Crop yield estimation method and device, electronic equipment and readable storage medium

Abstract

The application discloses a kind of crop yield estimation method, device, electronic equipment and readable storage medium, especially related to the field of intelligent agricultural technology, method includes: according to the multi-source remote sensing data of the region to be measured, the canopy structure parameter, photosynthetic function parameter and backscattering coefficient of target crop are obtained;According to canopy structure parameter, photosynthetic function parameter and backscattering coefficient, the biomass time series data of target crop is reconstructed;Initial model of yield estimation based on deep learning model and fusion target crop growth mechanism is constructed;According to the ground measured data of the region to be measured, the initial model of yield estimation is migrated and updated, and the local yield estimation model of the region to be measured is obtained, wherein the sample number of ground measured data is less than the preset number threshold;The field-level yield of the region to be measured is predicted by the local yield estimation model, and the yield estimation result and confidence interval are output in combination with uncertainty evaluation method.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method, apparatus, electronic device, and readable storage medium for estimating crop yield. Background Technology

[0002] Existing remote sensing estimation schemes for wheat yield have three common drawbacks:

[0003] 1. Limited data and incomplete characterization: Optical remote sensing is susceptible to weather interference; synthetic aperture radar (SAR) responses to biomass are scene-dependent; while solar-induced chlorophyll fluorescence (SIF) is directly related to photosynthesis, its spatial resolution is coarse. A single data source cannot comprehensively and robustly characterize the dynamic processes of crop canopy structure, photosynthetic function, and biomass accumulation.

[0004] 2. Weak model generalization ability: When the model is trained on a large number of ground samples, the accuracy will decrease significantly when applied across regions and years due to differences in climate, varieties and management measures.

[0005] 3. High dependence on calibration samples: When existing methods are deployed across regions, a large number of ground samples (tens to hundreds of quadrats) are usually required for model recalibration or training, resulting in long deployment cycles and high costs. At the same time, a single data source is difficult to stably represent structure-function-biomass processes under complex meteorological and management differences, making it difficult to meet the needs of rapid production estimation in operations.

[0006] Therefore, the industry urgently needs a yield estimation system that can comprehensively utilize the advantages of multi-source remote sensing and achieve rapid and high-precision migration with minimal ground sample support.

[0007] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to address the problem that existing crop yield assessment models have weak generalization ability, rely on a large number of sample calibrations, and cannot achieve fast and high-precision migration on a small number of sample grounds. The invention provides a crop yield estimation method, device, electronic device, and readable storage medium.

[0009] The first aspect of this invention provides a method for estimating crop yield, comprising: acquiring canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of a target crop based on multi-source remote sensing data of a region to be measured; reconstructing time-series biomass data of the target crop based on the canopy structure parameters, photosynthetic function parameters, and backscattering coefficients; constructing an initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop, wherein the initial yield estimation model is jointly trained using a physical constraint loss function to learn a nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop, the physical constraint loss function being constructed based on the consistency between the time-series biomass data and the growth mechanism; performing a migration update on the initial yield estimation model based on ground-measured data of the region to be measured to obtain a localized yield estimation model for the region to be measured, wherein the sample size of the ground-measured data is less than a preset threshold; predicting the field-level yield of the region to be measured using the localized yield estimation model, and outputting the yield estimation result and confidence interval by combining an uncertainty assessment method.

[0010] In one embodiment of the present invention, obtaining the canopy structure parameters, photosynthetic function parameters, and backscattering coefficient of the target crop based on multi-source remote sensing data of the area to be measured includes: inverting the multi-source remote sensing data to obtain the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficient. The inversion of the multi-source remote sensing data includes: determining the canopy structure parameters based on LiDAR point cloud data and optical image data of the area to be measured; determining the photosynthetic function parameters based on solar-induced chlorophyll fluorescence (SIF) data of the area to be measured; and determining the backscattering coefficient based on synthetic aperture radar (SAR) data.

[0011] In one embodiment of the present invention, the step of reconstructing the biomass time-series data of the target crop based on the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficient includes: fusing the photosynthetic function parameters and the canopy structure parameters through a spatiotemporal adaptive reflectance fusion model or a deep super-resolution network to generate the biomass time-series data.

[0012] In one embodiment of the present invention, the physical constraint loss function includes physical constraint terms including: a mechanism consistency constraint term, used to constrain the difference between the biomass predicted by the initial model of yield estimation and the prior biomass generated based on the growth mechanism, wherein the growth mechanism includes energy conversion law, phenological development law and harvest index conversion law; and a temporal smoothing and gradient constraint term, used to constrain the gradient of biomass change over time predicted by the initial model of yield estimation.

