A hierarchical missing perception rice yield prediction method based on an agronomic development chain

Abstract

The application provides a hierarchical missing perception rice yield prediction method based on an agronomic development chain, and belongs to the technical field of rice yield prediction. In stage one, a development mapping is trained by using double-period images of a green-up tillering period and a heading and fruiting period without a yield tag, so as to reduce the dependence on an expensive yield tag. In stage two, yield regression is jointly trained in multiple paths on a small amount of yield tags, and two-stage loss is separated and logic is clear. The green-up tillering period before row closing is selected, so that the extraction accuracy of density characteristics is fundamentally guaranteed. The technical bottleneck that the existing yield prediction model directly fails when time series data is interrupted is broken through. Path 3 is used as a full model, paths 1, 2 and 4 are used as three sub-models, a unified dimension feature compensation mechanism is combined, all physical missing states that may be encountered in field collection are completely covered, and strong system robustness is realized without increasing the model size.

Description

Technical Field

[0001] This invention belongs to the field of rice yield prediction technology and also to the interdisciplinary field of smart agriculture and computer vision. Specifically, it relates to a hierarchical missing perception method for rice yield prediction based on the agronomic development chain. Background Technology

[0002] Current deep learning methods for predicting rice yield mainly fall into two categories: Type 1 is yield estimation based on single-period images of the heading and grain-filling stage, which uses canopy images of the heading and grain-filling stage collected by UAVs as the sole input. Deep learning is used to extract features such as the number and size of rice panicles to directly predict the yield. This type of method is the mainstream research direction at present. Type 2 is yield estimation based on multi-time series remote sensing images, which uses satellite or UAV images of the entire rice growth period at multiple time points to extract time series vegetation index features and then inputs them into a sequence model to predict the yield. This requires complete and usable data for the entire growth period. However, in hilly and mountainous areas, the weather is unstable, and rain can lead to missing images. Long-term data acquisition results in high manpower consumption, large data processing volume, and high costs.

[0003] The aforementioned deep learning method for predicting rice yield has the following drawbacks:

[0004] Existing deep learning methods for predicting rice yield generally focus on images from the heading and grain-filling stage, implicitly assuming that the number of panicles directly reflects yield. However, the formula for rice yield composition shows that:

[0005] ;

[0006] Number of plants per unit area Planting density is the upper limit determinant of the effective panicle number and is an independent multiplicative factor in the yield formula. Existing deep learning methods for predicting rice yield only show how many panicles have been achieved, but ignore how many panicles could have been achieved based on the density, systematically losing information on the upper limit of production potential, leading to a decrease in prediction accuracy under different planting density patterns. The heading and grain-filling period (July-August) is often a period of heavy rain, drought, and fog in hilly and mountainous areas, with extremely limited drone flight windows. This is the most critical but also the most difficult period to acquire images. Existing deep learning methods for predicting rice yield lack any degradation inference strategies for missing data during the heading and grain-filling period, and therefore fail directly in this situation. Multi-time series methods have high acquisition costs and limited practicality. The manpower and equipment costs of multiple acquisitions throughout the entire growth period increase linearly with the number of acquisitions, and the loss of images at any time period leads to model failure. Satellite time series imagery in Southwest China is severely affected by cloud interference, limiting its practical application. When used on a large scale, the probability of missing images from multiple periods is relatively high.

[0007] In summary, existing deep learning methods for predicting rice yield do not explicitly model the key agronomical development chain from planting density to the number of effective panicles, and implicitly assume that key period images are complete and available, lacking hierarchical fault-tolerant inference capabilities. Existing technologies suffer from severe degradation when dealing with missing time-series data: once intermediate period data is missing, current practices either directly discard the field sample (leading to a sharp reduction in usable data), use mean interpolation to complete the data (introducing significant noise), or degenerate into training multiple independent single-period models (severing the temporal correlation). Existing methods lack a fine-grained fault-tolerant architecture that can dynamically and adaptively switch inference logic based on the data missing status and the reliability of prior features, failing to meet the practical deployment needs under complex agricultural climatic conditions. Summary of the Invention

[0008] To address the limitations of existing methods in acquiring incomplete field drone images, this paper proposes a four-path hierarchical missing-aware yield inference framework. This framework explicitly constructs an agronomic development chain mapping between planting density features and rice panicle features. It unifies the spatial dimensions of each path feature using a learnable missing-aware weight matrix and conducts joint training with shared parameters. Combined with small-sample regression of measured yield data using a five-point sampling method, this approach enables accurate field-level rice yield prediction.

[0009] The present invention employs the following technical solution.

[0010] A hierarchical missing-sensing method for rice yield prediction based on agronomic developmental chains includes:

[0011] The process consists of two training phases and one inference phase, executed sequentially. The two training phases are Phase 1 and Phase 2, respectively.

