A big data-based rapeseed oil content prediction method

Abstract

The application relates to the rapeseed oil content prediction technical field, and provides a rapeseed oil content prediction method based on big data, step 1: high-resolution satellite or unmanned aerial vehicle remote sensing images of a target area in a rapeseed full flowering period are acquired; a semantic segmentation algorithm is applied to process the images, and the spatial distribution map of a planting area is accurately extracted; based on the rapeseed flower spectral features and a phenology model in the images, the full flowering period time of each extracted planting area is identified and recorded; the application first realizes the large-scale and regional prediction of the core quality index of rapeseed kernels, i.e. oil content. By fusing the key meteorological stress indexes (such as effective accumulated temperature and diurnal temperature difference) in the full flowering period and the boll period, the yield potential signal based on the previous climate and the total yield of the current year and other multi-dimensional information, a prediction model reflecting the complex mechanism of rapeseed kernel oil formation is constructed, and the core defect that the existing satellite remote sensing technology can only predict the total yield but cannot obtain the oil content is effectively solved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method for predicting the oil content of rapeseed based on big data. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] Oilseed processing enterprises (whose main product is rapeseed oil) rely heavily on forecasting the supply and demand of raw rapeseed and finished rapeseed oil in their business decisions. The current common practice is to estimate the rapeseed planting area using satellite remote sensing imagery before the harvest season and combine this with historical yield data to predict the total rapeseed output for the year. Based on this output forecast, enterprises further infer the market supply of rapeseed oil, thereby predicting rapeseed oil price trends and planning subsequent rapeseed procurement fund allocation and strategies accordingly.

[0004] However, this total output-based forecasting method has significant limitations: it only reflects the "quantity" of rapeseed oil (or corresponding rapeseed) circulating in the market, and cannot obtain its core quality indicator—oil content. For large oilseed enterprises, the actual oil content (oil yield) of rapeseed is the key factor determining production costs. For example, if the average oil content of rapeseed decreases by 1% compared to previous years, it means that the enterprise needs to purchase and process approximately 1% more rapeseed to obtain the same amount of rapeseed oil. This will directly lead to:

[0005] Rising raw material procurement costs: More tons of rapeseed need to be purchased.

[0006] Increased operating costs: Significantly increases the cost burden of subsequent storage, transportation, and processing.

[0007] Therefore, relying solely on total production forecasts while ignoring fluctuations in oil content makes it difficult for companies to accurately estimate actual production costs, leading to significant deviations in cost budgets and severely impacting the optimization of procurement strategies and overall operational efficiency. Summary of the Invention

[0008] In view of this, the purpose of this application is to provide a method for predicting the oil content of rapeseed based on big data. The method for predicting the oil content of rapeseed based on big data disclosed in this application can achieve this.

[0009] The objective of this application is achieved through the following technical solution:

[0010] A method for predicting rapeseed oil content based on big data includes the following steps:

[0011] Step 1: Acquire high-resolution satellite or UAV remote sensing images of the target area during the rapeseed flowering period; apply semantic segmentation algorithms to process the images and accurately extract the spatial distribution map of the planting area; based on the spectral characteristics of rapeseed flowers and phenological models in the images, identify and record the flowering period time of each extracted planting area;

[0012] Step 2: Obtain high spatiotemporal resolution meteorological data of each planting area within the target region during the key growth stages before and after the full bloom period; perform spatiotemporal matching and fusion of the full bloom period time and corresponding temperature data of each area; based on the physiological and ecological research of rapeseed, calculate meteorological indicators related to oil content, and generate environmental stress / suitability information that characterizes the formation of oil content.

[0013] Step 3: Obtain historical climate data (such as precipitation, sunshine hours, accumulated temperature, etc.) of the target area during the early rapeseed growing season (from sowing to bolting); construct and train a support vector machine regression model, which uses the previous climate data as input to learn the nonlinear relationship between the data and historical rapeseed yield; use the trained support vector machine regression model to process the previous climate data of the target area in the current year and output information representing the yield potential predicted based on the previous climate conditions.