[0013] In one embodiment of the present invention, the step of migrating and updating the initial yield estimation model based on ground-measured data of the area to be measured includes at least one of the following: determining linear correction coefficients based on the ground-measured data and the predicted values ​​output by the initial yield estimation model, and performing weighted correction on the output of the initial model based on the linear correction coefficients; modeling the model parameters of the area to be measured as the sum of global parameters and random bias terms, updating the posterior distribution of the random bias terms based on the ground-measured data, and performing probability sampling prediction based on the updated posterior distribution; freezing the network parameters of a preset layer in the initial yield estimation model, updating the parameters of the unfrozen layers using gradient descent based on the ground-measured data, and introducing a mechanism consistency regularization during the update process. The process involves: first, constraining the yield estimation model; second, performing gradient updates in a finite number of steps based on the ground-based measured data, and calculating target parameters suitable for the test area; third, acquiring uncertainty indices or information gain indices for each grid point within the test area, and selecting points that meet preset conditions as the data collection targets based on the uncertainty indices of each grid point; fourth, updating the yield estimation model based on the measured data corresponding to the data collection targets; and fifth, generating virtual biomass samples with weakly labeled weights based on the crop growth mechanism model and the meteorological data of the test area, mixing the virtual biomass samples with the ground-based measured data, and jointly training the yield estimation model.

[0014] In one embodiment of the present invention, determining the linear correction coefficient based on the ground-measured data and the predicted value output by the initial model for yield estimation includes: calculating the linear correction coefficient using the total least squares method based on the ground-measured data and the predicted value.

[0015] In one embodiment of the present invention, the step of modeling the model parameters of the area to be measured as the sum of global parameters and random bias terms, and updating the posterior distribution of the random bias terms based on the ground measured data, includes: setting the prior distribution of the random bias terms as a normal distribution; and updating the posterior distribution based on Bayes' theorem.

[0016] A second aspect of the present invention provides a crop yield estimation device, comprising: an acquisition module for acquiring canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of a target crop based on multi-source remote sensing data of a region to be measured; a first processing module for reconstructing time-series biomass data of the target crop based on the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficients; and a second processing module for constructing an initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop, wherein the initial yield estimation model is jointly trained using a physical constraint loss function to learn the multi-source remote sensing data. The nonlinear mapping from the input features corresponding to the sensing data to the crop yield of the target crop, wherein the physical constraint loss function is constructed based on the consistency between the biomass time-series data and the growth mechanism; the update module is used to perform a migration update on the initial yield estimation model based on the ground measured data of the test area to obtain a localized yield estimation model for the test area, wherein the sample size of the ground measured data is less than a preset quantity threshold; the prediction module is used to predict the field-level yield of the test area through the localized yield estimation model, and output the yield estimation result and confidence interval in combination with the uncertainty assessment method.

[0017] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the crop yield estimation method as described in the first aspect.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the crop yield estimation method as described in the first aspect.

[0019] Compared with the prior art, the technical effects achieved by the present invention are as follows: Based on multi-source remote sensing data of the area to be measured, canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of the target crop are obtained. Based on these parameters, biomass time-series data of the target crop is reconstructed to determine data continuity under complex weather conditions. An initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop is constructed. This initial model is jointly trained using a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between the biomass time-series data and the growth mechanism. The initial yield estimation model is then updated using ground-based measured data of the area to be measured to obtain a localized yield estimation model for the area. The sample size of the ground-based measured data is less than a preset threshold. The localized yield estimation model is used to predict the field-level yield of the area to be measured, and the yield estimation results and confidence intervals are output using an uncertainty assessment method. By reducing the required calibration samples for the localized yield estimation model corresponding to the test area to a predetermined number, the generalization ability of the yield estimation model is improved, enabling rapid yield prediction of the target crop. This solves the problem that existing crop yield assessment models have weak generalization ability, rely on a large number of calibration samples, and cannot achieve rapid and high-precision transfer on a small number of sample areas. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for estimating crop yield according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of a crop yield estimation device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the frame of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0021] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0022] The technical solution of the present invention is illustrated below through specific embodiments. It should be understood that the one or more steps mentioned in the present invention do not preclude the existence of other methods and steps before or after the combined steps, or that other methods and steps may be inserted between these explicitly mentioned steps. It should also be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention. Unless otherwise stated, the numbering of each method step is only for the purpose of identifying each method step, and not for limiting the order of each method or limiting the scope of the present invention. Changes or adjustments to their relative relationships, without substantial changes to the technical content, can also be considered as within the scope of the present invention.