[0012] Phase 1 involves training a developmental mapping network on fields where both phase 1 and phase 2 images are available. This network maps the density feature vector from the tillering stage to the panicle feature vector at the heading and grain-filling stage. The training loss includes a feature reconstruction term and an agronomic upper bound constraint term. The agronomic upper bound constraint term penalizes cases where the number of effective panicles predicted by the mapping network exceeds the product of the number of plants per unit area and the upper bound of the maximum number of effective tillers per plant. For fields where both phase 1 and phase 2 data are available, the difference between the actual panicle features and the mapped prediction values ​​is calculated as the developmental bias vector. After training, the parameters of the developmental mapping network are frozen.

[0013] Phase two involves using all fields with true yield values ​​as training data. Based on the data availability indicators and development mapping reliability of each field, each field is automatically routed to one of four corresponding paths. After unifying the features to the same dimension through a path-specific learnable missing value compensation weight matrix, the features are input into a yield regression model with shared parameters for the four paths. The weighted sum of the losses of the four paths constitutes the joint training objective. After each mini-batch mixed sampling, a backpropagation is performed to update the shared regression parameters and the compensation matrix of the corresponding path. The true yield values ​​are obtained by a five-point sampling method, selecting five 1m samples from each field in a quincunx pattern. 2 The average value was calculated after the actual measurements were taken for each quadrat.

[0014] The inference phase includes selecting the corresponding path for new fields based on data availability indicators and development mapping reliability, and outputting the yield prediction value.

[0015] Preferably, in Stage 1, density images are collected during the tillering stage from transplanting to before canopy closure.

[0016] Preferred, prior to stage one, specifically includes:

[0017] Drone data collection was conducted only during two key growth stages: density images were collected during the tillering and greening stage, and panicle images were collected during the heading and grain-filling stage. Effective panicle number and grain number characteristics were extracted, and data availability indicators for each field were recorded. Any period can be missing. These are the data availability indicators for the tillering stage and the heading and grain-filling stage, respectively.

[0018] For fields that meet the conditions, the five-point sampling method is used to obtain the true yield value. (kg / mu) serves as the sole supervisory label for the stage two regression model. Specifically, five 1m² quadrats are selected from each field in a quincunx pattern. The crops are harvested, threshed, dried, and weighed at each quadrat. The average yield of the five quadrats is then converted to the yield per unit area. This true yield value is only used during the training phase; the predicted value can be output during the inference phase without requiring actual yield measurements.

[0019] The dataset is classified as follows:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] in, A set of fields for which bi-period images are available and without yield labels; A set of fields for which bi-period images are available and which have yield labels; This is a collection of fields where only heading and grain-filling period images are available and yield labels are provided; This is a collection of fields for which only images of the tillering stage are available and have yield labels. Images captured by drone during the tillering stage of the i-th field. Images of the i-th field taken by drone during the heading and fruiting stage. Let j be the true value of the measured yield of the j-th field. Images captured by drone during the tillering stage of the j-th field. Images captured by drone during the heading and fruiting stage of the j-th field. The images are taken by drone during the heading and fruiting stage of the k-th field. Let k be the true value of the measured yield of the kth field. Images captured by drone during the tillering stage of the lth field. This represents the true value of the measured yield of the l-th field.

[0025] Preferably, Phase One specifically includes:

[0026] Features from the two periods were extracted using a self-supervised pre-trained encoder on unlabeled field images:

[0027] ;

[0028] ;

[0029] in, This represents the density feature vector during the tillering stage; A feature extractor or encoder for images of the tillering stage; Images captured by drone during the tillering stage; This represents the characteristic vector of rice panicles during the heading and grain-filling stage. A feature extractor or encoder for images during the heading and fruiting stages; Images captured by drone during the heading and fruiting stage.

[0030] Preferably, Phase One also includes:

[0031] exist Training Developmental Mapping Network Do not use production labels:

[0032] ;

[0033] in The training loss of the developmental mapping network M is... This represents the characteristic vector of rice panicles during the heading and grain-filling stage. The predicted spikelet feature vector is output by the developmental mapping network M; This represents the density feature vector during the tillering stage; The weight of the agronomic constraint loss is set to control the proportion of the second term in the total loss (hyperparameter). For the hinge function; To predict rice panicle characteristics The estimated number of effective ears; This serves as an upper bound for the maximum effective tiller number per plant, preventing the mapping function from outputting biologically unreasonable predictions. The actual plant density was obtained from statistical analysis of images during the tillering stage.

[0034] Freeze after training convergence The parameters will no longer be updated.

[0035] Preferably, Phase One also includes:

[0036] right Calculate the developmental deviation vector for each field in the dataset:

[0037] ;

[0038] This is the developmental deviation vector.