[0014] Step 4: Collect data on rapeseed planting area and corresponding actual rapeseed oil production in the target area over the years to form a time series sample; construct and train a long short-term memory network model, which learns the temporal dependency between historical planting area and rapeseed oil production; input the rapeseed planting area extracted in Step 1 into the trained LSTM model to predict the total expected rapeseed production for the year.

[0015] Step 5: Integrate environmental stress / suitability information, yield potential information, and the expected total rapeseed yield at the feature level to form a multi-dimensional fused feature vector that comprehensively reflects the rapeseed growth status, environmental conditions, and yield expectations; construct and train a convolutional neural network model, which takes the fused feature vector as input and learns the complex mapping relationship between it and the average oil content of rapeseed in the target area; use the trained CNN model to process the fused feature vector of the current year, and finally output the predicted value of the average oil content of rapeseed in the target area.

[0016] Furthermore, step 1 includes the following steps:

[0017] Step 11: Obtain the flowering period range of the target area, acquire remote sensing image data within the flowering period range, and extract the infrared spectrum of the remote sensing image data to generate a spectral image;

[0018] Step 12: Use the FMask algorithm to extract the cloud region from the spectral image, obtain multiple spectral images, and fuse the spectral images with the cloud region removed to generate an upper-air image of the target area without cloud cover.

[0019] Step 13: Extract the rapeseed planting area from the aerial image based on the semantic segmentation model and generate a spatial distribution map of the target area;

[0020] Step 14: Generate the peak flowering time for each planting area based on the time when the rapeseed flower spectrum begins to appear in each planting area.

[0021] The technical solution provided in this application determines the rapeseed flowering period not based on data from meteorological or agricultural bureaus, but by whether the rapeseed flowers are in bloom. Therefore, in practice, it can accurately obtain the rapeseed flowering period for each planting field, avoiding the impact of broad statistical flowering periods on prediction accuracy.

[0022] Furthermore, step 12 includes the following steps:

[0023] Step 121: Detect the reflectance of each pixel in the spectral image at each wavelength.

[0024] Step 122: Set threshold ranges for the visible light band and the short-wave infrared band respectively; if the reflectance of a pixel in the visible light band is within the set threshold range, and the reflectance of a pixel in the short-wave infrared band is also within the threshold range, then the pixel is a strongly extracted edge.

[0025] If a pixel's reflectance is within the set threshold range only in the visible light band or in the short-wave infrared band, then the pixel is set as a weakly extracted edge.

[0026] Step 123: Connect weak extraction edges. If a weak extraction edge is connected to a strong extraction edge, then set the weak extraction edge as a strong extraction edge and treat all strong extraction edges as layer regions.

[0027] The technical solution provided in this application determines whether a region is cloud-covered based on reflectivity with high accuracy. Furthermore, different reflectivity thresholds are set for different wavelengths. Therefore, it effectively reduces judgment errors and increases the accuracy of cloud detection. This allows for the acquisition of accurate upper-air images of all locations within the target area.

[0028] Furthermore, semantic segmentation models include:

[0029] Input layer: used to input unobstructed high-altitude images, where each pixel in the high-altitude image data includes the reflectance spectrum of the corresponding region;

[0030] Encoding layer: High-altitude semantic features are extracted from the aerial images using a 4-level downsampling method; the 4-level downsampling includes two convolutional kernels and two max pooling operations;

[0031] Decoding layer: Performs 4 levels of upsampling on high-level semantic features to recover the spatial features of high-level semantic features; the 4 levels of upsampling include 2 transposed convolutions and 2 ordinary convolutions;

[0032] The output layer generates the probability of whether each pixel is a planting area based on the spatial features of high-level semantic features, and generates a spatial distribution map of the target area based on the probability of whether it is a planting area.

[0033] The technical solution provided in this application incorporates spectral information into aerial imagery data. Since rapeseed flowers are yellow when in bloom, exhibiting a very distinct yellow spectrum, using spectral information as a supplement to aerial imagery can increase the accuracy of planting area extraction. Furthermore, the encoding and decoding layers perform downsampling and upsampling respectively, effectively filtering feature information during this process to accurately generate a spatial distribution map of the target area.

