Soybean yield estimation method based on multi-dimensional collaborative characteristics and physical information neural network
By employing a multidimensional collaborative feature and physical information neural network approach, the dynamic variation and sample sparsity issues in soybean yield estimation in existing technologies have been addressed. This approach achieves high-precision and stable yield prediction, adapts to differences in growth patterns across different growth stages and regions, and enhances the model's generalization ability and interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies for estimating soybean yield using multispectral remote sensing data struggle to adapt to yield differences at different growth stages, ignore dynamic changes in crop growth, and suffer from insufficient model generalization ability due to sparse samples, thus failing to achieve high-precision and stable yield prediction.
This paper adopts a multidimensional collaborative feature and physical information neural network approach. By deploying a remote sensing data acquisition system to collect multispectral remote sensing images, an exponential feature expression module and a key growth period multidimensional collaborative feature extraction module are established. Combined with a hierarchical yield estimation physical information neural network, long short-term memory network and convolution operation are used to capture feature dependencies and cross-level correlations. Light energy utilization theory is introduced as a physical constraint to improve the model's generalization ability.
It achieves high-precision and robust soybean yield estimation, can adapt to differences in growth patterns of different varieties and regions, overcomes the problem of sample sparsity, and provides reliable agricultural management data support.
Smart Images
Figure CN121746945A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural remote sensing and precision agriculture technology, specifically a method for estimating soybean yield using multidimensional collaborative features and physical information neural networks. Background Technology
[0002] In agricultural production, the inaccuracy of existing technologies in practical applications is a frequent problem, easily leading to errors in national food security decision-making and chaos in agricultural production management. Therefore, accurate yield estimation is necessary as early as possible during the crop growing season. This problem is particularly prominent in major soybean-producing areas such as Northeast China, due to the short growth cycle, complex environmental background, and subtle differences in management practices. Agricultural satellites or drones can quickly collect multispectral remote sensing data of corresponding crops. Furthermore, the physiological and ecological parameters of soybeans dynamically evolve throughout their growth period, resulting in changes in canopy spectral reflectance characteristics. Therefore, it is essential to utilize multispectral remote sensing data to remotely monitor soybean growth, estimate yields as early as possible during the growing season, and guide decision-making, which can significantly improve the operational efficiency of agricultural management enterprises.
[0003] The current main method for estimating soybean yield using multispectral remote sensing data is to collect multispectral remote sensing data deployed in farmland using multispectral sensors and then analyze the data using artificial intelligence models. The main methods currently employed include: 1) Statistical regression models based on single or a few vegetation indices. This method mainly uses the values of single or a few vegetation indices such as NDVI at a specific time phase to establish a relationship with historical yield. The disadvantage of this method is that it oversimplifies the complex crop growth process, ignores differences in growth stages, fails to capture dynamic changes, and is insensitive to stress factors such as sudden weather changes and pests and diseases, resulting in severely insufficient and highly unstable yield estimation accuracy; 2) Yield estimation models based on traditional machine learning. This method uses models such as random forests and support vector machines to process multiple remote sensing index features. This method can relatively effectively address the problem of excessively high feature dimensionality; however, the model is a "black box" and lacks physical interpretability. More importantly, when applied to new varieties, new regions, or years with abnormal climates, due to the drift in data distribution, this model often performs well in laboratory experiments, but in practical applications, it is prone to overfitting to specific patterns in the training data, leading to a sharp decline in predictive performance or even complete failure. It is difficult to distinguish between high-yield and medium-high-yield models, as the yield difference is small and its pattern is similar to the normal state; both of the above two types of methods perform poorly. 3) Time-series prediction models based on deep learning: This method uses complex deep learning models such as RNN and LSTM to distinguish the core differences between high-yield and medium-high-yield models. The biggest problem with this model is that the scarcity of samples leads to overfitting, poor generalization ability, and yield estimation accuracy that is far from meeting the needs of practical applications.
[0004] Therefore, a method is needed that not only adapts to the differences in the contribution of different growth stages to yield, but also can deeply explore and quantify the dynamic changes in crop growth, especially the changes in growth rate and key turning points that determine the final yield. At the same time, it should effectively overcome the problem of insufficient model generalization ability caused by sparse samples, so as to achieve early and accurate estimation of soybean yield. Summary of the Invention
[0005] The purpose of this invention is to provide a soybean yield estimation method based on multidimensional collaborative features and physical information neural networks, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a soybean yield estimation method based on multidimensional collaborative features and physical information neural networks, comprising the following steps:
[0007] S1. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire growth period of soybeans. Collect multispectral remote sensing images and ground-measured yield data to form a historical data table RSList. Obtain the number of elements in the historical data table RSNum and initialize the multidimensional collaborative feature list SFList for key growth periods.
[0008] S2. Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, and the output of IFModel is the exponential feature representation output array IFOutput.
[0009] S3. Establish a multi-dimensional collaborative feature extraction module TCModel for critical growth period. The input of TCModel is a list of image objects with 20 elements TCInput. TCModel processes TCInput using IFModel. The output of TCModel is the multi-dimensional collaborative feature output result TCOutput for critical growth period.
[0010] S4. Use TCModel to process all contents of RSList and store the results in SFList;
[0011] S5. Establish a hierarchical physical information neural network PINN for yield estimation, and train PINN using SFList;
[0012] S6. Collect multispectral image time-series data TestRS of the plot to be predicted using a remote sensing data acquisition system, and use TCModel and PINN to determine and estimate soybean yield.
[0013] Preferably, step S1 specifically includes:
[0014] S101. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire growth period of soybeans. Combine the multispectral remote sensing images with ground-measured yield data to form a historical data table, RSList. RSList consists of the following two fields:
[0015] Multispectral image time series data RSData: RSData is a list of image objects with 20 elements, each containing four bands: red, green, blue and near-infrared. To improve the richness of feature expression, the image can preferably also contain a red edge band. Since the soybean growth period is about 120 days, and images are acquired every 6 days, this field corresponds to the entire growth cycle of soybean.
[0016] YieldObserved: This is a floating-point variable representing the actual yield per mu (unit of land area) of the plot corresponding to this sample, in kilograms per mu.