[0023] According to a first aspect of the present invention, a method for estimating crop yield is provided, such as... Figure 1 As shown, the method may specifically include the following steps: S101. Based on multi-source remote sensing data of the area to be measured, obtain the canopy structure parameters, photosynthetic function parameters and backscattering coefficient of the target crop. S102, based on canopy structure parameters, photosynthetic function parameters and backscattering coefficient, reconstruct the time series data of biomass of the target crop; S103, Construct an initial yield estimation model based on a deep learning model and integrating the growth mechanism of the target crop. The initial yield estimation model is jointly trained through a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between biomass time-series data and growth mechanism. S104. Based on the ground measured data of the area to be measured, the initial yield estimation model is migrated and updated to obtain the localized yield estimation model of the area to be measured, wherein the sample size of the ground measured data is less than the preset quantity threshold. S105 predicts field-level yields in the area to be tested using a localized yield estimation model, and outputs yield estimation results and confidence intervals by combining uncertainty assessment methods.

[0024] In this embodiment, multi-source remote sensing data includes, but is not limited to, LiDAR point cloud data, optical image data, thermal infrared image data, interferometric synthetic aperture radar (InSAR) data, and synthetic aperture radar (SAR) data.

[0025] By collecting multi-source remote sensing data of the area to be measured, parameter inversion is performed on the multi-source remote sensing data to obtain the canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of the target crop in the area to be measured. Canopy structure parameters include, but are not limited to, plant height, canopy width, leaf area index, and vegetation cover of the target crop.

[0026] In practical applications, there are two key data sources among the multi-source remote sensing data of the area to be measured. One type of data is functional data, which is high temporal resolution data, such as SIF data, with a frequency of daily or every few hours. However, its spatial resolution, for example, the scale corresponding to one pixel is tens of kilometers, making it impossible to determine the details of each field in the area to be measured. The other type of data is structural data, which is high spatial resolution data, including LiDAR point cloud data, SAR data, and high-resolution optical imagery. It has high spatial resolution but a low temporal frequency or is easily affected by cloud and rain weather.

[0027] To address the aforementioned issues, in this embodiment, biomass time-series data of the target crop is reconstructed using a preset fusion algorithm based on canopy structure parameters, photosynthetic function parameters, and backscattering coefficients. The spatial and temporal resolutions of the reconstructed biomass time-series data satisfy preset conditions. The preset fusion algorithm includes, but is not limited to, the spatiotemporal adaptive reflectivity fusion model STARFM and the Deep Super-Resolution Network.

[0028] Next, based on the input features corresponding to the multi-source sensor data, a preliminary yield extrapolation of the target crop is performed. The underlying architecture of the model can be an existing deep learning model, such as Transformer, Physical Information Neural Network (PINN), or Gradient Boosting Decision Tree (GBDT). Simultaneously, a physical constraint loss function is constructed using the reconstructed biomass time-series data and the growth mechanism of the target crop.

[0029] In this embodiment, the biological mechanism includes the physical boundaries of the target crop and prior knowledge, such as energy conversion, phenological constraints, and the harvest index HI. Phenological constraints ensure that the biomass of the target crop changes over time conforms to its growth curve, while the harvest index HI ensures that the yield of the target crop remains within a reasonable range. Then, the initial yield estimation model is jointly trained using a physical constraint loss function, and the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop is used to obtain the initial yield estimation model.

[0030] In this embodiment, the initial yield estimation model is migrated and updated based on the ground measured data of the area to be measured to obtain a localized yield estimation model for the area to be measured. The sample size of the ground measured data is less than a preset threshold, which can be set to 10, 15, etc., but is not limited in this embodiment.

[0031] In one example, a few-sample rapid calibration strategy reduces the number of regional calibration samples required for the test area from dozens to single digits, and provides plot-level uncertainty maps and reports. Few-sample rapid calibration strategies include, but are not limited to, prior offset correction, hierarchical Bayesian methods, layer selection fine-tuning, meta-learning rapid adaptation, active sampling, and conformal / quantile intervals.

[0032] Finally, the plot-level yield of the test area is predicted using a localized yield estimation model, and the yield estimation results and confidence intervals are output using an uncertainty assessment method. Specifically, multi-source sensor data of the test area are input into the localized yield estimation model to predict the yield per unit area of ​​each grid or plot in the test area and the corresponding confidence level.

[0033] In this embodiment, uncertainty assessment can be performed using hierarchical Bayesian sampling, quantile evaluation, and conformal prediction to provide a confidence interval for the prediction results of the localized yield estimation model.