[0039] Preferably, stage two includes the construction of a unified feature space along four paths, which is as follows:

[0040] All four paths unify the features to Dimension, input the same output regression model:

[0041] Path 3 is as follows:

[0042] ;

[0043] in This is the fusion feature vector for path 3. To be It is assembled sequentially;

[0044] Path 2 is as follows:

[0045] ;

[0046] in This is the fusion feature vector for path 2; For the order These three vector segments are concatenated in this way; the 0 on the left is the d-dimensional zero vector; the 0 on the right is the d-dimensional zero vector. The learnable compensation weight matrix for path 2;

[0047] Path 4 is as follows:

[0048] ;

[0049] in The feature vector for path 4 is fused; 0 is the d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 4;

[0050] Path 1 is as follows:

[0051] ;

[0052] in 0 represents the fusion feature vector of path 1; 0 represents the d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 1.

[0053] Preferably, stage two also includes four-path weighted joint training, which is as follows:

[0054] The four paths share parameters in the output regression model. Each of the three sub-models has its own independent compensation matrix. The total loss of joint training is a weighted sum of the losses of the four paths:

[0055] ;

[0056] in for of ; For setting of Set of for path of The key is to check whether the mapping is reliable. for ; for ; for path ; for of .

[0057] Preferably, the specific implementation method of joint training in Phase Two is as follows:

[0058] Each training mini-batch from , , Three datasets are sampled proportionally, and each sample is automatically routed to a corresponding path based on its data availability flag, and the feature vector of that path is calculated. Post-input shared model , to obtain the predicted value .

[0059] Preferably, the yield regression model described in Phase 2 is a random forest or gradient boosting tree, to adapt to the small sample constraint that the number of yield-labeled fields is usually less than 50; Path 1 and Path 4 use the same training data. When the mapping reliability switching mechanism is not activated simultaneously, the same batch of data is only routed to one of the paths, contributing to the loss.

[0060] A hierarchical missing-sensing rice yield prediction system based on agronomic developmental chains includes:

[0061] The data acquisition module is used to acquire drone images during the tillering and heading stages, record data availability indicators, and acquire true yield values ​​using the five-point sampling method.

[0062] The feature extraction module is used to extract density feature vectors and rice spike feature vectors from the two images, respectively.

[0063] The developmental mapping module is used to train a developmental mapping network with agronomic constraints on paired double-period fields. After training, the parameters are frozen and the developmental deviation vector is calculated.

[0064] The hierarchical path fusion module is used to automatically route each field to the corresponding path based on data availability indicators and developmental mapping reliability, and to unify features to the same dimension through a path-specific learnable missing compensation weight matrix.

[0065] The joint regression module is used to train a yield regression model with shared parameters using a mixed sampling mini-batch with unified dimensional features as input and measured yield as label, and to output the yield prediction value.

[0066] The beneficial effects of the present invention are as follows, compared with the prior art:

[0067] The first stage of this invention utilizes dual-period images without yield labels to train developmental mapping, reducing reliance on expensive yield labels. The second stage jointly trains yield regression on a small number of yield labels, with separate losses and clear logic between the two stages. Choosing the clearly identifiable tillering stage from transplanting to canopy closure rather than the peak tillering stage fundamentally ensures the accuracy of density feature extraction. This overcomes the technical bottleneck of existing sequence models failing directly when time-series data is interrupted. Path 3 serves as the full model, while paths 1, 2, and 4 serve as three sub-models. Combined with a unified feature compensation mechanism, it comprehensively covers all physical missing states that may be encountered in field collection, achieving extremely strong system robustness without increasing model size. Addressing the problem of expensive and scarce agricultural yield measurement labels, this invention achieves highly efficient cross-path knowledge transfer through a unique training paradigm of independent compensation for bottom-level paths and shared regression parameters for top-level paths. The ample training signal from Path 3 (complete two-stage process) significantly improves the prediction accuracy of each missing sub-model (Paths 1, 2, and 4) through the shared parameter Θ, avoiding the overfitting collapse problem that inevitably occurs with split training on small sample datasets of dozens of fields. The four paths share regression parameters, and the training signal from Path 3 indirectly improves the prediction quality of each sub-model through Θ; the compensation matrix for each path is optimized independently, flexibly learning feature compensation strategies under corresponding missing conditions. Inferring panicle characteristics from density through developmental mapping and retaining density as an independent feature is a degradation inference capability not found in existing methods. Compared to multiple samplings throughout the entire growth period, this invention only requires two key periods and allows for the absence of any one period, significantly reducing sampling costs while maintaining complete predictive capability. Attached Figure Description

[0068] Figure 1 This is a partial flowchart illustrating the principle of the hierarchical missing sense rice yield prediction method based on agronomic developmental chains in this invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0070] like Figure 1 As shown, this invention proposes a hierarchical missing sense method for rice yield prediction based on agronomic developmental chains, comprising the following steps:

[0071] The process consists of two training phases and one inference phase, executed sequentially. The two training phases are Phase 1 and Phase 2, respectively.