[0017] S102, the number of elements in the historical data table RSNum is equal to the number of elements in RSList;
[0018] S103, Establish a multi-dimensional collaborative feature list for critical growth stages. SFList is an empty list; each element of this list contains the following 6 fields, and its structure is designed to store multi-level features and yield labels extracted in subsequent steps.
[0019] The first-level contextual feature F1 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector;
[0020] The second-level contextual feature F2 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector.
[0021] The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector.
[0022] The fourth-level contextual feature of the output feature, F4, is a list of 20 elements, where each element is a 10-dimensional vector.
[0023] The fifth level contextual feature of the output feature, F5, is a list of 20 elements, where each element is a 10-dimensional vector.
[0024] YieldDecision: Expresses the predicted output as a floating-point variable;
[0025] S104, Construct a multi-dimensional collaborative feature list counter for key growth stages, SFCounter=1;
[0026] S105, create an element SFListItem of SFList, the element has 6 fields including:
[0027] The first-level contextual feature of the output feature F1 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0028] The second-level contextual feature of the output feature F2 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0029] The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0030] The fourth-level contextual feature of the output feature F4 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0.
[0031] The fifth-level contextual feature of the output feature F5 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0.
[0032] YieldDecision=0;
[0033] S106, Step S1 ends.
[0034] Preferably, step S2 specifically includes:
[0035] S201, Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, which is a single scene image containing four bands: red, green, blue, and near-infrared.
[0036] S202, the following formula is used to calculate IFInput, extract five core vegetation indices, and calculate their statistical characteristics within the image area to form a 10-dimensional feature vector to achieve a quantitative expression of crop growth status; the selection of the five core vegetation indices is to comprehensively characterize the crop physiological status from different dimensions: NDVI reflects greenness, EVI enhances high biomass sensitivity, SAVI suppresses soil noise, NDWI monitors water status, and NDRE targets chlorophyll content in the middle and late stages.
[0037] Assume that the single-view image area contains 1 pixel, for any pixel Its red light band value Near-infrared band value The green light band value is Blue light band value The red-edge band value is If the image does not have a red-edge band, then ;
[0038] Calculation of Normalized Difference Vegetation Index The formula is:
[0039] Calculate the enhanced vegetation index The formula is:
[0040] in , , ;
[0041] Calculate the soil-adjusted vegetation index The formula is:
[0042] in ;
[0043] Calculate the normalized differential moisture index The formula is:
[0044] Calculate the red-edge normalized vegetation index The formula is:
[0045] Calculate the mean of the Normalized Difference Vegetation Index (NDVI) The formula is:
[0046] Calculate the standard deviation of the normalized vegetation index (NDVI) The formula is:
[0047] Calculate the mean of the Enhanced Vegetation Index (EVI) The formula is:
[0048] Calculate the standard deviation of the Enhanced Vegetation Index (EVI) The formula is:
[0049] Calculate the mean of the soil-adjusted vegetation index (SAVI). The formula is:
[0050] Calculate the standard deviation of the soil-adjusted vegetation index (SAVI). The formula is:
[0051] Calculate the mean of the normalized differential moisture index (NDWI). The formula is:
[0052] Calculate the standard deviation of the normalized differential moisture index (NDWI). The formula is:
[0053] Calculate the mean of the red-edge normalized vegetation index (NDRE). The formula is:
[0054] Calculate the standard deviation of the red-edge normalized vegetation index (NDRE) The formula is:
[0055] S203, Create an exponential feature expression output array IFOutput with 10 elements;
[0056] S204, store the 10 results calculated in step S202 sequentially into IFOutput, which is structured as follows:
[0057] S205 outputs IFOutput as the result of IFModel.
[0058] Preferably, step S3 specifically includes:
[0059] S301, Establish the multi-dimensional collaborative feature extraction module TCModel for critical growth period. The input of TCModel is a list of image objects with 20 elements, TCInput.
[0060] S302, the first counter TCCounter1 of the key growth period multidimensional collaborative feature extraction module is 1; the list variable TCList of the key growth period multidimensional collaborative feature extraction module is an empty list used to temporarily store the feature vectors processed by IFModel.
[0061] S303, the first temporary variable TCTemp1 of the key growth period multidimensional collaborative feature extraction module is used to extract the TCCounter1 image object of TCInput;
[0062] S304, the second temporary variable TCTemp2 in the critical growth period multidimensional collaborative feature extraction module is processed using IFModel, where IFInput of IFModel = TCTemp1, and IFOutput is obtained.
[0063] S305, add TCTemp2 to TCList;
[0064] S306, TCCounter1=TCCounter1+1;
[0065] S307, if TCCounter1 is less than or equal to 20, go to S303; otherwise, go to S308.
[0066] S308, establish the multi-dimensional collaborative feature output result TCOutput for the critical growth period. The structure of TCOutput is the same as the structure of an element of SFList.
[0067] S309, TCOutput's F1 field = TCList; this feature is defined as a zero-order feature, and its physical meaning lies in characterizing the static physiological state of the crop at any time point, providing a benchmark for the extraction of subsequent higher-order dynamic features;
[0068] S310, the F2 field of TCOutput calculates the first-order difference feature of TCList. This feature is defined as a first-order feature, and its physical meaning is to characterize the growth rate. It is calculated using the following formula: Let the first-order difference feature in TCList be... The elements are , From 0 to 19;
[0069] and (Zero Vector); This feature represents the rate of change of adjacent time steps, i.e., the instantaneous growth rate, which directly reflects the growth rate of crops at specific stages such as the seedling and flowering stages;
[0070] S311, the F3 field of TCOutput is equal to the calculated second-order difference feature of TCList. This feature is defined as a second-order feature, and its physical meaning is to characterize the growth acceleration. It is calculated using the following formula:
[0071] and This feature represents growth acceleration and is extremely sensitive to the dynamic changes during key physiological transitions in crops, such as the transition from vegetative growth to reproductive growth. It is an important indicator for judging whether the growth process is normal.
[0072] S312, the F4 field of TCOutput is equal to the calculated third-order difference feature of TCList. This feature is defined as a third-order feature, and its physical meaning is to characterize the growth inflection point. It is calculated using the following formula:
[0073] and This characteristic is extremely sensitive to inflection points in the growth curve, such as the peak flowering period and the beginning of the grain filling period. These inflection points are often the key moments that determine the final number and weight of grains, and are crucial for distinguishing between high-yield and medium-high-yield plots.
