all images within the time range are extracted to form a growth period image set;

[0094] Repeat the above steps until all remote sensing image sets of key growth periods are extracted, and remote sensing image sets of four growth periods of crops, namely, greening period, jointing period, heading period and maturity period, are obtained.

[0095] The daily meteorological raster data, soil attribute raster layers, crop planting attribute data and growth period images are taken as the basic units in the spatial dimension and the days or growth periods as the units in the temporal dimension to construct a multidimensional spatiotemporal data cube integrating time and space.

[0096] Taking the pixels of remote sensing images as the basic unit, the daily meteorological raster data, soil attribute raster layer, crop planting attribute data and remote sensing image data are spatially registered and stacked in the spatial dimension to form multiple element layers with consistent spatial dimensions; in the temporal dimension, the daily meteorological data and remote sensing image data are aligned in units of days to form a meteorological element time series and a remote sensing image time series; the remote sensing image sets of key growth periods are aligned in units of growth periods to form a crop growth period time series;

[0097] The obtained multiple spatial element layers and time series are organized and integrated in the space-time dimension to construct a multidimensional data cube integrating space and time. Each pixel corresponds to multiple attribute values ​​at different time points, including meteorological, soil, crop planting and remote sensing attributes.

[0098] Through the above steps, historical meteorological data, soil data, crop planting data and remote sensing image data are integrated in time and space to construct a highly integrated multidimensional spatiotemporal data cube. This data cube fully considers the various factors that affect crop growth and yield formation, and achieves seamless data connection in time and space, laying a solid data foundation for subsequent deep learning model training and yield prediction.

[0099] In an optional embodiment,

[0100] The multidimensional spatiotemporal data cube is used as a training data set to train a fusion prediction model, the fusion prediction model is built based on a deep learning model, the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and the model parameters of the fusion prediction model are dynamically adjusted by an adaptive learning rate optimization algorithm, including:

[0101] The multidimensional spatiotemporal data cube is used as a training data set; a deep learning model is constructed, and a temporal convolutional network structure is adopted to respectively extract the local spatiotemporal correlation features and the global spatiotemporal correlation features of the training data set through a causal convolutional layer and a dilated convolutional layer, and an attention mechanism is introduced at the end of the network of the deep learning model to adaptively adjust the weights of the local spatiotemporal correlation features and the global spatiotemporal correlation features;

[0102] Construct a crop growth model, taking into account meteorological factors, soil properties and crop variety parameters, simulate photosynthesis, respiration, material accumulation and organ formation during crop growth and development, and dynamically adjust the adaptability of crop growth to adverse environmental conditions;

[0103] In the training process of the fusion prediction model, the field is used as the basic unit, and the variable sequence of crop growth status for consecutive years is paired with the image features of the corresponding period to construct a multi-dimensional spatiotemporal feature map with consistent time steps;

[0104] An adaptive learning rate optimization algorithm is used to train the crop growth model and the deep learning model separately. By setting different initial learning rates, learning rate decay methods, and regularization strategies, the training speed and generalization performance of the two sub-models are balanced.

[0105] In each training iteration step, the crop growth model is used to simulate the crop growth and development process, and the state variables of biomass and leaf area index in each growth period are output. They are then spliced ​​with the image features of the same period to obtain a multi-dimensional spatiotemporal feature map; the multi-dimensional spatiotemporal feature map is input into the deep learning model, and end-to-end joint training and parameter optimization are achieved through gradient back propagation.

[0106] Exemplarily, a multidimensional spatiotemporal data cube is used as a training data set to build a deep learning model, and a temporal convolutional network structure is used to extract the spatiotemporal correlation features of the data.

[0107] The multidimensional spatiotemporal data cube is divided into multiple time slices according to the time step. Each time slice contains multiple attribute layers such as meteorology, soil, crop planting and remote sensing, forming a four-dimensional tensor. The dimensions of the tensor are the number of time steps, the number of attribute layers, and the number of rows and columns in the spatial dimension. A deep learning model based on the temporal convolutional network (TCN) is designed. The model consists of multiple causal convolutional layers and dilated convolutional layers stacked alternately. The causal convolutional layer is used to extract the local spatiotemporal correlation features of the data. The causal convolution only considers the data before the current moment in the time dimension, avoiding the leakage of future information.