[0072] Phase 1 involves training a developmental mapping network on fields where both phase 1 and phase 2 images are available. This network maps the density feature vector from the tillering stage to the panicle feature vector at the heading and grain-filling stage. The training loss includes a feature reconstruction term and an agronomic upper bound constraint term. The agronomic upper bound constraint term penalizes cases where the number of effective panicles predicted by the mapping network exceeds the product of the number of plants per unit area and the upper bound of the maximum number of effective tillers per plant. For fields where both phase 1 and phase 2 data are available, the difference between the actual panicle features and the mapped prediction values ​​is calculated as the developmental bias vector. After training, the parameters of the developmental mapping network are frozen.

[0073] In a preferred but non-limiting embodiment of the present invention, density image acquisition is performed during the tillering period from transplanting to canopy closure rather than during the peak tillering period in Stage 1. At this time, the plants have not yet tillered extensively and the boundaries between plants are clear. The number of clumps can be accurately distinguished from the top view of the UAV, avoiding density extraction errors caused by overlapping stems and leaves during the tillering period.

[0074] In a preferred but non-limiting embodiment of the present invention, the process prior to stage one specifically includes:

[0075] Drone data collection was conducted only during two key growth stages: the tillering stage (clearly defined plant boundaries and accurate density extraction) and the heading and grain-filling stage (extracting images of rice panicles to determine the number of effective panicles and grains per panicle, and recording data availability indicators for each field). Any period can be missing. These are the data availability indicators for the tillering stage and the heading and grain-filling stage, respectively.

[0076] For fields that meet the conditions, the five-point sampling method is used to obtain the true yield value. (kg / mu), which means that five 1m² quadrats are selected in each field in a quincunx pattern, and the crops are harvested, threshed, dried and weighed. The average of the five quadrats is converted into the yield per unit area, which serves as the sole supervision label for the stage two regression model.

[0077] The dataset is classified as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] in, A set of fields for which bi-period images are available and without yield labels; A set of fields for which bi-period images are available and which have yield labels; This is a collection of fields where only heading and grain-filling period images are available and yield labels are provided; This is a collection of fields for which only images of the tillering stage are available and have yield labels. Images captured by drone during the tillering stage of the i-th field. Images of the i-th field taken by drone during the heading and fruiting stage. Let j be the true value of the measured yield of the j-th field. Images captured by drone during the tillering stage of the j-th field. Images captured by drone during the heading and fruiting stage of the j-th field. The images are taken by drone during the heading and fruiting stage of the k-th field. Let k be the true value of the measured yield of the kth field. Images captured by drone during the tillering stage of the lth field. This represents the true value of the measured yield of the l-th field.

[0083] Phase 1 use Phase Two Use .

[0084] In a preferred but non-limiting embodiment of the present invention, stage one specifically includes:

[0085] Features from the two periods were extracted using a self-supervised pre-trained encoder on unlabeled field images:

[0086] ;

[0087] ;

[0088] in, This represents the density feature vector during the tillering stage; A feature extractor or encoder for images of the tillering stage (the feature extractor or encoder for images of the tillering stage can be a CNN network); Images captured by drones during the tillering and greening stage, i.e., images of the tillering and greening stage. The feature extractor or encoder of the image during the tillering stage is fed into the image and outputs a d-dimensional vector. This represents the plant density or distribution information of this field; This represents the characteristic vector of rice panicles during the heading and grain-filling stage. For the feature extractor or encoder of the heading and fruiting stage image (the feature extractor or encoder of the heading and fruiting stage image can be a CNN network); Images captured by drones during the heading and grain-filling stage; that is, images captured by drones during the heading and grain-filling stage. Feature extractor or encoder fed into the heading and fruiting stage image Output a d-dimensional vector This represents information about the number, size, and distribution of rice ears.

[0089] In a preferred but non-limiting embodiment of the present invention, stage one further includes:

[0090] exist Training Developmental Mapping Network Do not use production labels:

[0091] ;

[0092] in The training loss (stage one loss) is used for the developmental mapping network M. This represents the characteristic vector of rice panicles during the heading and grain-filling stage. The predicted spikelet feature vector is output by the developmental mapping network M; This represents the density feature vector during the tillering stage; The weight of the agronomic constraint loss is set to control the proportion of the second term in the total loss (hyperparameter). This is the hinge function, which only penalizes cases where the limit is exceeded; if the limit is not exceeded, this term is 0. To predict rice panicle characteristics The estimated number of effective ears; This is the upper bound of the maximum number of effective tillers per plant (a constant determined by the variety, such as a maximum of 5 effective ears per plant), to prevent the mapping function from outputting biologically unreasonable predictions; The actual plant density (number of plants) was obtained from the images of the tillering stage.