[0074] S313, the F5 field of TCOutput is equal to the fourth-order difference feature of TCList. This feature is defined as a fourth-order feature, and its physical meaning lies in representing the macroscopic shape. It is calculated using the following formula:
[0075] and This feature is used to describe the macroscopic shape and overall trend of the growth curve, which helps to distinguish the growth patterns of different varieties or management strategies such as water and fertilizer regulation. For example, there will be significant differences in this feature between robust growth patterns and excessive growth patterns.
[0076] S314, TCOutput's F2 field = TCList - F2 field;
[0077] S315, TCOutput's F3 field = TCList - F3 field;
[0078] S316, TCOutput's F4 field = TCList - F4 field;
[0079] S317, TCOutput's F5 field = TCList - F5 field;
[0080] S318, the YieldDecision field of TCOutput is equal to 0;
[0081] S319, output TCOutput as the result of TCModel.
[0082] Preferably, step S4 specifically includes:
[0083] S401, the result processing counter CLCounter equals 1;
[0084] S402, Result processing: The first temporary variable CLTemp1 is equal to the CLCounter-th element retrieved from RSList;
[0085] S403, the result processing second temporary variable CLTemp2 is equal to the processing using TCModel, TCModel's TCInput is equal to CLTemp1.RSData, and the processing obtains TCOutput;
[0086] S404, the value of the CLCounter-th element of SFList is equal to CLTemp2;
[0087] S405, the YieldDecision of the CLCounter-th element of SFList is equal to the YieldObserved field of the CLCounter-th element of RSList;
[0088] S406, CLCounter equals CLCounter plus 1;
[0089] S407, if CLCounter is less than or equal to RSNum, go to S402; otherwise, go to S408.
[0090] S408, the processing of step S4 is complete.
[0091] Preferably, step S5 specifically includes:
[0092] S501, establish a hierarchical physical information neural network (PINN) model for yield estimation. This model contains the following structure, which is the key to achieving feature architecture synergy: the model architecture design aims to extract the temporal dependencies of features at each level, and capture complex patterns across levels through fusion and convolution operations, ultimately achieving accurate regression.
[0093] The model contains 5 input branches, each corresponding to a feature level F1 to F5;
[0094] Each branch first passes through a Long Short-Term Memory (LSTM) layer to effectively capture the temporal dependencies and long-term memory within the features of each level. Preferably, the LSTM layer contains 128 hidden units to extract the temporal dependencies of the features at each level and output a 100-dimensional feature vector. The output vectors of the five branches are concatenated along the feature dimension to form a 500-dimensional fused feature vector.
[0095] The fused feature vector is input into the first one-dimensional convolutional layer to extract local spatial patterns and cross-level associations from the fused high-dimensional features. Preferably, the first one-dimensional convolutional layer contains 64 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function. It is followed by a max pooling layer. Preferably, the pooling window size is 2 and the stride is 2.
[0096] The pooled features are input into a second one-dimensional convolutional layer to further deepen feature extraction and enhance the non-linear expressive power of the model. Preferably, the second one-dimensional convolutional layer contains 32 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function.
[0097] After the convolutional features are flattened, they are input into a multilayer perceptron (MLP) network as the final regressor, which maps the extracted abstract features to the output value. The network contains three fully connected layers with 256, 128 and 1 neurons respectively. The activation function of the first two layers is ReLU, and the last layer is a linear activation function. The final output is the predicted output.
[0098] S502, PINN’s five input branches are connected to the F1, F2, F3, F4 and F5 fields of SFList respectively, and PINN’s decision output is used to fit the YieldDecision field of SFList;
[0099] S503 uses SFList data to train PINN. During training, the loss function consists of both data loss and physical loss terms, enabling PINN to make decisions. Specifically:
[0100] (1) Define the data loss term Let the total number of training samples be... For the first For each sample, the neural network predicts the output as follows: The actual output was The data loss term is the average of the mean squared errors between the predicted and actual yields for all samples, and the formula is:
[0101] (2) Define the physical loss term The light energy utilization efficiency theory is introduced as a physical constraint. This theory is a classic model for estimating crop primary productivity and can provide neural networks with prior knowledge that does not depend on sparse samples. Specifically:
[0102] a. Based on the time series data of the NDVI mean at level F1, for the first... At each time point, the mean NDVI was Estimate each time point using empirical formulas. Photosynthetically active radiation absorption ratio :
[0103] in and These are empirical coefficients, for example, they can be taken as 1.25 and -0.15 respectively, and the specific values can be fine-tuned according to the region and crop variety;
[0104] b. Obtain each time point Incident photosynthetic effective radiation This data can be obtained from local weather station measurements or publicly available meteorological reanalysis data products such as ERA5_Land, and is measured in megajoules per square meter per day.
[0105] c. Let the light energy utilization efficiency be... The unit is grams of carbon per megajoule. This parameter can be obtained from relevant agricultural literature, and for soybeans, the value is usually between 1.0 and 2.0. The cumulative total primary productivity (GPP) over the entire growth period is calculated using the following formula:
[0106] d. Let the harvest index be... The unit is dimensionless, and for soybeans it is usually between 0.4 and 0.5, which can be determined through variety information; let the unit conversion factor be... This is used to convert carbon units into dry matter weight and ultimately into yield per acre (kg / acre); it converts GPP into physical yield estimates. The calculation formula is:
[0107] e. The physical loss term is the average of the mean squared errors between the neural network-predicted output and the physical estimated output for all samples, and the formula is:
[0108] (3) Define the total loss function By weighted fusion of data loss and physical loss, the model can fit real data while adhering to basic physical laws of crop growth, thereby suppressing overfitting and improving generalization ability when samples are sparse; the total loss function formula is:
[0109] Preferably, wherein , This weighting ratio can be adjusted according to the amount of data; the less data, the lower the weighting ratio. The weight can be appropriately increased;
[0110] S504, the process of step S5 is complete.
[0111] Preferably, step S6 specifically includes:
[0112] S601, using a remote sensing data acquisition system to collect time-series multispectral image data TestRS of the plot to be predicted;
[0113] S602, the first temporary variable PDTemp1 used for judgment is equal to the value of TCModel, TCInput of TCModel is equal to TestRS, and TCOutput is obtained after processing.