[0034] Furthermore, the loss function of the semantic segmentation model for:

[0035] ;

[0036] ;

[0037] ;

[0038] in, For binary cross-entropy loss, For Dice loss; N represents the total number of pixels in the aerial image, i represents the pixel index, and y represents the pixel index. i Indicates the real label, p i Indicates the predicted probability;

[0039] The semantic segmentation model updates the model parameters θ based on the loss function;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] in, Let represent the partial derivative of the loss function with respect to the parameters, and t represent the number of iterations. This represents the first-order moment decay rate. This represents the second-order moment decay rate. This represents the first-order moment estimate. This represents the second-order moment estimate. This represents the first moment of deviation correction. This represents the second moment of deviation correction. This represents the gradient symbol.

[0047] This invention achieves high-precision segmentation of planting areas by fusing multispectral and vegetation index data inputs, combined with U-Net symmetric encoding and decoding structure and skip connection mechanism; and effectively solves the class imbalance problem of insufficient rapeseed pixel proportion by combining a hybrid loss function, thereby improving the recognition rate of small plots.

[0048] Furthermore, step 2 includes the following steps:

[0049] Step 21: Obtain a spatial distribution map of the planting area and extract the rapeseed planting area from it;

[0050] Step 22: Obtain high spatiotemporal resolution meteorological data of the planting area, and calculate meteorological indicators of the planting area based on the high spatiotemporal resolution meteorological data;

[0051] Step 23: Input meteorological indicators into the transformation model to generate environmental stress / suitability information that characterizes the formation of oil content.

[0052] Furthermore, meteorological indicators ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] Where HS represents the calorific value, used to describe the effect of temperature on oil content, and T... t T represents the average temperature on day t during the peak flowering period. f This indicates the start time of peak bloom, and the blooming period lasts 20 days.

[0059] WBI stands for Moisture Content Index, used to describe the impact of moisture on oil content; P represents cumulative rainfall. This indicates the reference evaporation rate of rapeseed flowers;

[0060] DTR stands for Temperature Difference Index, used to describe the effect of temperature difference on oil content. This represents the highest temperature on day t during the peak flowering season. This represents the lowest temperature on day t during the peak flowering season.

[0061] The light index describes the effect of light on oil content. This represents the total solar radiation on day t during the peak flowering period;

[0062] a, b, c, and d represent the corresponding weight information.

[0063] The technical solution provided in this application includes meteorological indicators such as temperature, humidity, and light, which reflect the comprehensive impact of climate conditions on the oil content of rapeseed. Thus, based on the meteorological indicators, the approximate range of oil content of rapeseed grown in each region can be determined.

[0064] Step 3 includes the following steps:

[0065] Step 31: Collect rainfall, sunshine and accumulated temperature data of different planting areas in the early stage of rapeseed growing season to obtain historical climate data; use historical climate data as samples and rapeseed yield as labels to establish a database;

[0066] Step 32: Use the established database to train the SVM model;

[0067] Step 33: Obtain this year's precipitation, sunshine, and accumulated temperature data, create a template, and then input it into the SVM model to generate rapeseed yield, using rapeseed yield as yield potential information.

[0068] In some possible embodiments, although the rapeseed yield data obtained based on support vector machines is not the actual yield, the yield contains high-dimensional weather information. In essence, it is to reduce the dimensionality of the weather data in the early growing season, so as to accurately represent the implicit information of yield potential.

[0069] Furthermore, the SVM model is as follows:

[0070] ;

[0071] The constraints are:

[0072] ;

[0073] Where w represents the weight vector, b represents the bias term, and C represents the penalty coefficient. Denotes the first slack variable. Let y represent the second slack variable, n represent the number of templates, and y represent the second slack variable. i Indicates a label.

[0074] The beneficial effects of this application are as follows:

[0075] This invention marks the first time that large-scale, regionalized prediction of oil content—a core quality indicator of rapeseed—has been achieved. By integrating key meteorological stress indicators during the flowering and pod-forming stages (such as effective accumulated temperature and diurnal temperature range), yield potential signals based on previous climate data, and the projected total yield for the year, a predictive model reflecting the complex mechanism of rapeseed oil formation has been constructed. This effectively addresses the core deficiency of existing satellite remote sensing technology, which can only predict total yield but cannot determine oil content, providing the market with crucial "quality" dimension information and filling the information blind spot in traditional yield prediction systems.

