A rice lodging remote sensing evaluation method and system based on interactive stacked integration

Abstract

This invention relates to the field of agricultural remote sensing and information technology, and discloses a remote sensing assessment method and system for rice lodging based on interactive stacking integration. The key technical points are: acquiring field measured data and UAV multi-source images of the target rice area; extracting a multi-dimensional feature parameter set based on the UAV multi-source images; constructing a physiological structural resilience index with mechanistic constraints; constructing and training an interactive stacking integration model; and outputting rice lodging assessment results based on the trained model. Multi-dimensional features are extracted using UAV multi-source remote sensing technology. By constructing a physiological structural resilience index (PSRI) that considers both plant physiological function damage and canopy structure destruction, and by introducing the interactive stacking integration OEIS framework to explicitly mine the extrapolation prediction results of the base model and the high-order nonlinear correlation between the original highly correlated features, the limitations of traditional integration models in small-sample scenarios are effectively overcome, achieving high-precision quantitative assessment of rice lodging severity and accurate classification of recovery capacity.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and information technology, and more specifically, to a remote sensing assessment method and system for rice lodging based on interactive stacking integration. Background Technology

[0002] Rice is one of the world's most important staple crops, but extreme winds and precipitation events caused by climate change have increased the risk of lodging, which can lead to yield loss. While UAV remote sensing technology has been widely used for lodging monitoring, current assessments largely rely on single geometric or physiological indicators such as lodging angle, lacking comprehensive methods to simultaneously quantify canopy structural damage and physiological function decline. Furthermore, it is difficult to effectively distinguish between lodging types with similar phenotypes but significantly different recovery potentials (such as stem lodging versus stem bending). In addition, deep ensemble architectures heavily rely on large-scale labeled datasets when constructing remote sensing inversion models, limiting their effectiveness in small sample sizes and highly complex scenarios.

[0003] Traditional stacking ensemble methods often fail to fully explore the interaction patterns between features. Under limited sample conditions, they are prone to overfitting due to data leakage. At the same time, their "black box" fusion mechanism also hinders the transparent explanation of the model's decision-making logic.

[0004] Therefore, the present invention provides a remote sensing assessment method and system for rice lodging based on interactive stacking integration, which improves the above-mentioned technical problems. Summary of the Invention

[0005] This disclosure aims to address the shortcomings of existing technologies by providing a remote sensing assessment method and system for rice lodging based on interactive stacked integration. The invention employs UAV multi-source remote sensing technology to extract multidimensional features. It constructs a Physiological Structure Resilience Index (PSRI) that considers both plant physiological function impairment and canopy structure damage, and introduces an interactive stacked integration (OEIS) framework to explicitly mine the high-order nonlinear correlation between the base model's out-of-fold prediction results and highly correlated original features. This effectively overcomes the limitations of traditional ensemble models in small-sample scenarios, achieving high-precision quantitative assessment of rice lodging severity, accurate grading of recovery capacity, and precise differentiation of complex lodging types (such as stem lodging and stem breakage) that exhibit similar phenotypes but different mechanisms.

[0006] To achieve the above objectives, the present disclosure proposes the following technical solutions:

[0007] In a first aspect, embodiments of this disclosure propose a remote sensing assessment method for rice lodging based on interactive stacking integration, the method comprising the following steps:

[0008] S1. Acquire field measured data and UAV multi-source images of the target rice area, wherein the field measured data includes physiological function parameters and canopy structure parameters, and the UAV multi-source images include visible light images and multispectral images.

[0009] S2. Based on the UAV multi-source imagery, extract a multi-dimensional feature parameter set including canopy coverage to characterize the state of the rice canopy.

[0010] S3. Based on the field measurement data, the physiological structure resilience index is constructed using the entropy weight method with mechanistic constraints. The mechanistic constraints are upper limit threshold constraints imposed on the final weight of the canopy coverage index, and the physiological structure resilience index is used as the label value for model training.

[0011] S4. Using the multidimensional feature parameter set as input and the physiological structure resilience index as label value, construct and train an interactive stacked ensemble model; wherein, during the training process of the interactive stacked ensemble model, generate the out-of-fold prediction values ​​of each base model, select the original features that are highly correlated with the out-of-fold prediction values, multiply the out-of-fold prediction values ​​with the selected original features to construct an interaction term, and concatenate the interaction term with the original features to form an enhanced feature matrix, which is then input into the meta-model for fitting;

[0012] S5. Acquire multi-source UAV images of the area to be tested and extract the multi-dimensional feature parameter set to be tested. Input the set into the trained interactive stacked ensemble model and output the rice lodging assessment results of the area to be tested.

[0013] The upper limit threshold value is 0.465; After determining the final weights of each indicator, the physiological function index (PFI) and the canopy structure index (CSI) were calculated separately, and then the functional structure resilience index was constructed. Constructing a physiological structural resilience index model: ; in, For a specific combination coefficient The physiological structural resilience index and combination coefficient were calculated below. To maximize Pearson correlation coefficient with measured ear biomass Y To achieve the objective, the optimal combination coefficients are determined through grid search. : ; in, The parameter value operator is used to maximize the objective function; the optimal combination coefficients are... Substituting these values ​​into the physiological structure resilience index model yields the label values ​​used for model training.