[0114] S603, input PDTemp1 into the trained PINN model, and connect the five input branches of PINN to the F1, F2, F3, F4 and F5 fields of PDTemp1 respectively;
[0115] S604 is the decision output used to determine whether the second temporary variable PDTemp2 is equal to PINN;
[0116] S605, output PDTemp2 as the final estimated yield result, then proceed to S606;
[0117] S606, the patent description process is complete.
[0118] Compared with the prior art, the beneficial effects of the present invention are:
[0119] This invention establishes a multi-dimensional collaborative feature extraction module for key growth stages and a hierarchical yield estimation physical information neural network model. By using these two modules, features of multispectral remote sensing data are constructed hierarchically to explore the core differences in soybean yield formation under different growth conditions, thereby achieving high-precision and robust soybean yield estimation.
[0120] The hierarchical structure of this invention enables the description of the changes in multispectral remote sensing data across different growth stages. On one hand, this allows the neural network to consistently identify specific levels corresponding to key stages of yield formation, adapting to differences in growth patterns caused by different varieties and regions, and uncovering deeper biological information beyond a single index. On the other hand, the differences between levels calculated using formulas can more clearly present periodic variations without relying on samples, which helps address situations where ground samples are scarce. In particular, by introducing a physical information neural network, scientific constraints are imposed on the model, effectively suppressing overfitting caused by insufficient samples and significantly enhancing the model's generalization ability and interpretability. This invention enables more efficient and accurate estimation of soybean yield, providing more reliable and precise data support for agricultural production management. Attached Figure Description
[0121] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0122] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0123] Please see Figure 1 This invention provides a method for estimating soybean yield using multidimensional collaborative features and physical information neural networks, comprising the following steps:
[0124] S1. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire soybean growth period. Collect the multispectral remote sensing images and ground-measured yield data to form a historical data table RSList; obtain the number of elements in the historical data table RSNum; and initialize a multidimensional collaborative feature list SFList for key growth stages; specifically including:
[0125] S101. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire growth period of soybeans. Combine the multispectral remote sensing images with ground-measured yield data to form a historical data table, RSList. RSList consists of the following two fields:
[0126] Multispectral image time series data RSData: RSData is a list of image objects with 20 elements, each containing four bands: red, green, blue and near-infrared. To improve the richness of feature expression, the image can preferably also contain a red edge band. Since the soybean growth period is about 120 days, and images are acquired every 6 days, this field corresponds to the entire growth cycle of soybean.
[0127] YieldObserved: This is a floating-point variable representing the actual yield per mu (unit of land area) of the plot corresponding to this sample, in kilograms per mu.
[0128] S102, the number of elements in the historical data table RSNum is equal to the number of elements in RSList;
[0129] S103, Establish a multi-dimensional collaborative feature list for critical growth stages. SFList is an empty list; each element of this list contains the following 6 fields, and its structure is designed to store multi-level features and yield labels extracted in subsequent steps.
[0130] The first-level contextual feature F1 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector;
[0131] The second-level contextual feature F2 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector.
[0132] The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector.
[0133] The fourth-level contextual feature of the output feature, F4, is a list of 20 elements, where each element is a 10-dimensional vector.
[0134] The fifth level contextual feature of the output feature, F5, is a list of 20 elements, where each element is a 10-dimensional vector.
[0135] YieldDecision: Expresses the predicted output as a floating-point variable;
[0136] S104, Construct a multi-dimensional collaborative feature list counter for key growth stages, SFCounter=1;
[0137] S105, create an element SFListItem of SFList, the element has 6 fields including:
[0138] The first-level contextual feature of the output feature F1 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0139] The second-level contextual feature of the output feature F2 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0140] The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0.
[0141] The fourth-level contextual feature of the output feature F4 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0.
[0142] The fifth-level contextual feature of the output feature F5 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0.
[0143] YieldDecision=0;
[0144] S106, Step S1 ends;
[0145] S2. Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, and the output of IFModel is the exponential feature representation output array IFOutput; specifically including:
[0146] S201, Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, which is a single scene image containing four bands: red, green, blue, and near-infrared.
[0147] S202, the following formula is used to calculate IFInput, extract five core vegetation indices, and calculate their statistical characteristics within the image area to form a 10-dimensional feature vector to achieve a quantitative expression of crop growth status; the selection of the five core vegetation indices is to comprehensively characterize the crop physiological status from different dimensions: NDVI reflects greenness, EVI enhances high biomass sensitivity, SAVI suppresses soil noise, NDWI monitors water status, and NDRE targets chlorophyll content in the middle and late stages.
[0148] Assume that the single-view image area contains 1 pixel, for any pixel Its red light band value Near-infrared band value The green light band value is Blue light band value The red-edge band value is If the image does not have a red-edge band, then ;
[0149] Calculation of Normalized Difference Vegetation Index The formula is:
[0150] Calculate the enhanced vegetation index The formula is:
[0151] in , , ;
[0152] Calculate the soil-adjusted vegetation index The formula is:
[0153] in ;
[0154] Calculate the normalized differential moisture index The formula is:
[0155] Calculate the red-edge normalized vegetation index The formula is:
[0156] Calculate the mean of the Normalized Difference Vegetation Index (NDVI) The formula is:
[0157] Calculate the standard deviation of the normalized vegetation index (NDVI) The formula is:
[0158] Calculate the mean of the Enhanced Vegetation Index (EVI) The formula is:
[0159] Calculate the standard deviation of the Enhanced Vegetation Index (EVI) The formula is:
[0160] Calculate the mean of the soil-adjusted vegetation index (SAVI). The formula is:
[0161] Calculate the standard deviation of the soil-adjusted vegetation index (SAVI). The formula is:
[0162] Calculate the mean of the normalized differential moisture index (NDWI). The formula is:
[0163] Calculate the standard deviation of the normalized differential moisture index (NDWI). The formula is:
[0164] Calculate the mean of the red-edge normalized vegetation index (NDRE). The formula is:
[0165] Calculate the standard deviation of the red-edge normalized vegetation index (NDRE) The formula is:
[0166] S203, Create an exponential feature expression output array IFOutput with 10 elements;
[0167] S204, store the 10 results calculated in step S202 sequentially into IFOutput, which is structured as follows:
[0168] S205, output IFOutput as the result of IFModel;
[0169] S3. Establish a multi-dimensional collaborative feature extraction module TCModel for critical growth periods. The input of TCModel is a list of image objects with 20 elements, TCInput. TCModel processes TCInput using IFModel, and the output of TCModel is the multi-dimensional collaborative feature output result for critical growth periods, TCOutput. Specifically, this includes:
[0170] S301, Establish the multi-dimensional collaborative feature extraction module TCModel for critical growth period. The input of TCModel is a list of image objects with 20 elements, TCInput.