[0076] Forecasted oil content data enables oilseed processing companies to accurately quantify raw material needs and costs. By combining the predicted oil content with projected output, companies can accurately calculate the actual amount of rapeseed needed to produce the target rapeseed oil yield. For example, if a 1% drop in oil content is anticipated, companies can plan ahead for an increase of approximately 1% in raw material procurement and accurately assess the resulting additional storage space requirements, transportation costs, and processing energy consumption. This fundamentally avoids raw material shortages, significant cost budget deviations (e.g., a 1% fluctuation in oil content corresponds to approximately 1% in additional overall costs), and cash flow problems caused by unexpected fluctuations in oil content, greatly improving the accuracy of cost budgeting and operational stability.

[0077] Based on dual forecasts of "yield + oil content," enterprises can make more scientific and optimized decisions across the entire supply chain. On the procurement side, a "premium pricing" strategy can be implemented, accurately pricing and securing high-quality sources based on predicted oil content. On the production side, oil pressing plans can be rationally arranged, and process parameters (such as pre-pressing temperature) adjusted to suit raw material quality, optimizing equipment utilization and energy consumption. On the warehousing and logistics side, inventory turnover and transportation scheduling can be optimized based on raw material quality and demand forecasts. On the financial side, acquisition funds, production reserves, and risk management funds can be allocated more rationally. This decision-making based on accurate quality information significantly improves the enterprise's resource utilization efficiency, market responsiveness, and overall competitiveness. Attached Figure Description

[0078] Figure 1 This is a flowchart of a method for predicting rapeseed oil content based on big data. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0080] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0081] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0082] Example 1

[0083] The first embodiment of this application discloses a method for predicting the oil content of rapeseed based on big data, including:

[0084] Step 1: Acquire high-resolution satellite or UAV remote sensing images of the target area during the rapeseed flowering period; apply semantic segmentation algorithms to process the images and accurately extract the spatial distribution map of the planting area; based on the spectral characteristics of rapeseed flowers and phenological models in the images, identify and record the flowering period time of each extracted planting area.

[0085] The peak flowering period is when rapeseed blooms, and it has a significant impact on rapeseed yield and oil content. If favorable weather conditions are not present during the peak flowering period, it can lead to reduced yield and lower oil content.

[0086] This application uses aerial imagery to determine the planting area. Rapeseed fields visible in aerial imagery are considered large-scale rapeseed fields, and the planting techniques for these fields are fixed. Therefore, this application does not consider the influence of planting conditions, but only analyzes the impact of climate conditions on oil content.

[0087] Step 11: Obtain the flowering period range of the target area, acquire remote sensing image data within the flowering period range, and extract the infrared spectrum of the remote sensing image data to generate a spectral image.

[0088] A spectral image is essentially the reflected light information of each pixel. By using the reflected light, we can distinguish the approximate composition of the region corresponding to each pixel. Examples include: grassland, rapeseed fields, buildings, and highways.

[0089] Step 12: Use the FMask algorithm to extract the cloud region from the spectral image, obtain multiple spectral images, and fuse the spectral images with the cloud region removed to generate an upper-air image of the target area without cloud cover.

[0090] Step 12 includes the following steps:

[0091] Step 121: Detect the reflectance of the spectral image at each pixel in each band.

[0092] Step 122: Set threshold ranges for the visible light band and the short-wave infrared band respectively; if the reflectance of a pixel in the visible light band is within the set threshold range, and the reflectance of a pixel in the short-wave infrared band is also within the threshold range, then the pixel is a strongly extracted edge.

[0093] If a pixel's reflectance is within the set threshold range only in the visible light band or in the short-wave infrared band, then the pixel is set as a weakly extracted edge.

[0094] Step 123: Connect weak extraction edges. If a weak extraction edge is connected to a strong extraction edge, then set the weak extraction edge as a strong extraction edge and treat all strong extraction edges as layer regions.

[0095] In this application, the determination of whether something is a cloud based on reflectivity mainly utilizes the fact that clouds have a uniform color, resulting in a relatively high range of reflected light. A weak edge is the area around the cloud that is not thick enough to see below. Therefore, if a weak edge can connect to a strong edge, it means that the weak edge is part of the strong edge; otherwise, it is not.