[0108] The dilated convolution layer is used to extract the global spatiotemporal correlation features of the data. The dilated convolution captures the long-distance spatiotemporal dependencies by expanding the receptive field of the convolution kernel. The attention mechanism is introduced at the end of the network to adaptively adjust the weights of local and global features. The weight parameters and bias terms of the attention layer are used to perform element-by-element multiplication to weightedly fuse the local and global features.

[0109] Construct a crop growth model to simulate the key physiological processes during crop growth and development.

[0110] Select a process model suitable for the target crop, such as WOFOST, CERES, etc., as the basis of the crop growth model; consider the impact of key meteorological factors (temperature, radiation, precipitation, etc.), soil properties (water, nutrients, etc.) and crop variety parameters (development rate, distribution coefficient, etc.) on crop growth in the crop growth model; simulate the key physiological processes in the growth and development of crops, including photosynthesis, respiration, dry matter accumulation and organ formation, and calculate the relationship between net photosynthetic rate, maintenance respiration rate and growth respiration rate through the rate of change of crop dry matter weight; dynamically adjust the adaptability of crop growth to adverse environmental conditions (high temperature, drought, etc.) by introducing environmental stress factors, and adjust the maximum photosynthetic rate according to the stress factor functions of temperature, water and nitrogen.

[0111] In the training process of the fusion prediction model, the field is taken as the basic unit to construct a multi-dimensional spatiotemporal feature map with consistent time steps.

[0112] The study area is divided into several plots, each of which contains multiple pixels to ensure that the soil properties and crop planting conditions in the same plot are relatively consistent; for each plot, a sequence of crop growth state variables (such as biomass, leaf area index, etc.) for consecutive years is extracted, and paired with the remote sensing image features of the corresponding period in time and space to form a multidimensional spatiotemporal feature sequence with consistent time steps; the multidimensional spatiotemporal feature sequence of each plot is arranged and combined according to year and growing period to construct a five-dimensional tensor, and the dimensions of the tensor are the number of plots, the number of years, the number of growing periods, the number of multidimensional elements, and the time step.

[0113] An adaptive learning rate optimization algorithm is used to train the crop growth model and the deep learning model separately to balance the training speed and generalization performance of the two sub-models.

[0114] Initialize the weight parameters of the crop growth model and the deep learning model, set different initial learning rates, learning rate decay methods and regularization strategies; use the adaptive learning rate optimization algorithm to train the two sub-models separately, and dynamically adjust the size and change trend of the learning rate according to the performance of the model on the validation set; in the training process of the crop growth model, set a smaller initial learning rate and a faster learning rate decay rate to ensure that the model converges smoothly under physical constraints and avoids overfitting; in the training process of the deep learning model, set a larger initial learning rate and a slower learning rate decay rate to accelerate the model's learning of complex spatiotemporal patterns and improve generalization performance; introduce L1 and L2 regularization terms to control the parameter scale of the two sub-models and reduce the risk of overfitting.

[0115] In each training iteration, the crop growth model is used to simulate the crop growth and development process, and end-to-end joint training and parameter optimization are performed with the deep learning model.

[0116] In each training iteration step, the crop growth model is first used to simulate the crop growth and development process, and state variables such as biomass and leaf area index in each growth period are output; the state variables output by the crop growth model are spliced ​​with the remote sensing image features of the same period to obtain a multi-dimensional spatiotemporal feature map; the multi-dimensional spatiotemporal feature map is input into the deep learning model, and after multiple layers of temporal convolution and attention operations, high-level semantic features are extracted, and a fully connected layer and a linear regression layer are connected at the end of the network to predict crop yield; the prediction loss is calculated using the real yield label, and the mean square error is used as the loss function; the physiological process loss is calculated using the simulated value of the crop growth model, and the relative error is used as the loss function to guide the model to follow the law of crop growth and development; the prediction loss and the physiological process loss are weighted and summed to obtain the total training objective function, and the relative importance of the two parts of the loss is controlled by the weight coefficient; the gradient of the total loss to the model parameters is calculated using the adaptive learning rate optimization algorithm, and the joint training and parameter update of the crop growth model and the deep learning model are realized through gradient back propagation.

[0117] The above steps are iteratively executed to achieve end-to-end joint training of crop growth models and deep learning models. In each iteration, the crop growth model is first used to simulate the crop growth and development process, output physiological state variables, and splice them with remote sensing image features to form a multi-dimensional spatiotemporal feature map; then the multi-dimensional spatiotemporal feature map is input into the deep learning model to extract spatiotemporal semantic features and predict yield; finally, the prediction loss and physiological process loss are calculated, and the parameters of the two sub-models are optimized through gradient back propagation. This joint training paradigm can make full use of the complementary advantages of crop growth models and deep learning models, taking into account both the physical mechanism of crop growth and development and the complex patterns contained in spatiotemporal big data, thereby significantly improving the accuracy and reliability of crop yield prediction.