[0171] S302, the first counter TCCounter1 of the key growth period multidimensional collaborative feature extraction module is 1; the list variable TCList of the key growth period multidimensional collaborative feature extraction module is an empty list used to temporarily store the feature vectors processed by IFModel.
[0172] S303, the first temporary variable TCTemp1 of the key growth period multidimensional collaborative feature extraction module is used to extract the TCCounter1 image object of TCInput;
[0173] S304, the second temporary variable TCTemp2 in the critical growth period multidimensional collaborative feature extraction module is processed using IFModel, where IFInput of IFModel = TCTemp1, and IFOutput is obtained.
[0174] S305, add TCTemp2 to TCList;
[0175] S306, TCCounter1=TCCounter1+1;
[0176] S307, if TCCounter1 is less than or equal to 20, go to S303; otherwise, go to S308.
[0177] S308, establish the multi-dimensional collaborative feature output result TCOutput for the critical growth period. The structure of TCOutput is the same as the structure of an element of SFList.
[0178] S309, TCOutput's F1 field = TCList; this feature is defined as a zero-order feature, and its physical meaning lies in characterizing the static physiological state of the crop at any time point, providing a benchmark for the extraction of subsequent higher-order dynamic features;
[0179] S310, the F2 field of TCOutput calculates the first-order difference feature of TCList. This feature is defined as a first-order feature, and its physical meaning is to characterize the growth rate. It is calculated using the following formula: Let the first-order difference feature in TCList be... The elements are , From 0 to 19;
[0180] and (Zero Vector); This feature represents the rate of change of adjacent time steps, i.e., the instantaneous growth rate, which directly reflects the growth rate of crops at specific stages such as the seedling and flowering stages;
[0181] S311, the F3 field of TCOutput is equal to the calculated second-order difference feature of TCList. This feature is defined as a second-order feature, and its physical meaning is to characterize the growth acceleration. It is calculated using the following formula:
[0182] and This feature represents growth acceleration and is extremely sensitive to the dynamic changes during key physiological transitions in crops, such as the transition from vegetative growth to reproductive growth. It is an important indicator for judging whether the growth process is normal.
[0183] S312, the F4 field of TCOutput is equal to the calculated third-order difference feature of TCList. This feature is defined as a third-order feature, and its physical meaning is to characterize the growth inflection point. It is calculated using the following formula:
[0184] and This characteristic is extremely sensitive to inflection points in the growth curve, such as the peak flowering period and the beginning of the grain filling period. These inflection points are often the key moments that determine the final number and weight of grains, and are crucial for distinguishing between high-yield and medium-high-yield plots.
[0185] S313, the F5 field of TCOutput is equal to the fourth-order difference feature of TCList. This feature is defined as a fourth-order feature, and its physical meaning lies in representing the macroscopic shape. It is calculated using the following formula:
[0186] and This feature is used to describe the macroscopic shape and overall trend of the growth curve, which helps to distinguish the growth patterns of different varieties or management strategies such as water and fertilizer regulation. For example, there will be significant differences in this feature between robust growth patterns and excessive growth patterns.
[0187] S314, TCOutput's F2 field = TCList - F2 field;
[0188] S315, TCOutput's F3 field = TCList - F3 field;
[0189] S316, TCOutput's F4 field = TCList - F4 field;
[0190] S317, TCOutput's F5 field = TCList - F5 field;
[0191] S318, the YieldDecision field of TCOutput is equal to 0;
[0192] S319, output TCOutput as the result of TCModel;
[0193] S4. Process all contents of RSList using TCModel and store the results in SFList; specifically including:
[0194] S401, the result processing counter CLCounter equals 1;
[0195] S402, Result processing: The first temporary variable CLTemp1 is equal to the CLCounter-th element retrieved from RSList;
[0196] S403, the result processing second temporary variable CLTemp2 is equal to the processing using TCModel, TCModel's TCInput is equal to CLTemp1.RSData, and the processing obtains TCOutput;
[0197] S404, the value of the CLCounter-th element of SFList is equal to CLTemp2;
[0198] S405, the YieldDecision of the CLCounter-th element of SFList is equal to the YieldObserved field of the CLCounter-th element of RSList;
[0199] S406, CLCounter equals CLCounter plus 1;
[0200] S407, if CLCounter is less than or equal to RSNum, go to S402; otherwise, go to S408.
[0201] S408, Step S4 processing ends;
[0202] S5. Establish a hierarchical physical information neural network (PINN) for yield estimation, and train PINN using SFList; specifically including:
[0203] S501, establish a hierarchical physical information neural network (PINN) model for yield estimation. This model contains the following structure, which is the key to achieving feature architecture synergy: the model architecture design aims to extract the temporal dependencies of features at each level, and capture complex patterns across levels through fusion and convolution operations, ultimately achieving accurate regression.
[0204] The model contains 5 input branches, each corresponding to a feature level F1 to F5;
[0205] Each branch first passes through a Long Short-Term Memory (LSTM) layer to effectively capture the temporal dependencies and long-term memory within the features of each level. Preferably, the LSTM layer contains 128 hidden units to extract the temporal dependencies of the features at each level and output a 100-dimensional feature vector. The output vectors of the five branches are concatenated along the feature dimension to form a 500-dimensional fused feature vector.
[0206] The fused feature vector is input into the first one-dimensional convolutional layer to extract local spatial patterns and cross-level associations from the fused high-dimensional features. Preferably, the first one-dimensional convolutional layer contains 64 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function. It is followed by a max pooling layer. Preferably, the pooling window size is 2 and the stride is 2.
[0207] The pooled features are input into a second one-dimensional convolutional layer to further deepen feature extraction and enhance the non-linear expressive power of the model. Preferably, the second one-dimensional convolutional layer contains 32 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function.