[0096] Step 13: Extract the rapeseed planting area from the aerial image based on the semantic segmentation model and generate a spatial distribution map of the target area.

[0097] Semantic segmentation models include:

[0098] Input layer: used to input unobstructed high-altitude images, where each pixel in the high-altitude image data includes the reflectance spectrum of the corresponding region;

[0099] Encoding layer: High-altitude semantic features are extracted from the aerial images using a 4-level downsampling method; the 4-level downsampling includes two convolutional kernels and two max pooling operations;

[0100] Decoding layer: Performs 4 levels of upsampling on high-level semantic features to recover the spatial features of high-level semantic features; the 4 levels of upsampling include 2 transposed convolutions and 2 ordinary convolutions;

[0101] The output layer generates the probability of whether each pixel is a planting area based on the spatial features of high-level semantic features, and generates a spatial distribution map of the target area based on the probability of whether it is a planting area.

[0102] Loss function of semantic segmentation model for:

[0103] ;

[0104] ;

[0105] ;

[0106] in, For binary cross-entropy loss, For Dice loss; N represents the total number of pixels in the aerial image, i represents the pixel index, and y represents the pixel index. i Indicates the real label, p i Indicates the predicted probability;

[0107] The semantic segmentation model updates the model parameters θ based on the loss function;

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] in, Let represent the partial derivative of the loss function with respect to the parameters, and t represent the number of iterations. This represents the first-order moment decay rate. This represents the second-order moment decay rate. This represents the first-order moment estimate. This represents the second-order moment estimate. This represents the first moment of deviation correction. This represents the second moment of deviation correction. This represents the gradient symbol.

[0115] Step 14: Generate the peak flowering time for each planting area based on the time when the rapeseed flower spectrum begins to appear in each planting area.

[0116] After rapeseed flowers bloom, their color changes from green to yellow in aerial images, corresponding to a change in the spectral density. Therefore, the peak bloom time for each rapeseed planting area can be determined based on the spectral changes in the planting area. Compared to the average bloom time provided by the agricultural bureau, the peak bloom time measured by aerial imagery will be more accurate.

[0117] Step 2: Obtain high spatiotemporal resolution meteorological data of each planting area within the target region during the key growth stages before and after the full bloom period; perform spatiotemporal matching and fusion of the full bloom period time and corresponding temperature data of each area; based on rapeseed physiological and ecological research, calculate meteorological indicators related to oil content, and generate environmental stress / suitability information that characterizes the formation of oil content.

[0118] Step 2 includes the following steps:

[0119] Step 21: Obtain a spatial distribution map of the planting area and extract the rapeseed planting area from it.

[0120] Step 22: Obtain high spatiotemporal resolution meteorological data of the planting area, and calculate meteorological indicators of the planting area based on the high spatiotemporal resolution meteorological data;

[0121] Meteorological indicators ;

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] Where HS represents the calorific value, used to describe the effect of temperature on oil content, and T... t T represents the average temperature on day t during the peak flowering period. f This indicates the start time of peak bloom, and the blooming period lasts 20 days.

[0128] WBI stands for Moisture Content Index, used to describe the impact of moisture on oil content; P represents cumulative rainfall. This indicates the reference evaporation rate of rapeseed flowers;

[0129] DTR stands for Temperature Difference Index, used to describe the effect of temperature difference on oil content. This represents the highest temperature on day t during the peak flowering season. This represents the lowest temperature on day t during the peak flowering season.

[0130] The light index describes the effect of light on oil content. This represents the total solar radiation on day t during the peak flowering period.

[0131] Meteorological indicators primarily consider factors related to the flowering period. This is mainly because the flowering period is a crucial stage for rapeseed pollination and also a critical stage for seed embryo development. Poor meteorological indicators during the flowering period will inevitably lead to impaired embryo development, resulting in a simultaneous decrease in yield and oil content. Especially regarding oil content, although poorly developed embryos may form rapeseed, the oil content of the rapeseed is likely to be significantly affected.

[0132] Step 23: Input meteorological indicators into the transformation model to generate environmental stress / suitability information that characterizes the formation of oil content.