[0118] In an optional embodiment,

[0119] The agricultural knowledge graph is introduced into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, and the crop growth and estimated yield of each grid in the specified area are inferred. The estimated yield of each grid is spatially interpolated to generate the yield distribution map of the specified area, including:

[0120] Construct an agricultural knowledge graph, extract agricultural concepts and semantic relationships from unstructured texts through natural language processing technology, design an agricultural ontology model, and use ontology construction tools to create the basic framework of the agricultural knowledge graph;

[0121] Based on the constructed agricultural knowledge graph, a rule-based reasoning method is used to mine the constraint rules implicit in the agricultural knowledge graph. According to the crop growth cycle, phenological period division, and the relationship between agricultural operations and yield, a series of constraint rule sets covering different constraint dimensions are constructed.

[0122] Using the trained fusion prediction model, the yield prediction value of each grid in the specified area is obtained, the specified area is divided into several grid units of equal size, each grid is assigned a unique number, the entities in the agricultural knowledge graph are spatially matched with the corresponding grids, and a mapping table between entities and grids is established to realize the association between grid-level prediction results and constraint rules;

[0123] For each grid, retrieve the entities and their attribute values ​​related to the grid from the agricultural knowledge graph, match the attribute values ​​with the constraint rules, and when the attribute values ​​meet the condition part of the constraint rules, trigger the conclusion part of the rule to correct the prediction results of the grid. By applying the constraint rules one by one, a systematic correction of the prediction results of the entire specified area can be achieved.

[0124] The spatial interpolation method is used to spatially interpolate the corrected prediction results, and the GIS tool is used to generate the yield distribution map corresponding to the interpolated corrected prediction results.

[0125] Exemplarily, an agricultural knowledge graph is constructed, agricultural concepts and semantic relations are extracted, an ontology model is designed, and a knowledge graph framework is created.

[0126] Collect unstructured text data related to the agricultural field, such as agricultural popular science articles, expert interview records, pest and disease control manuals, etc., to build an agricultural corpus; use natural language processing technology to pre-process the agricultural corpus, including word segmentation, part-of-speech tagging, named entity recognition, etc., to extract key concepts and entities in the agricultural field; use semantic role labeling, dependency syntax analysis and other technologies to identify the semantic relationship between agricultural concepts, such as "crop-planting time", "pest and disease-control methods", etc.;

[0127] Design an agricultural ontology model, define the core concepts, attributes and relationships in the agricultural field, such as "crops", "soil", "weather", "planting", "harvest", etc., and build a basic framework of agricultural knowledge; use ontology construction tools to import the extracted agricultural concepts, entities and semantic relationships into the ontology model to form a structured and semantic agricultural knowledge graph.

[0128] Based on the agricultural knowledge graph, a rule-based reasoning method is used to mine implicit constraint rules.

[0129] Analyze the entities, attributes and relationships in the agricultural knowledge graph, summarize the empirical laws and causal relationships in the agricultural production process, such as "sowing time-growth process-maturity period", "irrigation frequency-soil moisture-yield", etc.; define the condition part and conclusion part of a series of constraint rules based on the knowledge of agricultural experts. The condition part describes the prerequisite for the triggering of the rule, and the conclusion part describes the correction method for the prediction result after the rule is triggered;

[0130] The constraint rules are classified and organized according to applicable objects, constraint dimensions, etc. to form a hierarchical and modular rule base to facilitate the management, retrieval and updating of rules; using rule engines such as Drools, the constraint rules are converted into executable reasoning rules to realize rule mining and reasoning of agricultural knowledge graphs.

[0131] The fusion prediction model is used to obtain the grid-level yield prediction value, and the entities in the agricultural knowledge graph are spatially matched with the corresponding grids.

[0132] The specified area is divided into grid units of equal size, each grid is assigned a unique number, and a mapping relationship between the grid number and the geographic coordinates is established; using the trained fusion prediction model, the multi-source remote sensing data, meteorological data, soil data, etc. of each grid in the specified area are input to obtain the grid-level yield prediction value; according to the geographic attributes of the entities in the agricultural knowledge graph, such as land location, administrative divisions, etc., the entities are spatially matched with the corresponding grids, and a mapping table between entities and grids is established; through the entity-grid mapping table, the association between the raster-level prediction results and the constraint rules in the agricultural knowledge graph is realized, providing a basis for the subsequent correction of the prediction results.