[0034] This invention, through its embodiments, acquires canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of a target crop based on multi-source remote sensing data of the area to be measured. It then reconstructs the biomass time-series data of the target crop based on these parameters to determine data continuity under complex weather conditions. An initial yield estimation model is constructed based on a deep learning model and integrates the growth mechanism of the target crop. This initial model is jointly trained using a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between the biomass time-series data and the growth mechanism. The initial yield estimation model is then updated using ground-based measured data of the area to be measured to obtain a localized yield estimation model for the area. The sample size of the ground-based measured data is less than a preset threshold. Finally, the localized yield estimation model is used to predict the field-level yield of the area to be measured, and the yield estimation results and confidence intervals are output using an uncertainty assessment method. By reducing the required calibration samples for the localized yield estimation model corresponding to the test area to a predetermined number, the generalization ability of the yield estimation model is improved, enabling rapid yield prediction of the target crop. This solves the problem that existing crop yield assessment models have weak generalization ability, rely on a large number of calibration samples, and cannot achieve rapid and high-precision transfer on a small number of sample areas.

[0035] Optionally, in this embodiment, step S101 includes, but is not limited to: inverting multi-source remote sensing data to obtain canopy structure parameters, photosynthetic function parameters, and backscattering coefficients. The inversion of the multi-source remote sensing data includes: determining canopy structure parameters based on LiDAR point cloud data and optical image data of the area to be measured; determining photosynthetic function parameters based on solar-induced chlorophyll fluorescence (SIF) data of the area to be measured; and determining backscattering coefficients based on synthetic aperture radar (SAR) data.

[0036] Specifically, by using LiDAR point cloud data and calculating the difference between the digital surface model (DSM) and the digital terrain model (DTM), a canopy height model (CHM) is generated, thereby accurately extracting wheat plant height information.

[0037] Using optical imagery data (such as multispectral or hyperspectral imagery), vegetation indices are calculated through band calculations, and leaf area index (LAI), canopy width, and vegetation cover are retrieved using a radiative transfer model. Fluorescence radiation intensity in specific bands (such as the O2-A band near 760 nm) is extracted from SIF data acquired by a hyperspectral sensor. Backscattering coefficients under VV polarization (vertical emission-vertical reception) and VH polarization (vertical emission-horizontal reception) modes are extracted from SAR data after radiometric calibration and filtering.

[0038] In one example, taking a winter wheat field in the jointing stage as the test area, a LiDAR-equipped drone was first used to fly over the field, measuring the average plant height as 45 cm. Simultaneously, optical images of the area were acquired, and the leaf area index (LAI) was retrieved as 3.5. For photosynthetic function, SIF data of the test area was acquired, showing high fluorescence intensity, indicating vigorous wheat growth. To address the lack of optical data during cloudy or rainy days, SAR data was used, and the VH polarization backscattering coefficient was extracted as -12 dB to assist in correcting canopy structure information.

[0039] Through the above embodiments, multi-source remote sensing data are inverted to obtain canopy structure parameters, photosynthetic function parameters, and backscattering coefficients. These parameters together constitute a multidimensional feature set that includes both three-dimensional morphology and physiological activity.

[0040] Optionally, in this embodiment, step S102 includes, but is not limited to: fusing photosynthetic function parameters and canopy structure parameters through a spatiotemporal adaptive reflectivity fusion model or a deep super-resolution network to generate biomass time-series data.

[0041] In practical applications, SIF data typically has high spatiotemporal resolution (high frequency, low resolution, e.g., daily or multi-day data) and is used to capture rapid changes in wheat growth trends. LiDAR and high-resolution optical imagery have high spatial resolution (low frequency, high resolution) and are used to characterize field boundaries and the spatial heterogeneity of crop growth. Therefore, in this embodiment, a spatiotemporal adaptive reflectivity fusion model (STARFM) or a deep super-resolution network (such as SRResNet) is employed. Using high SIF data as the temporal variation benchmark and canopy structure parameters as the spatial texture benchmark, the temporal variation information of the SIF data is mapped onto the pixels of the high-resolution imagery to reconstruct the biomass time-series data of the target crop.

[0042] In one example, suppose we need to estimate wheat biomass on April 10th (cloudy, no optical satellite data). We acquire high-resolution optical imagery (high spatial resolution) from April 1st and continuous SIF data (high temporal resolution) from April 1st to April 10th. Using a pre-trained deep learning model (e.g., a deep super-resolution network), we learn the mapping between optical texture and SIF intensity on April 1st. Then, we input the SIF data from April 10th into the deep learning model. By learning the spatial texture structure from April 1st, we fill the high-resolution grid with the photosynthetic physiological changes of April 10th, generating a biomass distribution map that includes both the physiological state on April 10th and the spatial details of April 1st.