[0093] Freeze after training convergence The parameters will no longer be updated.

[0094] In a preferred but non-limiting embodiment of the present invention, stage one further includes:

[0095] right Calculate the developmental deviation vector for each field in the dataset:

[0096] ;

[0097] This is the developmental deviation vector.

[0098] It should be noted that the developmental mapping network can be a developmental neural network.

[0099] Phase two involves using all fields with true yield values ​​as training data, i.e.

[0100] Based on the data availability flag and development mapping reliability of each field, each field is automatically routed to one of the four corresponding paths: The field in the middle is routed to path 3; The field in the middle is routed to path 2;

[0101] The fields are routed to either path 4 or path 1 based on mapping reliability. Features are unified to the same dimension using a path-specific learnable missing value compensation weight matrix, and then input into a yield regression model with shared parameters for the four paths. The weighted sum of the losses from the four paths constitutes the joint training objective. After each mini-batch mixed sampling, a backpropagation is performed to simultaneously update the shared regression parameters and the compensation matrix of the corresponding path. The true yield values ​​are obtained using a five-point sampling method, selecting five 1m samples from each field in a quincunx pattern. 2 The average value was calculated after the actual measurements were taken for each quadrat.

[0102] In a preferred but non-limiting embodiment of the present invention, stage two includes the construction of a four-path unified feature space, which is as follows:

[0103] All four paths unify the features to Dimension, input the same output regression model:

[0104] Path 3 (full model) is as follows:

[0105] ;

[0106] in The fused feature vector of path 3 is input into the output regression model. To be It is assembled sequentially; all three feature segments are real observations, and no compensation matrix is ​​required.

[0107] Path 2 (sub-model B) is as follows:

[0108] ;

[0109] in Input the fused feature vector of path 2 (available only during the heading and grain-filling stage) into the yield regression model; For the order These three vector segments are concatenated in this way: the 0 on the left is a d-dimensional zero vector, representing the missing density characteristics of the tillering stage; the 0 on the right is a d-dimensional zero vector, representing the missing developmental deviation characteristics. The learnable compensation weight matrix for path 2 is 3d*3d; its function is to reweight the rice spike features to compensate for the impact of missing density and bias information.

[0110] By reweighting the characteristics of rice panicles, the optimal feature compensation strategy is learned when density information and developmental deviations are missing.

[0111] Path 4 (sub-model C, enabled when mapping is reliable) is as follows:

[0112] ;

[0113] in The fused feature vector of path 4 (available only during the tillering stage, enabled when mapping is reliable) is input into the yield regression model; 0 represents the d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 4, with dimensions 3d*3d, is used to correct the distribution difference between the predicted rice spike features and the true features.

[0114] Density characteristics Retaining this information reflects the agronomical meaning of planting density as an independent multiplicative factor in the yield formula; using the predictive output of the developmental mapping network. Substitute for real rice ear characteristics ; due to reality Unavailable, developmental deviation vector Unable to calculate, corresponding zero-valued vector; learnable compensation weight matrix. The statistical distribution differences between the predicted characteristics and the actual rice panicle characteristics are compensated for, as well as the information loss introduced by setting the developmental deviation end to zero.

[0115] Path 1 (sub-model A, enabled when mapping is unreliable) is as follows:

[0116] ;

[0117] in The fused feature vector of path 1 is input into the output regression model; 0 is the d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 1 is 3d*3d; it is used to reweight the density features to compensate for the lack of rice spike and deviation information.

[0118] The density features are reweighted as a conservative alternative when path 4 is unavailable.

[0119] Both Path 1 and Path 4 have the same data availability condition (only images from the tillering stage are available). The difference lies in the reliability judgment of the developmental mapping network: when the number of paired fields used for Phase 1 training is not less than the preset reliability threshold, the developmental mapping network has sufficient generalization ability, and Path 4 is enabled to map and predict rice panicle characteristics. The missing true rice panicle features are replaced; when the number of paired double-cropping fields is below this threshold, the mapping network's generalization ability is insufficient, and path 4 degenerates into path 1, relying only on true density features. A prediction is made as a conservative alternative. Path 1 and Path 4 are not activated simultaneously during the same training process, and the same batch of data is routed to only one path, contributing to the loss.

[0120] In a preferred but non-limiting embodiment of the present invention, stage two further includes four-path weighted joint training, which is as follows:

[0121] The four paths share parameters in the output regression model. Each of the three sub-models has its own independent compensation matrix. The total loss of joint training is a weighted sum of the losses of the four paths:

[0122] ;

[0123] in for of ; For setting of Set of for path of The key is to check whether the mapping is reliable. for ; for ; for path ; for of .