[0133] The prediction results of each grid are systematically corrected and constraint rules are applied to correct the predicted values.

[0134] Traverse each grid in the specified area, retrieve the entity and its attribute value related to the grid from the entity-grid mapping table according to the grid number; match the retrieved attribute value with the rule conditions in the constraint rule library to determine whether the triggering condition of the rule is met; when the attribute value meets the condition part of the constraint rule, trigger the conclusion part of the rule, and correct the prediction result of the grid according to the correction method defined in the conclusion part, such as linear scaling, threshold truncation, smoothing, etc.; apply the constraint rules one by one according to the priority and scope of application of the rules to achieve systematic correction of all grid prediction results in the entire specified area; reorganize the corrected raster prediction results into a rule matrix to prepare for subsequent spatial interpolation processing.

[0135] The spatial interpolation method is used to process the corrected prediction results to generate a yield distribution map.

[0136] According to the geographical characteristics and grid size of the designated area, select appropriate spatial interpolation methods, such as inverse distance weighted method, Kriging method, spline function method, etc.; use the coordinates of the grid center point and the corrected prediction value as the sample point, use the selected interpolation method to perform spatial interpolation on the entire designated area, and estimate the yield value of the unknown position; use cross-validation and other methods to evaluate the accuracy and reliability of the interpolation results, adjust the interpolation parameters or change the interpolation method when necessary; use GIS tools to visualize the interpolation results and generate thematic maps reflecting the yield distribution characteristics of the designated area, such as graded color maps, contour maps, etc.; overlay the yield distribution map with remote sensing images, geographic base maps, etc., analyze the relationship between yield distribution and factors such as natural geography and farming system, and provide an intuitive and explainable reference for agricultural management decisions.

[0137] By integrating agricultural knowledge graphs with deep learning models, this solution realizes the organic combination of prior knowledge in the agricultural field and data-driven models. The agricultural knowledge graph provides structured and semantic agricultural rules to guide the model to learn the empirical laws in the agricultural production process; the deep learning model extracts the key features of crop growth and yield formation from massive spatiotemporal data to reveal the inherent patterns of yield distribution. The complementary integration of the two makes the prediction results of the model more in line with the actual agricultural production, the prediction error is effectively controlled, and the applicability and interpretability of the solution are significantly enhanced. At the same time, the visual display of the yield distribution map provides an intuitive and operational decision-making auxiliary tool for agricultural authorities, which promotes the transformation and implementation of the model application results.

[0138] In an optional embodiment,

[0139] For each grid, the entities and their attribute values ​​related to the grid are retrieved from the agricultural knowledge graph, and the attribute values ​​are matched with the constraint rules. When the attribute values ​​meet the condition part of the constraint rules, the conclusion part of the rule is triggered to correct the prediction results of the grid. By applying the constraint rules one by one, the systematic correction of the prediction results of the entire specified area is achieved, including:

[0140] Preprocess the constraint rules corresponding to the agricultural knowledge graph, convert the condition part of the constraint rules into the form of attribute value range, and convert the conclusion part into the form of correction coefficient to achieve quantitative representation of the constraint rules;

[0141] For each grid, retrieve the entity corresponding to the grid from the agricultural knowledge graph, and obtain the attributes and attribute values ​​associated with the entity to form a grid-level attribute value vector;

[0142] Index the preprocessed constraint rules, use the condition part of the constraint rules as the index key, use the attribute type as the dimension of the rule tree, and use the attribute value range as the node of the rule tree to build a multi-dimensional rule tree;

[0143] Use attribute value vectors and rule trees for fast matching, adopt a priority-based matching strategy, assign priority weights according to the confidence and specificity of constraint rules, give priority to matching constraint rules with priority weights greater than the preset weight threshold, and find all constraint rules that meet the conditions;

[0144] For the successfully matched constraint rules, their conclusion part is triggered, and the yield prediction value of the grid is corrected according to the correction coefficient of the constraint rule. When multiple constraint rules are met in the same grid, the weighted average method is adopted to comprehensively consider the correction coefficient and priority weight of each constraint rule, calculate the final correction coefficient, and realize the coordinated correction of multiple constraint rules.