[0133] The conversion model is primarily a proportional conversion model, where the meteorological indicator S is a specific number. Based on the magnitude of the number, the meteorological indicator is converted into a corresponding percentage. This percentage represents the environmental stress / suitability information affecting oil content formation. In practice, this percentage reflects the oil content information of rapeseed in the corresponding planting area.

[0134] Step 3: Obtain historical climate data (such as precipitation, sunshine hours, accumulated temperature, etc.) of the target area during the early rapeseed growing season (from sowing to bolting); construct and train a support vector machine regression model, which uses the previous climate data as input to learn the nonlinear relationship between the data and historical rapeseed yield; use the trained support vector machine regression model to process the previous climate data of the target area in the current year and output information representing the yield potential predicted based on the previous climate conditions.

[0135] Step 3 includes the following steps:

[0136] Step 31: Collect rainfall, sunshine and accumulated temperature data of different planting areas in the early stage of rapeseed growing season to obtain historical climate data; use historical climate data as samples and rapeseed yield as labels to establish a database;

[0137] Step 32: Use the established database to train the SVM model;

[0138] The SVM model is:

[0139] ;

[0140] The constraints are:

[0141] ;

[0142] Where w represents the weight vector, b represents the bias term, and C represents the penalty coefficient. Denotes the first slack variable. Let y represent the second slack variable, n represent the number of templates, and y represent the second slack variable. i Indicates a label.

[0143] Step 33: Obtain this year's precipitation, sunshine, and accumulated temperature data, create a template, and then input it into the SVM model to generate rapeseed yield, using rapeseed yield as yield potential information.

[0144] Yield potential information is mainly derived from climatic conditions to determine the possible yield of rapeseed. This yield is not the total yield of all rapeseed in the entire target area, but rather the yield per acre (mu). In other words, it reflects the impact of climatic information on rapeseed yield per acre. This application chooses Support Vector Machine (SVM) for prediction because SVM can use planar segmentation to regress complex sample data to specific yield information.

[0145] The yield potential information here represents the probability label of the corresponding yield. In a specific yield prediction scheme, the SVM model outputs the probability of the extracted features corresponding to different labels, which is then input into the actual labels after linear regression. The yield potential information in this scheme is actually the classification information with implied corresponding labels extracted by the SVM model.

[0146] Step 4: Collect data on rapeseed planting area and corresponding actual rapeseed oil production in the target area over the years to form a time series sample; construct and train a long short-term memory network model, which learns the temporal dependency between historical planting area and rapeseed oil production; input the rapeseed planting area extracted in Step 1 into the trained LSTM model to predict the total expected rapeseed production for the year.

[0147] Collect the following data for the target area over several consecutive years: Rapeseed planting area: the actual rapeseed planting area each year (unit: hectares or mu). Actual rapeseed oil yield: the total rapeseed yield harvested in the corresponding year in this area (unit: tons). Sample construction: Sort the collected data by year to form a time series sample set.

[0148] Construct a Long Short-Term Memory (LSTM) network model. LSTM is a variant of recurrent neural networks specifically designed for learning long-term dependencies in time series. Use the mean squared error as the loss function to train the LSTM model. After training, input the rapeseed planting area of ​​the target region extracted / obtained in step 1 for the current year into the trained LSTM model to obtain the estimated total rapeseed yield for the year.

[0149] Step 5: Integrate environmental stress / suitability information, yield potential information, and the expected total rapeseed yield at the feature level to form a multi-dimensional fused feature vector that comprehensively reflects the rapeseed growth status, environmental conditions, and yield expectations; construct and train a convolutional neural network model, which takes the fused feature vector as input and learns the complex mapping relationship between it and the average oil content of rapeseed in the target area; use the trained CNN model to process the fused feature vector of the current year, and finally output the predicted value of the average oil content of rapeseed in the target area.

[0150] Environmental stress / suitability information is a two-dimensional information. The first dimension of this two-dimensional information is the coordinates of the planting area, and the second dimension is the implicit information of the oil content of rapeseed in the corresponding planting area.

[0151] Yield potential information is a two-dimensional information. The first dimension of this two-dimensional information is the coordinates of the planting area, and the second dimension is the yield potential information (yield per acre) of the corresponding planting area.

[0152] The total estimated rapeseed production is a specific figure.