[0208] After the convolutional features are flattened, they are input into a multilayer perceptron (MLP) network as the final regressor, which maps the extracted abstract features to the output value. The network contains three fully connected layers with 256, 128 and 1 neurons respectively. The activation function of the first two layers is ReLU, and the last layer is a linear activation function. The final output is the predicted output.
[0209] S502, PINN’s five input branches are connected to the F1, F2, F3, F4 and F5 fields of SFList respectively, and PINN’s decision output is used to fit the YieldDecision field of SFList;
[0210] S503 uses SFList data to train PINN. During training, the loss function consists of both data loss and physical loss terms, enabling PINN to make decisions. Specifically:
[0211] (2) Define the data loss term Let the total number of training samples be... For the first For each sample, the neural network predicts the output as follows: The actual output was The data loss term is the average of the mean squared errors between the predicted and actual yields for all samples, and the formula is:
[0212] (2) Define the physical loss term The light energy utilization efficiency theory is introduced as a physical constraint. This theory is a classic model for estimating crop primary productivity and can provide neural networks with prior knowledge that does not depend on sparse samples. Specifically:
[0213] a. Based on the time series data of the NDVI mean at level F1, for the first... At each time point, the mean NDVI was Estimate each time point using empirical formulas. Photosynthetically active radiation absorption ratio :
[0214] in and These are empirical coefficients, for example, they can be taken as 1.25 and -0.15 respectively, and the specific values can be fine-tuned according to the region and crop variety;
[0215] b. Obtain each time point Incident photosynthetic effective radiation This data can be obtained from local weather station measurements or publicly available meteorological reanalysis data products such as ERA5_Land, and is measured in megajoules per square meter per day.
[0216] c. Let the light energy utilization efficiency be... The unit is grams of carbon per megajoule. This parameter can be obtained from relevant agricultural literature, and for soybeans, the value is usually between 1.0 and 2.0. The cumulative total primary productivity (GPP) over the entire growth period is calculated using the following formula:
[0217] d. Let the harvest index be... The unit is dimensionless, and for soybeans it is usually between 0.4 and 0.5, which can be determined through variety information; let the unit conversion factor be... This is used to convert carbon units into dry matter weight and ultimately into yield per acre (kg / acre); it converts GPP into physical yield estimates. The calculation formula is:
[0218] e. The physical loss term is the average of the mean squared errors between the neural network-predicted output and the physical estimated output for all samples, and the formula is:
[0219] (3) Define the total loss function By weighted fusion of data loss and physical loss, the model can fit real data while adhering to basic physical laws of crop growth, thereby suppressing overfitting and improving generalization ability when samples are sparse; the total loss function formula is:
[0220] Preferably, wherein , This weighting ratio can be adjusted according to the amount of data; the less data, the lower the weighting ratio. The weight can be appropriately increased;
[0221] S504, the processing step S5 is complete;
[0222] S6. Collect time-series multispectral image data (TestRS) of the plot to be predicted using a remote sensing data acquisition system; use TCModel and PINN to determine and estimate soybean yield; specifically including:
[0223] S601, using a remote sensing data acquisition system to collect time-series multispectral image data TestRS of the plot to be predicted;
[0224] S602, the first temporary variable PDTemp1 used for judgment is equal to the value of TCModel, TCInput of TCModel is equal to TestRS, and TCOutput is obtained after processing.
[0225] S603, input PDTemp1 into the trained PINN model, and connect the five input branches of PINN to the F1, F2, F3, F4 and F5 fields of PDTemp1 respectively;
[0226] S604 is the decision output used to determine whether the second temporary variable PDTemp2 is equal to PINN;
[0227] S605, output PDTemp2 as the final estimated yield result, then proceed to S606;
[0228] S606, the patent description process is complete.
[0229] Example:
[0230] To verify the effectiveness of the above soybean yield estimation method, a soybean planting base was selected as the experimental area. The area has flat terrain, uniform soil type, and consistent planting varieties, which makes it easy to control variables to accurately evaluate the performance of the estimation method.
[0231] First, multispectral image time-series data of the entire soybean growing season in the experimental area were acquired using a remote sensing data acquisition system. The image data were collected at 10-day intervals to ensure complete recording of all key stages of soybean growth. The acquired multispectral image time-series data were then processed according to the method described above to construct multidimensional collaborative features. Specifically, following steps S301 to S319, TCModel was used to extract and process features from the multispectral image time-series data, resulting in TCOutput, which includes zero-order, first-order, second-order, third-order, and fourth-order features.
[0232] Next, following steps S401 to S408, multiple sets of multispectral image time-series data (simulating RSList) are processed using TCModel, and the results are stored in SFList to provide data support for the subsequent training of the PINN model. When constructing SFList, it is ensured that the YieldDecision field of each element accurately records the actual yield data of the corresponding plot. This data is obtained through field sampling and measurement to guarantee the accuracy and reliability of the data.
[0233] Then, following steps S501 to S504, a hierarchical physical information neural network (PINN) model for yield estimation is established. During model training, the SFList data is divided into training, validation, and test sets according to a certain ratio. The training set is used for learning and optimizing model parameters, the validation set is used to adjust the model's hyperparameters to prevent overfitting, and the test set is used to evaluate the model's final performance. During training, parameters such as the learning rate are dynamically adjusted based on changes in the total loss function (Total_Loss) to improve the model's training efficiency and convergence.
[0234] After multiple iterations of training, training was stopped when the model's performance metrics (such as mean squared error and mean absolute error) on the validation set no longer showed significant improvement, resulting in a well-trained PINN model. The trained PINN model was evaluated using a test set. The results showed that the model's estimated soybean yield had a high correlation with the actual yield, and both the mean squared error and mean absolute error were kept within a small range. This indicates that the model can accurately capture the key features of soybean growth and achieve precise estimation of soybean yield.
[0235] Finally, following steps S601 to S606, multispectral image time-series data TestRS of the plots to be predicted (plots within the experimental area not involved in model training) were collected using a remote sensing data acquisition system. The soybean yield was then assessed and estimated using TCModel and the trained PINN model. The estimated results were compared with the actual yield measured in the field, further validating the effectiveness and reliability of the method in practical applications. This embodiment demonstrates that the soybean yield estimation method based on multidimensional collaborative features and physical information neural networks can fully utilize multidimensional information from remote sensing data, combined with the physical laws of crop growth, to achieve accurate soybean yield estimation, providing strong support for agricultural production decisions.