[0145] Exemplarily, the constraint rules in the agricultural knowledge graph are preprocessed to achieve quantitative representation of the rules.

[0146] The condition part of the constraint rule is converted into the form of attribute value range, and the attribute type and threshold involved in the condition are extracted to form a triple of "attribute type-upper threshold-lower threshold" for subsequent rule matching; the conclusion part of the constraint rule is converted into the form of correction coefficient, and a correction coefficient between 0 and 1 is designed according to the correction method and correction strength described in the conclusion. The larger the correction coefficient, the greater the correction range of the prediction result;

[0147] For condition descriptions that cannot be directly quantified, such as qualitative language expressions, text mining technology is used to extract keywords, and corresponding quantification standards are formulated based on the common usage of keywords in the agricultural field to convert qualitative descriptions into quantitative attribute value ranges; the quantified constraint rules are regularized to unify the value ranges and measurement units of attribute values ​​and correction coefficients to facilitate subsequent rule matching and correction of prediction results.

[0148] For each grid, the agricultural knowledge graph is retrieved to obtain the attribute value vector at the grid level.

[0149] According to the geographical location and boundary range of the grid, the basic attributes such as the administrative division, land use type, soil type, etc. to which the grid belongs are determined; in the agricultural knowledge graph, with the grid as the center, geographic spatial indexing technology is used to quickly retrieve agricultural entities that overlap or are adjacent to the grid location, such as land, crops, farming activities, etc.; for each retrieved agricultural entity, its associated attribute information is obtained, and the specific values ​​of the numerical attributes are extracted to form an attribute value vector, in which each element of the vector corresponds to an attribute type, and the value of the element is the specific value of the attribute; the attribute value vector is normalized to eliminate the differences in different attribute dimensions and numerical ranges, thereby improving the accuracy and efficiency of subsequent matching.

[0150] Index the preprocessed constraint rules and build a multi-dimensional rule tree.

[0151] Taking the condition part of the constraint rule as the index key, extract all the attribute types involved in the condition as the dimensions of the rule tree, and each dimension corresponds to an attribute type; based on the range of the attribute value in the condition, divide each dimension into several intervals, and each interval corresponds to an attribute value range, forming a multi-level tree structure; insert the constraint rules into the corresponding tree nodes according to the attribute types and attribute value ranges involved in their condition parts, so as to realize the distributed storage and fast retrieval of the rules in the tree; prune and optimize the rule tree, merge similar attribute value ranges, reduce the depth of the tree and the number of nodes, and improve the matching efficiency.

[0152] Use the attribute value vector to quickly match the rule tree to find all constraint rules that meet the conditions.

[0153] For each grid, its attribute value vector is matched with the rule tree, and the depth-first search algorithm is used to quickly locate the matching leaf node along the trunk and branches of the tree; in the matching process, a priority-based matching strategy is adopted, and the priority weight is calculated according to the confidence and specificity of the constraint rule, and the rules with weights greater than the preset threshold are matched first to reduce the number of matches; for each matched rule, the attribute value range in the rule condition is compared with the value of the corresponding attribute in the grid attribute value vector to determine whether the triggering condition of the rule is met; all constraint rules that meet the conditions are screened out to form a candidate rule set, which provides a basis for subsequent correction of prediction results.

[0154] The successfully matched constraint rules are triggered to coordinately correct the prediction results of the raster.

[0155] For each constraint rule in the candidate rule set, extract the correction coefficient corresponding to its conclusion part, multiply it by the priority weight of the rule, and get the weighted correction coefficient; weighted average the weighted correction coefficients of all rules to get the comprehensive correction coefficient, and the value of the weight is proportional to the confidence and specificity of the rule; use the comprehensive correction coefficient to correct the yield prediction value of the grid, and adjust the prediction value to a reasonable range through multiplication or addition operation; rewrite the corrected grid prediction value into the grid attribute table, update the yield attribute value of the grid, and complete the correction of the prediction result of a single grid; repeat the above steps for all grids in the specified area to achieve systematic correction of the prediction results of the entire area and improve the accuracy and reliability of the yield prediction.

[0156] By preprocessing and quantifying the constraint rules in the agricultural knowledge graph, the qualitative rule description is converted into a quantitative attribute value range and correction coefficient, providing a numerical judgment basis for rule matching and triggering. The rule tree index and priority-based matching strategy are adopted to achieve efficient and accurate rule matching, reduce rule conflicts and computational overhead. When multiple constraint rules are triggered at the same time, the correction opinions of different rules are integrated by weighted average to achieve collaborative correction of constraint rules.