[0043] Optionally, in this embodiment, the physical constraint loss function includes the following physical constraint terms: The mechanism consistency constraint term is used to constrain the difference between the biomass predicted by the initial model for yield estimation and the prior biomass generated based on the growth mechanism, where the growth mechanism includes the laws of energy conversion, the laws of phenological development, and the laws of harvest index conversion. Specifically, the mechanism consistency constraints include, but are not limited to, energy conversion law constraints and harvest index conversion law constraints. Energy conversion law constraints ensure that the accumulation of the target crop should be approximately equal to the product of light energy utilization efficiency and absorbed photosynthetically active radiation. Harvest index conversion law constraints stipulate that the HI (the ratio of the target crop's final yield to its biomass at maturity) should be within a biologically reasonable range; for example, the HI for wheat is typically between 0.35 and 0.55.

[0044] Specifically, the physical constraint loss function L The mathematical expression is as follows:

[0045] in, To predict output for the model, To measure the actual yield, The biomass time series output by the model. For mechanism-based biomass priors, Representing the time gradient of biomass, For growth rate functions based on meteorology and management, , These are the weighting coefficients.

[0046] In one example, during the training of the initial yield estimation model, the predicted yield for a certain plot of land is 750 kg / mu based on the input remote sensing features. However, the mechanistic model, based on local meteorological data (sunlight, temperature), calculates that the theoretical upper limit for this wheat variety under ideal conditions is only 600 kg / mu. At this point, the mechanistic consistency constraint term generates a large loss value, forcing the initial yield estimation model to adjust its parameters during backpropagation, thus lowering the predicted value.

[0047] Temporal smoothing and gradient constraint terms are used to constrain the gradient of biomass change over time predicted by the initial model for yield estimation. Specifically, temporal smoothing and gradient constraints are used to constrain the first derivative (growth rate) of the target crop's biomass over time. Crop growth is continuous, and biomass cannot undergo drastic changes in a short period of time (except for harvesting). Temporal smoothing and gradient constraints are used to penalize abnormal fluctuations in the model output.

[0048] In one example, if the initial yield estimation model predicts that wheat biomass on day 5 of the jointing stage suddenly increases by 50% compared to day 4, which violates the natural growth rate, the time-series smoothing and gradient constraint terms will produce a loss value.

[0049] Optionally, in this embodiment, step S104 includes at least one of the following: The linear correction coefficients are determined based on the ground-measured data and the predicted values ​​output by the initial model for yield estimation, and the output results of the initial model are weighted and corrected based on the linear correction coefficients; preferably, the linear correction coefficients are calculated by the total least squares method based on the ground-measured data and the predicted values.

[0050] Specifically, when using linear correction coefficients for migration updates, considering that the ground-based measured data itself also has measurement errors, the total least squares (TLS) method is used to calculate the linear correction coefficients.

[0051] In one example, the predicted yield of a certain plot of land is X, and the actual measured yield is Y. The correction factor is calculated using the total least squares method. k and b ( It can tolerate noise in both the predicted value X and the measured value Y, and calculates a straight line that geometrically minimizes the sum of squared distances to all data points, thus obtaining a more robust linear correction coefficient.

[0052] The model parameters of the area to be tested are modeled as the sum of global parameters and random bias terms. The posterior distribution of the random bias terms is updated based on the ground measured data, and probability sampling prediction is performed based on the updated posterior distribution. In one example, the parameters of the localized production prediction model for the region to be tested are set as follows: The model is as follows:

[0053] in, These are the globally universal parameters trained in the source domain. This is a random bias term specific to the region to be tested.

[0054] Assume the prior distribution is a random bias term. It follows a normal distribution with a mean of 0 and a covariance. The posterior distribution of the random bias term is updated using Bayes' theorem:

[0055] in, These are ground-measured data.

[0056] The network parameters of the preset layer in the initial model for frozen yield estimation are updated by gradient descent based on ground-measured data, and mechanism consistency constraints are introduced during the update process. Specifically, only the last K layers of the initial model (such as the fusion layer and the regression head) are unfrozen in terms of yield estimation, while the remaining layers in the model remain frozen. A small learning rate and L2 regularization are used, and a mechanism consistency constraint term is added to suppress overfitting.

[0057] Based on the initial parameters obtained from the initial model of yield estimation in the source domain multi-task learning, a finite number of gradient updates are performed according to the ground measured data to calculate the target parameters adapted to the test area. Specifically, in this embodiment, a gradient-based meta-learning strategy is used to perform gradient updates in a limited number of steps based on ground-based measured data, such as the Model Independent Meta-Learning (MAML) algorithm.