[0236] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A soybean yield estimation method based on multidimensional collaborative features and physical information neural networks, characterized by: Includes the following steps: S1. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire growth period of soybeans. Collect multispectral remote sensing images and ground-measured yield data to form a historical data table RSList. Obtain the number of elements in the historical data table RSNum and initialize the multidimensional collaborative feature list SFList for key growth periods. S2. Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, and the output of IFModel is the exponential feature representation output array IFOutput. S3. Establish a multi-dimensional collaborative feature extraction module TCModel for critical growth period. The input of TCModel is a list of image objects with 20 elements TCInput. TCModel processes TCInput using IFModel. The output of TCModel is the multi-dimensional collaborative feature output result TCOutput for critical growth period. S4. Use TCModel to process all contents of RSList and store the results in SFList; S5. Establish a hierarchical physical information neural network PINN for yield estimation, and train PINN using SFList; S6. Collect multispectral image time-series data TestRS of the plot to be predicted using a remote sensing data acquisition system, and use TCModel and PINN to determine and estimate soybean yield.
2. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S1 specifically includes: S101. Deploy a remote sensing data acquisition system in the target farmland area to collect multispectral remote sensing images covering the entire growth period of soybeans. Combine the multispectral remote sensing images with ground-measured yield data to form a historical data table, RSList. RSList consists of the following two fields: Multispectral image time series data RSData: RSData is a list of image objects with 20 elements, each containing four bands: red, green, blue and near-infrared. To improve the richness of feature expression, the image can preferably also contain a red edge band. Since the soybean growth period is about 120 days, and images are acquired every 6 days, this field corresponds to the entire growth cycle of soybean. YieldObserved: This is a floating-point variable representing the actual yield per mu (unit of land area) of the plot corresponding to this sample, in kilograms per mu. S102, the number of elements in the historical data table RSNum is equal to the number of elements in RSList; S103, Establish a multidimensional collaborative feature list for key growth stages. SFList is an empty list; each element of the list contains the following 6 fields. The first-level contextual feature F1 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector; The second-level contextual feature F2 of the output feature is a list of 20 elements, where each element is a 10-dimensional vector. The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector. The fourth-level contextual feature of the output feature, F4, is a list of 20 elements, where each element is a 10-dimensional vector. The fifth level contextual feature of the output feature, F5, is a list of 20 elements, where each element is a 10-dimensional vector. YieldDecision: Expresses the predicted output as a floating-point variable; S104, Construct a multi-dimensional collaborative feature list counter for key growth stages, SFCounter=1; S105, create an element SFListItem of SFList, the element has 6 fields including: The first-level contextual feature of the output feature F1 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0. The second-level contextual feature of the output feature F2 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0. The third-level contextual feature of the output feature F3 is a list of 20 elements, where each element is a 10-dimensional vector and all elements inside the vector are 0. The fourth-level contextual feature of the output feature F4 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0. The fifth-level contextual feature of the output feature F5 is a list of 20 elements, where each element is a 10-dimensional vector, and all elements inside the vector are 0. YieldDecision=0; S106, Step S1 ends.
3. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S2 specifically includes: S201, Establish the exponential feature representation module IFModel. The input of IFModel is the exponential feature representation input image IFInput, which is a single scene image containing four bands: red, green, blue, and near-infrared. S202, the following formula is used to calculate IFInput, extract five core vegetation indices, and calculate their statistical characteristics in the image area to form a 10-dimensional feature vector; the selection of the five core vegetation indices is to comprehensively characterize the physiological state of crops from different dimensions: NDVI reflects greenness, EVI enhances high biomass sensitivity, SAVI suppresses soil noise, NDWI monitors water status, and NDRE targets chlorophyll content in the middle and late stages. Assume that the single-view image area contains 1 pixel, for any pixel Its red light band value Near-infrared band value The green light band value is Blue light band value The red-edge band value is If the image does not have a red-edge band, then ; Calculation of Normalized Difference Vegetation Index The formula is: , Calculate the enhanced vegetation index The formula is: , in , , ; Calculate the soil-adjusted vegetation index The formula is: , in ; Calculate the normalized differential moisture index The formula is: , Calculate the red-edge normalized vegetation index The formula is: , Calculate the mean of the Normalized Difference Vegetation Index (NDVI) The formula is: , Calculate the standard deviation of the normalized vegetation index (NDVI) The formula is: , Calculate the mean of the Enhanced Vegetation Index (EVI) The formula is: , Calculate the standard deviation of the Enhanced Vegetation Index (EVI) The formula is: , Calculate the mean of the soil-adjusted vegetation index (SAVI). The formula is: , Calculate the standard deviation of the soil-adjusted vegetation index (SAVI). The formula is: , Calculate the mean of the normalized differential moisture index (NDWI). The formula is: , Calculate the standard deviation of the normalized differential moisture index (NDWI). The formula is: , Calculate the mean of the red-edge normalized vegetation index (NDRE). The formula is: , Calculate the standard deviation of the red-edge normalized vegetation index (NDRE) The formula is: , S203, Create an exponential feature expression output array IFOutput with 10 elements; S204, store the 10 results calculated in step S202 sequentially into IFOutput, which is structured as follows: ， S205 outputs IFOutput as the result of IFModel.
4. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S3 specifically includes: S301, Establish the multi-dimensional collaborative feature extraction module TCModel for critical growth period. The input of TCModel is a list of image objects with 20 elements, TCInput. S302, the first counter TCCounter1 of the key growth period multidimensional collaborative feature extraction module is 1; the list variable TCList of the key growth period multidimensional collaborative feature extraction module is an empty list used to temporarily store the feature vectors processed by IFModel. S303, the first temporary variable TCTemp1 of the key growth period multidimensional collaborative feature extraction module is used to extract the TCCounter1 image object of TCInput; S304, the second temporary variable TCTemp2 in the critical growth period multidimensional collaborative feature extraction module is processed using IFModel, where IFInput of IFModel = TCTemp1, and IFOutput is obtained. S305, add TCTemp2 to TCList; S306, TCCounter1=TCCounter1+1; S307, if TCCounter1 is less than or equal to 20, go to S303; otherwise, go to S308. S308, establish the multi-dimensional collaborative feature output result TCOutput for the critical growth period. The structure of TCOutput is the same as the structure of an element of SFList. S309, TCOutput's F1 field = TCList; S310, the F2 field of TCOutput calculates the first-order difference feature of TCList. This feature is defined as a first-order feature, and its physical meaning is to characterize the growth rate. It is calculated using the following formula: Let the first-order difference feature in TCList be... The elements are , From 0 to 19; , and ; S311, the F3 field of TCOutput is equal to the calculated second-order difference feature of TCList. This feature is defined as a second-order feature, and its physical meaning is to characterize the growth acceleration. It is calculated using the following formula: , and ; S312, the F4 field of TCOutput is equal to the calculated third-order difference feature of TCList. This feature is defined as a third-order feature, and its physical meaning is to characterize the growth inflection point. It is calculated using the following formula: , and ; S313, the F5 field of TCOutput is equal to the fourth-order difference feature of TCList. This feature is defined as a fourth-order feature, and its physical meaning lies in representing the macroscopic shape. It is calculated using the following formula: , and ; S314, TCOutput's F2 field = TCList - F2 field; S315, TCOutput's F3 field = TCList - F3 field; S316, TCOutput's F4 field = TCList - F4 field; S317, TCOutput's F5 field = TCList - F5 field; S318, the YieldDecision field of TCOutput is equal to 0; S319, output TCOutput as the result of TCModel.
5. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S4 specifically includes: S401, the result processing counter CLCounter equals 1; S402, Result processing: The first temporary variable CLTemp1 is equal to the CLCounter-th element retrieved from RSList; S403, the result processing second temporary variable CLTemp2 is equal to the processing using TCModel, TCModel's TCInput is equal to CLTemp1.RSData, and the processing obtains TCOutput; S404, the value of the CLCounter-th element of SFList is equal to CLTemp2; S405, the YieldDecision of the CLCounter-th element of SFList is equal to the YieldObserved field of the CLCounter-th element of RSList; S406, CLCounter equals CLCounter plus 1; S407, if CLCounter is less than or equal to RSNum, go to S402; otherwise, go to S408. S408, the processing of step S4 is complete.
6. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S5 specifically includes: S501, Establish a hierarchical physical information neural network (PINN) model for yield estimation. This model includes the following structure. The model contains 5 input branches, each corresponding to a feature level F1 to F5; Each branch first passes through a Long Short-Term Memory (LSTM) layer to effectively capture the temporal dependencies and long-term memory within the features of each level. Preferably, the LSTM layer contains 128 hidden units to extract the temporal dependencies of the features at each level and output a 100-dimensional feature vector. The output vectors of the five branches are concatenated along the feature dimension to form a 500-dimensional fused feature vector. The fused feature vector is input into the first one-dimensional convolutional layer to extract local spatial patterns and cross-level associations from the fused high-dimensional features. Preferably, the first one-dimensional convolutional layer contains 64 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function. It is followed by a max pooling layer. Preferably, the pooling window size is 2 and the stride is 2. The pooled features are input into a second one-dimensional convolutional layer to further deepen feature extraction and enhance the non-linear expressive power of the model. Preferably, the second one-dimensional convolutional layer contains 32 convolutional kernels of size 3 with a stride of 1 and uses the ReLU activation function. After the convolutional features are flattened, they are input into a multilayer perceptron (MLP) network as the final regressor, which maps the extracted abstract features to the output value. The network contains three fully connected layers with 256, 128 and 1 neurons respectively. The activation function of the first two layers is ReLU, and the last layer is a linear activation function. The final output is the predicted output. S502, PINN’s five input branches are connected to the F1, F2, F3, F4 and F5 fields of SFList respectively, and PINN’s decision output is used to fit the YieldDecision field of SFList; S503 uses SFList data to train PINN. During training, the loss function consists of both data loss and physical loss terms, enabling PINN to make decisions. Specifically: (1) Define the data loss term Let the total number of training samples be... For the first For each sample, the neural network predicts the output as follows: The actual output was The data loss term is the average of the mean squared errors between the predicted and actual yields for all samples, and the formula is: , (2) Define the physical loss term Introducing light energy utilization theory as a physical constraint; specifically: a. Based on the time series data of the NDVI mean at level F1, for the first... At each time point, the mean NDVI was Estimate each time point using empirical formulas. Photosynthetically active radiation absorption ratio : , in and This is an empirical coefficient; b. Obtain each time point Incident photosynthetic effective radiation ; c. Let the light energy utilization efficiency be... The unit is grams of carbon per megajoule. This parameter can be obtained from relevant agricultural literature, and for soybeans, the value is usually between 1.0 and 2.
0. The cumulative total primary productivity (GPP) over the entire growth period is calculated using the following formula: , d. Let the harvest index be... The unit is dimensionless, and for soybeans it is usually between 0.4 and 0.5, which can be determined through variety information; let the unit conversion factor be... This is used to convert carbon units into dry matter weight and ultimately into yield per acre (kg / acre); it converts GPP into physical yield estimates. The calculation formula is: , e. The physical loss term is the average of the mean squared errors between the neural network-predicted output and the physical estimated output for all samples, and the formula is: , (3) Define the total loss function By weighted fusion of data loss and physical loss, the model can fit real data while adhering to basic physical laws of crop growth, thereby suppressing overfitting and improving generalization ability when samples are sparse; the total loss function formula is: , Preferably, wherein , This weighting ratio can be adjusted according to the amount of data; the less data, the lower the weighting ratio. The weight can be appropriately increased; S504, the processing of step S5 is complete.
7. The soybean yield estimation method based on multidimensional collaborative features and physical information neural networks according to claim 1, characterized in that: Step S6 specifically includes: S601, using a remote sensing data acquisition system to collect time-series multispectral image data TestRS of the plot to be predicted; S602, the first temporary variable PDTemp1 used for judgment is equal to the value of TCModel, TCInput of TCModel is equal to TestRS, and TCOutput is obtained after processing. S603, input PDTemp1 into the trained PINN model, and connect the five input branches of PINN to the F1, F2, F3, F4 and F5 fields of PDTemp1 respectively; S604 is the decision output used to determine whether the second temporary variable PDTemp2 is equal to PINN; S605, output PDTemp2 as the final estimated yield result, then proceed to S606; S606, the patent description process is complete.