[0157] In another embodiment of the present application, the present application takes eggplant crops as the research object, obtains regional climate data, indoor environment data, soil data, crop growth data and water and fertilizer regulation data for consecutive years under different greenhouse conditions and different crops in a specified area, and dynamically adjusts the model parameters of the fusion prediction model through the adaptive learning rate optimization algorithm through the accumulated IoT sensor collection data, intelligent control equipment collection data and growth statistics in multiple crops; obtains the greenhouse planting conditions, crops, real-time meteorological data, indoor environment data, water and fertilizer regulation data and growth statistics in the specified area, and inputs them into a pre-built fusion prediction model, and introduces an agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, and infers the predicted maturity period and output.

[0158] Specifically, obtain climate data of the specified area, including temperature, humidity, precipitation, light intensity, etc. The data source can be a meteorological station, satellite remote sensing or regional weather forecast system. Collect indoor environmental data under different greenhouse conditions, including temperature, humidity, CO2 concentration, light intensity, etc. The data source is mainly the IoT sensors installed in the greenhouse. Collect soil data of different crops, including soil temperature, humidity, pH value, nutrient content, etc. The data source is soil sensors and laboratory test results.

[0159] Through IoT sensors and manual statistics, we can obtain the growth data of eggplant and tomato crops, such as growth cycle, leaf area index, plant height, etc. We can collect water and fertilizer control data of different crops, including irrigation amount, fertilizer amount, fertilization frequency, etc. The data comes from the intelligent irrigation system and fertilization system. We can preprocess the collected multi-source data, including data cleaning, missing value filling, outlier detection and normalization. We can integrate the preprocessed regional climate data, indoor environment data, soil data, crop growth data and water and fertilizer control data to construct a comprehensive data set.

[0160] Based on deep learning and machine learning algorithms, a fusion prediction model is constructed. The model input is multi-source fusion data, and the output is the maturity period and output of eggplant crops. The fusion prediction model is optimized using an adaptive learning rate optimization algorithm. By dynamically adjusting the learning rate of the model, the convergence speed and prediction accuracy of the model are improved.

[0161] Divide the historical data set into a training set and a validation set to train the fusion prediction model. Use the adaptive learning rate optimization algorithm to dynamically adjust the model parameters to ensure the stability and convergence of the model. Use the validation set to verify the trained model and evaluate the prediction performance of the model. If the model performance does not meet the requirements, adjust the model structure or optimize the algorithm and retrain the model.

[0162] Obtain greenhouse planting conditions, crop rotation, real-time meteorological data, indoor environmental data, water and fertilizer control data, and growth statistics of the specified area, and input them into the pre-built fusion prediction model. Introduce the agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the model. The agricultural knowledge graph contains the growth laws, environmental influencing factors, and their interrelationships of eggplant and tomato crops. Through the correction of the fusion prediction model and the agricultural knowledge graph, the predicted maturity period and output of eggplant and tomato crops are inferred. According to the prediction results, precise management and regulation are carried out to improve crop yield and quality.

[0163] Figure 2 FIG. 1 is a schematic diagram of the structure of a crop yield prediction system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0164] The first unit is used to obtain historical meteorological data, soil data, crop planting data and remote sensing image data of a designated area for many consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and water conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data according to time series to construct a multidimensional spatiotemporal data cube;

[0165] The second unit is used to use the multidimensional spatiotemporal data cube as a training data set to train a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, and the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm;

[0166] The third unit is used to obtain real-time meteorological data, soil data and remote sensing image data of the specified area, input them into a pre-built fusion prediction model, and introduce the agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, infer the crop growth and estimated yield of each grid in the specified area, perform spatial interpolation on the estimated yield of each grid, and generate a yield distribution map of the specified area.