[0058] Furthermore, the loss function used during layer selection fine-tuning is... Includes regularization terms, as follows:

[0059] in, For the network parameter weights of the unfrozen layers, This is the regularization coefficient, used to suppress overfitting under conditions of few samples.

[0060] In practical applications, the yield data of target crops in different regions over historical years are divided into multiple source domain tasks, and initial parameters sensitive to yield prediction tasks are trained. For the yield estimation task of the test area, the system no longer trains from scratch. Starting with the initial parameters, it performs gradient descent updates using a very small amount of ground-based measured data from the test area to obtain target parameters adapted to that test area, thus obtaining a localized yield estimation model for the test area.

[0061] In one example, suppose that meta-training has been completed based on historical data of a certain area over the past 5 years (2015-2020), and the initialization parameter A has been obtained. If it is necessary to estimate the yield of the target crop in the test area, the measured yield data of 5 representative quadrats are collected in the test area. By loading the initialization parameter A into the initialization yield estimation model, the measured yield data of 5 representative quadrats are used as the model input, the learning rate is set to 0.001, and only 3 gradient updates are performed to obtain the localized yield estimation model of the test area.

[0062] Optionally, in this embodiment, the uncertainty index of each grid point in the area to be measured is obtained, and points that meet the preset conditions are selected as the collection targets of ground measurement data based on the uncertainty index of each grid point. The initial model for yield estimation is updated by collecting the measurement data corresponding to the collection targets. Specifically, the uncertainty index of each grid point within the test area is calculated based on methods such as variance in hierarchical Bayesian inference. Based on the principle of maximizing information gain, points meeting preset conditions are selected as the targets for collecting ground-based measured data. These preset conditions include, but are not limited to, model prediction accuracy being less than a preset accuracy threshold, and / or, geographic spatial distribution being greater than a preset density threshold. Then, the measured data corresponding to the collection targets are acquired, and the initial yield model is updated based on this measured data.

[0063] Based on the crop growth mechanism model and meteorological data of the area to be measured, virtual biomass samples with weak label weights are generated. The virtual biomass samples are then mixed with ground-measured data to jointly train the initial model for yield estimation.

[0064] When the amount of measured yield data for the target area is small, such as fewer than three samples, it can easily lead to model overfitting. To address this issue, in this embodiment, daily meteorological data (temperature, radiation, precipitation) and basic physiological parameters of the crop variety are input into the target area. The mechanistic model then generates a biomass time series (growth curve) corresponding to the target crop. A certain amount of random noise is added based on the biomass time series to generate dozens of virtual biomass samples. These virtual biomass samples are then mixed with ground-based measured data to jointly train the initial yield estimation model.

[0065] In one example, it is assumed that ground-based data from only two ground measurement points can be obtained for a certain area, which is insufficient to meet the minimum training requirements of the neural network. Weather station data for the area over the past three months is obtained, and a wheat mechanism model is used to calculate that, under these meteorological conditions, the theoretical biomass at the jointing stage should be between 150 and 400 grams per square meter. Fifty virtual biomass data points are generated.

[0066] A second aspect of this application also provides a crop yield estimation device, such as... Figure 2 As shown, the device may specifically include: 1) Acquisition module 20 is used to acquire canopy structure parameters, photosynthetic function parameters and backscattering coefficients of target crops based on multi-source remote sensing data of the area to be measured; 2) The first processing module 22 is used to reconstruct the biomass time-series data of the target crop based on the canopy structure parameters, the photosynthetic function parameters and the backscattering coefficient; 3) The second processing module 24 is used to construct an initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop. The initial yield estimation model is jointly trained through a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between the biomass time series data and the growth mechanism. 4) Update module 26 is used to migrate and update the initial production estimation model based on the ground measured data of the area to be tested, so as to obtain a localized production estimation model for the area to be tested, wherein the sample size of the ground measured data is less than a preset quantity threshold. 5) Prediction module 28 is used to predict the field-level yield of the area to be tested through the localized yield estimation model, and output the yield estimation results and confidence intervals in combination with the uncertainty assessment method.

[0067] Optionally, in this embodiment, the acquisition module 20 includes: The first processing submodule is used to invert the multi-source remote sensing data to obtain the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficient, wherein, The first processing submodule includes: The first determining unit is used to determine the canopy structure parameters based on the LiDAR point cloud data and optical image data of the area to be tested. The second determining unit is used to determine the photosynthetic function parameters based on the solar-induced chlorophyll fluorescence (SIF) data of the area to be tested. The third determining unit is used to determine the backscattering coefficient based on synthetic aperture radar (SAR) data.