[0153] Therefore, when fusing environmental stress / suitability information and yield potential information, which are two-dimensional information, it is only necessary to align the coordinates of the planting area to form a coordinate plane for the planting area. Then, three channels of information are added to each pixel on the coordinate plane. These channels represent the implicit information of rapeseed oil content, yield potential information, and the expected total rapeseed yield, respectively. Thus, the multi-dimensional fused feature vector obtained after feature fusion is actually a high-dimensional matrix vector. The predicted average oil content is used as the label information, and the multi-dimensional fused feature vector is used as the input information of the CNN model. This allows the model to output the predicted average oil content of rapeseed in the target area.

[0154] A CNN model consists of a multi-channel input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0155] The input layer is a multi-channel input layer, which is used to input three feature vectors of the multi-dimensional fused feature vector.

[0156] The convolutional layer uses convolutional kernels to automatically extract local spatial patterns contained in multi-dimensional fused feature vectors.

[0157] Pooling layers perform average pooling on the convolutional features extracted by the convolutional layers, outputting implicit information related to the label information.

[0158] The output layer uses the sigmoid function to regress the latent information to the corresponding average oil content prediction value.

[0159] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting rapeseed oil content based on big data, characterized in that, include: Step 1: Acquire high-resolution satellite or UAV remote sensing images of the target area during the rapeseed flowering period; apply semantic segmentation algorithms to process the images and accurately extract the spatial distribution map of the planting area; based on the spectral characteristics of rapeseed flowers and phenological models in the images, identify and record the flowering period time of each extracted planting area; Step 2: Obtain high spatiotemporal resolution meteorological data of each planting area within the target region during the key growth stages before and after the full bloom period; perform spatiotemporal matching and fusion of the full bloom period time and corresponding temperature data of each area; based on the physiological and ecological research of rapeseed, calculate meteorological indicators related to oil content, and generate environmental stress / suitability information that characterizes the formation of oil content. Step 3: Obtain historical climate data for the target area during the early rapeseed growing season, which is from sowing to bolting. Historical climate data includes precipitation, sunshine hours, and accumulated temperature. Construct and train a support vector machine regression model. This model uses the previous climate data as input to learn the nonlinear relationship between the previous climate data and historical rapeseed yield. Use the trained support vector machine regression model to process the previous climate data of the target area in the current year and output information representing the yield potential predicted based on the previous climate conditions. Step 4: Collect data on rapeseed planting area and corresponding actual rapeseed oil production in the target area over the years to form a time series sample; construct and train a long short-term memory network model, which learns the temporal dependency between historical planting area and rapeseed oil production; input the rapeseed planting area extracted in Step 1 into the trained LSTM model to predict the total expected rapeseed production for the year. Step 5: Integrate environmental stress / suitability information, yield potential information, and the expected total rapeseed yield at the feature level to form a multi-dimensional fused feature vector that comprehensively reflects the rapeseed growth status, environmental conditions, and yield expectations; construct and train a convolutional neural network model, which takes the fused feature vector as input and learns the complex mapping relationship between it and the average oil content of rapeseed in the target area; use the trained CNN model to process the fused feature vector of the current year, and finally output the predicted value of the average oil content of rapeseed in the target area; Step 1 includes the following steps: Step 11: Obtain the flowering period range of the target area, acquire remote sensing image data within the flowering period range, and extract the infrared spectrum of the remote sensing image data to generate a spectral image; Step 12: Use the FMask algorithm to extract the cloud region from the spectral image, obtain multiple spectral images, and fuse the spectral images with the cloud region removed to generate an upper-air image of the target area without cloud cover. Step 13: Extract the rapeseed planting area from the aerial image based on the semantic segmentation model and generate a spatial distribution map of the target area; Step 14: Generate the peak flowering time for each planting area based on the time when the rapeseed flower spectrum begins to appear in each planting area.