[0124] In a preferred but non-limiting embodiment of the present invention, the joint training in stage two is specifically implemented as follows:

[0125] Each training mini-batch from , , Three datasets are sampled proportionally, and each sample is automatically routed to a corresponding path based on its data availability flag, and the feature vector of that path is calculated. Post-input shared model , to obtain the predicted value A single backpropagation operation updates the following parameters simultaneously:

[0126] Shared regression parameters (Samples from all paths participate in gradient updates); Compensation matrix for path 2 (only (Sample contribution gradient in the data); compensation matrix for path 4 (only (Enable the sample contribution gradient of path 4); compensation matrix of path 1. (only (Samples from path 1 contribute gradients); samples from path 3 do not contain a compensation matrix, and backpropagation only updates the gradient. Parameter sharing allows the training signal from path 3, when there is sufficient data, to be transmitted. This indirectly improves the prediction quality of each sub-model, resulting in cross-path knowledge transfer effects.

[0127] Regression model Random forests or gradient boosting trees are used to accommodate the small sample constraints that typically involve fewer than 50 yield-labeled plots.

[0128] In a preferred but non-limiting embodiment of the present invention, the yield regression model in stage two is a random forest or gradient boosting tree to adapt to the small sample constraint that the number of yield-labeled plots is usually less than 50; path 1 and path 4 use the same training data. When the mapping reliability switching mechanism is not activated simultaneously, the same batch of data is only routed to one of the paths, contributing to the loss.

[0129] The inference phase includes selecting the corresponding path for the new field based on data availability indicators and outputting the predicted yield value.

[0130] In addressing the issue of missing key-period images in multi-time series prediction, existing methods typically employ simple numerical interpolation or directly discard missing samples. This invention, considering the reality of incomplete field data collection, constructs a multi-path dynamic routing mechanism. This mechanism not only distinguishes between specific missing periods (corresponding to paths 2 and 4) but also introduces the reliability of the inference results from the preceding mapping network as a judgment criterion, providing a fallback alternative (corresponding to path 1). This hierarchical missing-aware architecture enables the model to adaptively switch inference paths under different data missing states, improving the system's fault tolerance in practical deployment. Facing the problem of inconsistent input feature dimensions caused by different missing conditions, the conventional approach is to construct multiple independent network branches. This invention, through a structured feature zero-padding and reweighting strategy, utilizes zero vectors and path-specific learnable weight matrices (…). ), mapping all inputs of different dimensions to a unified Dimensional space. For example, the structured fusion formula designed in path 4. This not only achieved dimensional alignment, but also made the matrix... This process learns and compensates for statistical distribution differences between predicted and true features. Obtaining agricultural field-level yield labels (such as measured data from five-point sampling) is costly, resulting in typically small sample sizes. Splitting samples into four independent models for training based on data missing states can easily lead to overfitting. This invention employs a weighted joint training mechanism, allowing samples with different missing states to be trained together in a mini-batch. Each sample independently updates its corresponding path's underlying compensation matrix, but collectively backpropagates to update the shared regression parameters at the top level. This mechanism enables cross-path feature representation transfer, allowing the yield mapping patterns learned by fields with complete data from both periods to assist feature learning in fields with missing data from a single period, thus mitigating the overfitting problem under small sample conditions.

[0131] This invention also proposes a hierarchical missing sense rice yield prediction system based on agronomic developmental chains, comprising:

[0132] The data acquisition module is used to collect drone images and record data availability indicators during the tillering and heading / filling stages after transplanting and before canopy closure, as well as to collect true yield values ​​using the five-point sampling method.

[0133] The feature extraction module is used to extract density feature vectors and rice spike feature vectors from the two images, respectively.

[0134] The developmental mapping module is used to train a developmental mapping network with agronomic constraints on paired double-period fields. After training, the parameters are frozen and the developmental deviation vector is calculated.

[0135] The hierarchical path fusion module is used to automatically route each field to the corresponding path based on data availability indicators and developmental mapping reliability, and to unify features to the same dimension through a path-specific learnable missing compensation weight matrix.

[0136] The joint regression module is used to train a yield regression model with shared parameters using a mixed sampling mini-batch with unified dimensional features as input and measured yield as label, and to output the yield prediction value.

[0137] This invention selects the tillering stage after transplanting and before canopy closure, rather than the peak tillering stage, for density feature collection: During the tillering stage, the plants have just established roots and stabilized, and have not yet tillered extensively to the point of canopy closure. The gaps between plants are clear, and the boundaries of individual plants are well-defined in the UAV top-down view. One response point corresponds to one cluster, making density extraction accurate and reliable. During the tillering stage, plants tiller rapidly, with multiple stems overlapping, making it difficult to accurately distinguish the number of clusters from the UAV perspective, resulting in large density extraction errors.