[0068] Optionally, in this embodiment, the first processing module 22 includes: The second processing submodule is used to characterize the temporal variation trend of crop growth based on the photosynthetic function parameters. The third processing submodule is used to characterize the spatial details of the crop canopy based on the canopy structure parameters and the backscattering coefficient. The fourth processing submodule is used to fuse the photosynthetic function parameters and the canopy structure parameters using a spatiotemporal adaptive reflectivity fusion model or a deep super-resolution network to generate the biomass time-series data.

[0069] Optionally, in this embodiment, the physical constraint loss function includes the following physical constraint terms: The mechanism consistency constraint term is used to constrain the difference between the biomass predicted by the initial model for yield estimation and the prior biomass generated based on the growth mechanism, wherein the growth mechanism includes energy conversion law, phenological development law and harvest index conversion law; The time-series smoothing and gradient constraint terms are used to constrain the gradient of biomass change over time predicted by the initial model for yield estimation.

[0070] Optionally, in this embodiment, the update module 26 includes at least one of the following: The first update submodule is used to determine the linear correction coefficient based on the ground measured data and the predicted value output by the initial model of yield estimation, and to perform weighted correction on the output of the initial model based on the linear correction coefficient. The second update submodule is used to model the model parameters of the area to be tested as the sum of global parameters and random bias terms, update the posterior distribution of the random bias terms according to the ground measured data, and perform probability sampling prediction based on the updated posterior distribution. The third update submodule is used to freeze the network parameters of the preset layer in the initial model of the yield estimation, update the parameters of the unfrozen layer by gradient descent according to the ground measured data, and introduce mechanism consistency constraints in the update process. The fourth update submodule is used to perform a finite number of gradient updates based on the initial parameters obtained by the initial model in the source domain multi-task learning based on the ground measured data, and calculate the target parameters adapted to the test area. The fifth update submodule is used to obtain the uncertainty index of each grid point in the area to be measured, and to select points that meet the preset conditions as the collection targets of the ground measured data according to the uncertainty index of each grid point, and to update the yield estimation initial model through the measured data corresponding to the collection targets. The sixth update submodule is used to generate a virtual biomass sample with weak label weights based on the crop growth mechanism model and the meteorological data of the area to be measured, and to mix the virtual biomass sample with the ground measured data to jointly train the yield estimation initial model.

[0071] Optionally, in this embodiment, determining the linear correction coefficient based on the measured ground data and the predicted value output by the initial model for yield estimation includes: calculating the linear correction coefficient using the total least squares method based on the measured ground data and the predicted value.

[0072] Optionally, in this embodiment, the step of modeling the model parameters of the area to be tested as the sum of global parameters and random bias terms, and updating the posterior distribution of the random bias terms based on the ground-measured data, includes: setting the prior distribution of the random bias terms as a normal distribution; and updating the posterior distribution based on Bayes' theorem.

[0073] A third aspect of this application also provides an electronic device, such as... Figure 3 As shown, it includes: at least one processor 301; and a memory 302 communicatively connected to the at least one processor 301; wherein the memory 302 stores instructions executable by the at least one processor 301, the instructions being executed by the at least one processor 301 to enable the at least one processor 301 to perform the crop yield estimation methods in the above embodiments.

[0074] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0075] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0076] A fourth aspect of this application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements an embodiment of the above-described crop yield estimation method.

[0077] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means implemented in one or more flows and / or blocks of a flowchart. Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing one or more processes and / or blocks in the flowchart. Figure 1 The steps of the function specified in one or more boxes.

[0080] The foregoing description of specific exemplary embodiments of this application is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for estimating crop yield, characterized in that, include: Based on multi-source remote sensing data of the area to be measured, obtain the canopy structure parameters, photosynthetic function parameters and backscattering coefficient of the target crop; Based on the canopy structure parameters, photosynthetic function parameters, and backscattering coefficient, the biomass time-series data of the target crop are reconstructed. An initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop is constructed. The initial yield estimation model is jointly trained through a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between the biomass time-series data and the growth mechanism. Based on the ground-measured data of the area to be tested, the initial production estimation model is migrated and updated to obtain a localized production estimation model for the area to be tested, wherein the sample size of the ground-measured data is less than a preset quantity threshold. The localized yield estimation model is used to predict the field-level yield of the area to be tested, and the yield estimation results and confidence intervals are output by combining the uncertainty assessment method. The physical constraint loss function includes the following physical constraint terms: The mechanism consistency constraint term is used to constrain the difference between the biomass predicted by the initial model for yield estimation and the prior biomass generated based on the growth mechanism, wherein the growth mechanism includes energy conversion law, phenological development law and harvest index conversion law; The time-series smoothing and gradient constraint terms are used to constrain the gradient of biomass change over time predicted by the initial model for yield estimation.