[0167] According to a third aspect of the embodiments of the present invention,

[0168] An electronic device is provided, comprising:

[0169] processor;

[0170] a memory for storing processor-executable instructions;

[0171] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0172] According to a fourth aspect of the embodiments of the present invention,

[0173] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0174] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0175] Finally, it should be noted that the above embodiments 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crop yield prediction method based on artificial intelligence, characterized in that: include: Acquire historical meteorological data, soil data, crop planting data and remote sensing image data of the designated area for many consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and hydraulic conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data in time series to construct a multidimensional spatiotemporal data cube; The multidimensional spatiotemporal data cube is used as a training data set to train a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, and the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm; The multidimensional spatiotemporal data cube is used as a training data set; a deep learning model is constructed, and a temporal convolutional network structure is adopted to respectively extract the local spatiotemporal correlation features and the global spatiotemporal correlation features of the training data set through a causal convolutional layer and a dilated convolutional layer, and an attention mechanism is introduced at the end of the network of the deep learning model to adaptively adjust the weights of the local spatiotemporal correlation features and the global spatiotemporal correlation features; Construct a crop growth model, taking into account meteorological factors, soil properties and crop variety parameters, simulate photosynthesis, respiration, material accumulation and organ formation during crop growth and development, and dynamically adjust the adaptability of crop growth to adverse environmental conditions; In the training process of the fusion prediction model, the field is used as the basic unit, and the variable sequence of crop growth status for consecutive years is paired with the image features of the corresponding period to construct a multi-dimensional spatiotemporal feature map with consistent time steps; An adaptive learning rate optimization algorithm is used to train the crop growth model and the deep learning model separately. By setting different initial learning rates, learning rate decay methods, and regularization strategies, the training speed and generalization performance of the two sub-models are balanced. In each training iteration, the crop growth model is used to simulate the crop growth and development process, and the state variables of biomass and leaf area index in each growth period are output, which are then spliced ​​with the image features of the same period to obtain a multi-dimensional spatiotemporal feature map. Input the multi-dimensional spatiotemporal feature map into the deep learning model, and achieve end-to-end joint training and parameter optimization through gradient back propagation; Real-time meteorological data, soil data and remote sensing image data of the specified area are obtained, and input into a pre-built fusion prediction model. The agricultural knowledge graph is introduced into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model. The crop growth and estimated yield of each grid in the specified area are inferred, and the estimated yield of each grid is spatially interpolated to generate a yield distribution map of the specified area.

2. The method according to claim 1, characterized in that The historical meteorological data, soil data, crop planting data and remote sensing image data are aligned in time series to construct a multidimensional spatiotemporal data cube including: Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data, the daily meteorological grid data including temperature, precipitation and sunshine hours; Using the Kriging interpolation method, in combination with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute grid layer consistent with the spatial resolution of the remote sensing image data, wherein the soil attribute grid layer includes soil fertility, soil moisture, temperature and hydraulic conductivity; Spatially matching the row and column numbers of the crop planting data with the remote sensing image data, assigning crop variety, sowing date and planting density attributes to each pixel, and estimating missing data using a multiple interpolation method to obtain crop planting attribute data; Perform sliding window segmentation on the remote sensing image data on the time axis to extract growth period images corresponding to the crop greening period, jointing period, heading period and maturity period; The daily meteorological grid data, the soil attribute grid layer, the crop planting attribute data and the growth period image are constructed with pixels as basic units in the spatial dimension and days or growth periods as units in the temporal dimension to form a multidimensional spatiotemporal data cube integrating time and space.

3. The method according to claim 2, characterized in that Performing inverse distance weighted interpolation on the historical meteorological data to generate daily meteorological grid data consistent with the spatial resolution of the remote sensing image data includes: ； in, Indicates the meteorological location to be interpolated S 0 Estimated values ​​of meteorological elements at n represents the number of meteorological stations involved in the interpolation, Z(S i ) Indicates i Weather stations S i The meteorological element observation values ​​at d i0 Indicates the position to be interpolated S 0 With i Weather stations S i The Euclidean distance between p represents the distance weighting parameter; Using the Kriging interpolation method, combined with the digital elevation model and land use data, the soil data is spatially interpolated to generate a soil attribute raster layer consistent with the spatial resolution of the remote sensing image data, including: ； in, T(L 0 ) Indicates the soil position to be interpolated L 0 Estimated values ​​of meteorological elements at m Indicates the number of soil sampling points involved in interpolation, k j Indicates j The Kriging weight coefficient of each sampling point is determined by the semivariogram and the spatial position relationship of the sampling points. ST(L j ) Indicates j Soil sampling points L j The measured value of the attribute at .