[0138] Using the full information path as the full model, three sub-models are set up to form a hierarchical structure of the full model and the three sub-models. The data conditions, usage information, and training data sources for the four paths are shown in Table 1 below:

[0139] Table 1

[0140]

[0141] Path 3 features do not require a compensation matrix, as all three feature segments are true observations; paths 1, 2, and 4 each have independent compensation matrices. , , W_{pe}, in stage two and regression parameters Synchronous end-to-end optimization. Path 1 and Path 4 have the same data collection conditions (only during the tillering stage), the difference being that Path 4 additionally utilizes a frozen developmental mapping network. Inferring predicted rice panicle characteristics. This is based on the number of paired double-phase fields used for Phase 1 training. If the reliability falls below a preset threshold, path 4 degenerates into path 1. Path 1 and path 4 use the same training data. However, under the mapping reliability switching mechanism, they are not activated simultaneously, and the same batch of data only contributes to the loss by taking one of the paths. Phase 1 uses... (All bi-period data are available for field plots; yield labels are not required) Training The training dataset for stage two product regression is independent of, but overlaps with, the training dataset for stage two product regression. (Also, labeled samples are provided for path 3).

[0142] The symbols of this invention are defined as follows:

[0143] Drone images of the tillering stage; Drone images of the heading and fruiting stage; Tillering period

[0144] The eigenvector of density; This represents the feature vector of a rice panicle during the heading and grain-filling stage. For developmental mapping functions (used during training in phase one and freezing in phase two); The rice spike characteristics are mapped and predicted. This is the developmental deviation vector; This serves as a marker for the availability of data during the tillering and heading / filling stages. These are the dedicated learnable missing compensation weight matrices for paths 1, 2, and 4, respectively. To integrate feature vectors with a unified dimension for input regression models; For the output regression model, the parameters are... Shared by four paths; For the first The loss weight coefficient for each path; The true yield (kg / mu) is the actual measured yield using the five-point sampling method. To predict output for the model; The agronomic upper limit coefficient for the maximum effective tiller number per plant; This is the weighting coefficient for agronomic constraint losses.

[0145] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0146] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0147] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0148] In any case, the language can be either compiled or interpreted.

[0149] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0150] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0151] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0152] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0153] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0154] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A hierarchical missing perception based on agronomic development chain for rice yield prediction method, characterized in that, include: The process consists of two training phases and one inference phase, executed sequentially. The two training phases are Phase 1 and Phase 2, respectively. Phase 1 involves training a developmental mapping network on fields where images from both periods are available, from the density feature vector at the tillering stage to the panicle feature vector at the heading and grain-filling stage. The training loss includes a feature reconstruction term and an agronomic upper bound constraint term, where the agronomic upper bound constraint term penalizes cases where the number of effective panicles predicted by the mapping network exceeds the product of the number of plants per unit area and the upper bound of the maximum number of effective tillers per plant. For fields with available bi-period data, the difference between the actual rice panicle characteristics and the mapped predicted values ​​is calculated as the developmental deviation vector; after training, the parameters of the developmental mapping network are frozen. Stage two, which includes all the training data with yield true value, according to the data availability flag and development mapping reliability of each field, the automatic routing of each field to one of the four paths, through the path-specific learnable missing compensation weight matrix, the features are unified to the same dimension, and then input into the yield regression model with shared parameters of four paths; The loss weighted sum of the four paths constitutes the joint training target, and each mini-batch mixed sampling is updated simultaneously by one backward propagation to update the shared regression parameters and the compensation matrix of the corresponding path, wherein the yield true value is obtained by five-point sampling method, and 5 1m 2 Quotient after taking the average of the five quadrats respectively measured The inference phase includes selecting the corresponding path for the new field based on data availability indicators and outputting the predicted yield value.

2. The hierarchical missing perception rice yield prediction method based on agronomic development chain according to claim 1, characterized in that, In Phase 1, density images were collected during the tillering stage from transplanting to before canopy closure.

3. The hierarchical missing perception rice yield prediction method based on agronomic development chain according to claim 2, characterized in that, Prior to Phase One, the following were included: Only two key growth periods are executed by unmanned aerial vehicle collection tasks, that is, the density image of rice is collected in the regreening tillering period of the key growth period, and the density characteristics of the rice field are extracted; the rice ear image is collected in the heading and fruiting period of the key growth period, and the number of effective ears and the number of grains per ear are extracted, and the data availability flag of each field is recorded , allowing any period to be missing, respectively, the data availability flag of the regreening tillering period and the heading and fruiting period For the qualified field, five-point sampling method is used to obtain the yield true value As the only supervised label for the phase two regression model, The dataset is classified as follows: ； ； ； ； in, A set of fields for which bi-period images are available and without yield labels; A set of fields for which bi-period images are available and which have yield labels; This is a collection of fields where only heading and grain-filling period images are available and yield labels are provided; This is a collection of fields for which only images of the tillering stage are available and have yield labels. Images captured by drone during the tillering stage of the i-th field. Images of the i-th field taken by drone during the heading and fruiting stage. Let j be the true value of the measured yield of the j-th field. Images captured by drone during the tillering stage of the j-th field. Images captured by drone during the heading and fruiting stage of the j-th field. The images are taken by drone during the heading and fruiting stage of the k-th field. Let k be the true value of the measured yield of the kth field. Images captured by drone during the tillering stage of the lth field. This represents the true value of the measured yield of the l-th field.

4. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 3, characterized in that, Phase One specifically includes: Features from the two periods were extracted using a self-supervised pre-trained encoder on unlabeled field images: ； ； in, This represents the density feature vector during the tillering stage; A feature extractor or encoder for images of the tillering stage; Images captured by drone during the tillering stage; This represents the characteristic vector of rice panicles during the heading and grain-filling stage. A feature extractor or encoder for images during the heading and fruiting stages; Images captured by drone during the heading and fruiting stage.

5. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 4, characterized in that, Phase one also includes: exist Training Developmental Mapping Network Do not use production labels: ； in The training loss of the developmental mapping network M is... This represents the characteristic vector of rice panicles during the heading and grain-filling stage. The predicted spikelet feature vector is output by the developmental mapping network M; This represents the density feature vector during the tillering stage; The weight of the agronomic constraint loss is set to control the proportion of the second term in the total loss (hyperparameter). For the hinge function; To predict rice panicle characteristics The estimated number of effective ears; This serves as an upper bound for the maximum effective tiller number per plant, preventing the mapping function from outputting biologically unreasonable predictions. The actual plant density was obtained from statistical analysis of images during the tillering stage. Freeze after training convergence The parameters will no longer be updated.

6. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 5, characterized in that, Phase one also includes: right Calculate the developmental deviation vector for each field in the dataset: ； This is the developmental deviation vector.

7. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 6, characterized in that, Phase two includes the construction of a unified feature space along four paths, which is as follows: All four paths unify the features to Dimension, input the same output regression model: Path 3 is as follows: ； in This is the fusion feature vector for path 3. To be It is assembled sequentially; Path 2 is as follows: ； in This is the fusion feature vector for path 2; For the order These three vector segments are concatenated in this way; the 0 on the left is the d-dimensional zero vector; the 0 on the right is the d-dimensional zero vector. The learnable compensation weight matrix for path 2; Path 4 is as follows: ； in Filtrate feature vectors for path 4; 0 It is a d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 4; Path 1 is as follows: ； in The fusion feature vector for path 1; 0 It is a d-dimensional zero vector. For the order These three vector segments are then concatenated. The learnable compensation weight matrix for path 1.

8. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 7, characterized in that, Phase two also includes four-path weighted joint training, which is as follows: The four paths share parameters in the output regression model. Each of the three sub-models has its own independent compensation matrix. The total loss of joint training is a weighted sum of the losses of the four paths: ； in for of ; For setting of Set of for path of Heavy, To determine if the mapping is reliable for ; for ; for path ; for of .

9. The method for predicting rice yield based on hierarchical missing sensory pathways according to claim 8, characterized in that, The specific implementation method of the joint training in Phase Two is as follows: Each training mini-batch from , , Three datasets are sampled proportionally, and each sample is automatically routed to a corresponding path based on its data availability flag, and the feature vector of that path is calculated. Post-input shared model , to obtain the predicted value ; The yield regression model described in Phase 2 is a random forest or gradient boosting tree to accommodate the small sample constraint that the number of yield-labeled plots is usually less than 50; Path 1 and Path 4 use the same training data. When the mapping reliability switching mechanism is not activated simultaneously, the same batch of data is only routed to one of the paths, contributing to the loss.

10. A hierarchical missing sense rice yield prediction system based on agronomic developmental chains, characterized in that, include: The data acquisition module is used to acquire drone images during the tillering and heading stages, record data availability indicators, and acquire true yield values ​​using the five-point sampling method. The feature extraction module is used to extract density feature vectors and rice spike feature vectors from the two images, respectively. The developmental mapping module is used to train a developmental mapping network with agronomic constraints on paired double-period fields. After training, the parameters are frozen and the developmental deviation vector is calculated. The hierarchical path fusion module is used to automatically route each field to the corresponding path based on data availability indicators and developmental mapping reliability, and to unify features to the same dimension through a path-specific learnable missing compensation weight matrix. The joint regression module is used to train a yield regression model with shared parameters using a mixed sampling mini-batch with unified dimensional features as input and measured yield as label, and to output the yield prediction value.

Images