2. The method according to claim 1, characterized in that, The process of acquiring canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of the target crop based on multi-source remote sensing data of the area to be measured includes: The multi-source remote sensing data is inverted to obtain the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficient, wherein, Inversion of the multi-source remote sensing data includes: The canopy structure parameters are determined based on the LiDAR point cloud data and optical image data of the area to be tested. The photosynthetic function parameters were determined based on the solar-induced chlorophyll fluorescence (SIF) data of the area to be tested. The backscattering coefficients are determined based on synthetic aperture radar (SAR) data.

3. The method according to claim 1, characterized in that, The process of reconstructing the biomass time-series data of the target crop based on the canopy structure parameters, the photosynthetic function parameters, and the backscattering coefficient includes: The photosynthetic function parameters and the canopy structure parameters are fused using a spatiotemporal adaptive reflectance fusion model or a deep super-resolution network to generate the biomass time-series data.

4. The method according to claim 1, characterized in that, The migration update of the initial yield estimation model based on the ground-measured data of the area to be measured includes at least one of the following: The linear correction coefficients are determined based on the ground-measured data and the predicted values ​​output by the initial production estimation model, and the output of the initial model is weighted and corrected based on the linear correction coefficients. The model parameters of the area to be tested are modeled as the sum of global parameters and random bias terms. The posterior distribution of the random bias terms is updated based on the ground measured data, and probability sampling prediction is performed based on the updated posterior distribution. The network parameters of the preset layer in the initial model for yield estimation are frozen, and the parameters of the unfrozen layer are updated by gradient descent based on the ground measured data, with mechanism consistency constraints introduced during the update process. Based on the initial parameters obtained by the initial model for yield estimation in the source domain multi-task learning, a finite number of gradient updates are performed according to the ground measured data to calculate the target parameters adapted to the test area. The uncertainty index of each grid point in the area to be measured is obtained, and points that meet the preset conditions are selected as the collection targets of the ground measured data based on the uncertainty index of each grid point. The initial model of yield estimation is updated by the measured data corresponding to the collection targets. Based on the crop growth mechanism model and the meteorological data of the area to be measured, a virtual biomass sample with weak label weights is generated. The virtual biomass sample is then mixed with the ground-measured data to jointly train the initial model for yield estimation.

5. The method according to claim 4, characterized in that, The step of determining the linear correction coefficient based on the ground-measured data and the predicted value output by the initial yield estimation model includes: The linear correction coefficient is calculated using the total least squares method based on the measured ground data and the predicted value.

6. The method according to claim 4, characterized in that, The step of modeling the parameters of the area to be measured as the sum of global parameters and random bias terms, and updating the posterior distribution of the random bias terms based on the ground-measured data, includes: The prior distribution of the random bias term is set to a normal distribution; The posterior distribution is updated based on Bayes' theorem.

7. A crop yield estimation device, characterized in that, include: The acquisition module is used to acquire canopy structure parameters, photosynthetic function parameters, and backscattering coefficients of the target crop based on multi-source remote sensing data of the area to be measured. The first processing module is used to reconstruct the biomass time-series data of the target crop based on the canopy structure parameters, the photosynthetic function parameters and the backscattering coefficient; The second processing module is used to construct an initial yield estimation model based on a deep learning model and incorporating the growth mechanism of the target crop. The initial yield estimation model is jointly trained through a physical constraint loss function to learn the nonlinear mapping from the input features corresponding to the multi-source remote sensing data to the crop yield of the target crop. The physical constraint loss function is constructed based on the consistency between the biomass time series data and the growth mechanism. The update module is used to migrate and update the initial yield estimation model based on the ground measured data of the area to be tested, so as to obtain a localized yield estimation model for the area to be tested, wherein the sample size of the ground measured data is less than a preset quantity threshold. The prediction module is used to predict the field-level yield of the area to be tested using the localized yield estimation model, and output the yield estimation results and confidence intervals in combination with the uncertainty assessment method. The physical constraint loss function includes the following physical constraint terms: The mechanism consistency constraint term is used to constrain the difference between the biomass predicted by the initial model for yield estimation and the prior biomass generated based on the growth mechanism, wherein the growth mechanism includes energy conversion law, phenological development law and harvest index conversion law; The time-series smoothing and gradient constraint terms are used to constrain the gradient of biomass change over time predicted by the initial model for yield estimation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the crop yield estimation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the crop yield estimation method as described in any one of claims 1 to 6.

Images