4. The method according to claim 1, characterized in that: The agricultural knowledge graph is introduced into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, and the crop growth and estimated yield of each grid in the specified area are inferred. The estimated yield of each grid is spatially interpolated to generate the yield distribution map of the specified area, including: Construct an agricultural knowledge graph, extract agricultural concepts and semantic relationships from unstructured texts through natural language processing technology, design an agricultural ontology model, and use ontology construction tools to create the basic framework of the agricultural knowledge graph; Based on the constructed agricultural knowledge graph, a rule-based reasoning method is used to mine the constraint rules implicit in the agricultural knowledge graph. According to the crop growth cycle, phenological period division, and the relationship between agricultural operations and yield, a series of constraint rule sets covering different constraint dimensions are constructed. Using the trained fusion prediction model, the yield prediction value of each grid in the specified area is obtained, the specified area is divided into several grid units of equal size, each grid is assigned a unique number, the entities in the agricultural knowledge graph are spatially matched with the corresponding grids, and a mapping table between entities and grids is established to realize the association between grid-level prediction results and constraint rules; For each grid, retrieve the entities and their attribute values ​​related to the grid from the agricultural knowledge graph, match the attribute values ​​with the constraint rules, and when the attribute values ​​meet the condition part of the constraint rules, trigger the conclusion part of the rule to correct the prediction results of the grid. By applying the constraint rules one by one, a systematic correction of the prediction results of the entire specified area can be achieved. The spatial interpolation method is used to spatially interpolate the corrected prediction results, and the GIS tool is used to generate the yield distribution map corresponding to the interpolated corrected prediction results.

5. The method according to claim 4, characterized in that For each grid, the entities and their attribute values ​​related to the grid are retrieved from the agricultural knowledge graph, and the attribute values ​​are matched with the constraint rules. When the attribute values ​​meet the condition part of the constraint rules, the conclusion part of the rule is triggered to correct the prediction results of the grid. By applying the constraint rules one by one, the systematic correction of the prediction results of the entire specified area is achieved, including: Preprocess the constraint rules corresponding to the agricultural knowledge graph, convert the condition part of the constraint rules into the form of attribute value range, and convert the conclusion part into the form of correction coefficient to achieve quantitative representation of the constraint rules; For each grid, retrieve the entity corresponding to the grid from the agricultural knowledge graph, and obtain the attributes and attribute values ​​associated with the entity to form a grid-level attribute value vector; Index the preprocessed constraint rules, use the condition part of the constraint rules as the index key, use the attribute type as the dimension of the rule tree, and use the attribute value range as the node of the rule tree to build a multi-dimensional rule tree; Use attribute value vectors and rule trees for fast matching, adopt a priority-based matching strategy, assign priority weights according to the confidence and specificity of constraint rules, give priority to matching constraint rules with priority weights greater than the preset weight threshold, and find all constraint rules that meet the conditions; For the successfully matched constraint rules, their conclusion part is triggered, and the yield prediction value of the grid is corrected according to the correction coefficient of the constraint rule. When multiple constraint rules are met in the same grid, the weighted average method is adopted to comprehensively consider the correction coefficient and priority weight of each constraint rule, calculate the final correction coefficient, and realize the coordinated correction of multiple constraint rules.

6. A crop yield prediction system based on artificial intelligence, used to implement the method according to any one of claims 1 to 5, characterized in that: include: The first unit is used to obtain historical meteorological data, soil data, crop planting data and remote sensing image data of a designated area for many consecutive years, wherein the historical meteorological data includes daily temperature, precipitation, sunshine hours and wind speed; the soil data includes soil fertility, soil moisture, temperature and water conductivity; the crop planting data includes crop variety, sowing date, planting density, irrigation amount and fertilizer amount; the remote sensing image data includes satellite remote sensing images and color images taken by drones; align the historical meteorological data, soil data, crop planting data and remote sensing image data according to time series to construct a multidimensional spatiotemporal data cube; The second unit is used to use the multidimensional spatiotemporal data cube as a training data set to train a fusion prediction model, wherein the fusion prediction model is constructed based on a deep learning model, and the fusion prediction model is used to extract growth feature vectors of crop growth and development processes and perform yield prediction, and dynamically adjust model parameters of the fusion prediction model through an adaptive learning rate optimization algorithm; The third unit is used to obtain real-time meteorological data, soil data and remote sensing image data of the specified area, input them into a pre-built fusion prediction model, and introduce the agricultural knowledge graph into the fusion prediction model to constrain and correct the prediction results of the fusion prediction model, infer the crop growth and estimated yield of each grid in the specified area, perform spatial interpolation on the estimated yield of each grid, and generate a yield distribution map of the specified area.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images