2. The method for predicting rapeseed oil content based on big data according to claim 1, characterized in that, Step 12 includes the following steps: Step 121: Detect the reflectance of each pixel in the spectral image at each wavelength. Step 122: Set threshold ranges for the visible light band and the short-wave infrared band respectively; if the reflectance of a pixel in the visible light band is within the set threshold range, and the reflectance of a pixel in the short-wave infrared band is also within the threshold range, then the pixel is a strongly extracted edge. If a pixel has a reflectance in the visible light band within the set threshold range or a reflectance in the short-wave infrared band within the threshold range, then the pixel is set as a weakly extracted edge. Step 123: Connect weak extraction edges. If a weak extraction edge is connected to a strong extraction edge, then set the weak extraction edge as a strong extraction edge and treat all strong extraction edges as cloud regions.

3. The method for predicting rapeseed oil content based on big data according to claim 1, characterized in that, Semantic segmentation models include: Input layer: used to input unobstructed high-altitude images, wherein each pixel in the high-altitude image data includes the reflectance spectrum of the corresponding region; Encoding layer: High-altitude semantic features are extracted from the aerial images using a 4-level downsampling method; the 4-level downsampling includes two convolutional kernels and two max pooling operations; Decoding layer: Performs 4 levels of upsampling on high-level semantic features to recover the spatial features of high-level semantic features; the 4 levels of upsampling include 2 transposed convolutions and 2 ordinary convolutions; The output layer generates the probability of whether each pixel is a planting area based on the spatial features of high-level semantic features, and generates a spatial distribution map of the target area based on the probability of whether it is a planting area.

4. The method for predicting rapeseed oil content based on big data according to claim 3, characterized in that, Loss function of semantic segmentation model for: ； ； ； Among them, L BCE For binary cross-entropy loss, For Dice loss; N represents the total number of pixels in the aerial image, i represents the pixel index, and y represents the pixel index. i Indicates the real label, p i Indicates the predicted probability; The semantic segmentation model updates the model parameters θ based on the loss function; ； ； ； ； ； ； in, Let represent the partial derivative of the loss function with respect to the parameters, and t represent the number of iterations. This represents the first-order moment decay rate. This represents the second-order moment decay rate. This represents the first-order moment estimate. This represents the second-order moment estimate. This represents the first moment of deviation correction. This represents the second moment of deviation correction. This represents the gradient symbol.

5. The method for predicting rapeseed oil content based on big data according to claim 4, characterized in that, Step 2 includes the following steps: Step 21: Obtain a spatial distribution map of the planting area and extract the rapeseed planting area from it; Step 22: Obtain high spatiotemporal resolution meteorological data of the planting area, and calculate meteorological indicators of the planting area based on the high spatiotemporal resolution meteorological data; Step 23: Input meteorological indicators into the transformation model to generate environmental stress / suitability information that characterizes the formation of oil content.

6. The method for predicting rapeseed oil content based on big data according to claim 5, characterized in that, Meteorological indicators ; ; ; ; ； ； Where HS represents the calorific value, used to describe the effect of temperature on oil content, and T... t T represents the average temperature on day t during the peak flowering period. f This indicates the start time of peak bloom, and the blooming period lasts 20 days. WBI stands for Moisture Content Index, used to describe the impact of moisture on oil content; P represents cumulative rainfall. This indicates the reference evaporation rate of rapeseed flowers; DTR stands for Temperature Difference Index, used to describe the effect of temperature difference on oil content. This represents the highest temperature on day t during the peak flowering season. This represents the lowest temperature on day t during the peak flowering season. The light index describes the effect of light on oil content. This represents the total solar radiation on day t during the peak flowering period; a, b, c, and d represent the corresponding weight information.

7. The method for predicting rapeseed oil content based on big data according to claim 4, characterized in that, Step 3 includes the following steps: Step 31: Collect rainfall, sunshine and accumulated temperature data of different planting areas in the early stage of rapeseed growing season to obtain historical climate data; use historical climate data as samples and rapeseed yield as labels to establish a database; Step 32: Use the established database to train the SVM model; Step 33: Obtain this year's precipitation, sunshine, and accumulated temperature data, create a template, and then input it into the SVM model to generate rapeseed yield, using rapeseed yield as yield potential information.

8. The method for predicting rapeseed oil content based on big data according to claim 7, characterized in that, The SVM model is: ； The constraints are: ； Where w represents the weight vector, b represents the bias term, and C represents the penalty coefficient. Denotes the first slack variable. Let y represent the second slack variable, n represent the number of templates, and y represent the second slack variable. i Indicates a label.

Images