[0014] As a preferred technical solution of the present invention, in S2, the extracted multidimensional feature parameter set includes: vegetation index, texture features, canopy height and canopy coverage;

[0015] The formula for calculating the canopy height CH is as follows:

[0016] CH=DSM-DEM;

[0017] Wherein, DSM is a digital surface model generated based on the visible light image, and DEM is a digital elevation model;

[0018] Canopy coverage The acquisition process is as follows: A segmentation model of rice plants and soil background is constructed using the Excess Green Index (EXG), followed by binarization separation and calculation.

[0019] ;

[0020] ;

[0021] Wherein, G, R, and B are the gray values ​​of the green, red, and blue channels of the pixels in the visible light image, respectively; This represents the number of vegetation pixels in the binarized image. This represents the total number of pixels in the target rice paddy area.

[0022] As a preferred technical solution of the present invention, in S3, the process of constructing the physiological structure resilience index using the entropy weight method with mechanistic constraints includes:

[0023] The field measured data were divided into a physiological function subset including plant relative chlorophyll content, plant green leaf nitrogen concentration, and plant water content, and a canopy structure subset including leaf area index, canopy height, and canopy coverage.

[0024] Set the canopy coverage as a negative indicator and the other parameters as positive indicators, and perform range normalization.

[0025] The standardized formula for a positive indicator is:

[0026] ;

[0027] The standardized formula for negative indicators is:

[0028] ; Map the normalized data to a probability distribution. : ;

[0029] in, Let be the original observation value of the i-th sample and the j-th indicator. and These are the minimum and maximum values ​​of the indicator across all samples, respectively, where n is the number of samples. This is the standardized data.

[0030] As a preferred technical solution of the present invention, the information entropy and effective information content of each indicator are calculated, and an upper limit threshold constraint is introduced in the basic entropy weight calculation:

[0031] Information entropy of the j-th indicator and the amount of effective information The calculation formula is:

[0032] ;

[0033] ;

[0034] in, is the dimensionless coefficient of entropy; Let be the natural logarithm to the base e;

[0035] The basic entropy weight formula is:

[0036] ;

[0037] in, The effective information content of the j-th indicator; For the first The effective information content of each indicator; m is the number of indicators;

[0038] When calculating the weights of the canopy structure subset, if the weight of the canopy coverage calculated by the basic entropy weight formula exceeds the set upper limit threshold, the weight of the canopy coverage is forcibly fixed to the upper limit threshold, and the weights of the remaining indicators in the canopy structure subset are redistributed proportionally.

[0039] As a preferred embodiment of the present invention, the specific process of constructing and training the interactive stacking ensemble model in S4 is as follows:

[0040] Multiple machine learning algorithms were selected as base models for training.

[0041] Generate the out-of-bounds predictions of each base model under outer-layer cross-validation;

[0042] Calculate the Pearson correlation coefficient between the out-of-range predicted values ​​of each base model and the original features, and select the top three highly correlated features;

[0043] The predicted values ​​from the out-of-bounds region are multiplied by the selected features to construct an interaction term, forming an enhanced feature matrix;

[0044] The enhanced feature matrix is ​​input into the meta-model to complete the fitting.

[0045] As a preferred technical solution of the present invention, in S5, the rice lodging assessment results include lodging severity, recovery level and lodging type;

[0046] The method for outputting the lodging type is as follows: using the physiological structural resilience index of the non-lodging control area as the benchmark value, calculate the ratio of the predicted value of the test area to the benchmark value; use the ratio as the input feature to input the classifier in the trained interactive stacked ensemble model, and output the test pixel as stem lodging type 45°, stem lodging type 90°, or stem break type 90°.

[0047] Secondly, embodiments of this disclosure propose a remote sensing assessment system for rice lodging based on interactive stacking integration. The system is used to implement a remote sensing assessment method for rice lodging based on interactive stacking integration. The system includes:

[0048] The data acquisition and feature extraction module is used to acquire field measured data and UAV multi-source images of the target rice area, and extract a multi-dimensional feature parameter set including canopy coverage to characterize the state of the rice canopy.

[0049] The comprehensive index construction module is used to construct a physiological structure resilience index based on the field measured data using an entropy weight method that introduces mechanistic constraints. The mechanistic constraints are an upper limit threshold constraint imposed on the final weight of the canopy coverage index, and the physiological structure resilience index is used as the label value for model training.

[0050] The interactive stacked ensemble model construction module is used to construct and train an interactive stacked ensemble model by taking the multidimensional feature parameter set as input and the physiological structure resilience index as the label value. During the training process, the out-of-fold prediction values ​​of each base model are generated, the original features with high correlation to the out-of-fold prediction values ​​are selected and multiplied with the out-of-fold prediction values ​​to construct an interaction term, and the interaction term is concatenated with the original features to form an enhanced feature matrix, which is then input into the meta-model for fitting.

[0051] The lodging status assessment module is used to acquire multi-source UAV images of the area to be tested and extract the multi-dimensional feature parameter set to be tested. The data is then input into a trained interactive stacked ensemble model, which outputs the lodging assessment results of rice in the area to be tested.

[0052] Thirdly, embodiments of this disclosure propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a remote sensing assessment method for rice lodging based on interactive stacking integration.

[0053] Fourthly, embodiments of this disclosure provide an electronic device, including: a processor and a memory; the memory is used to store executable instructions of the processor; wherein the processor is configured to execute a rice lodging remote sensing assessment method based on interactive stacking integration by executing the executable instructions.

[0054] In summary, the present invention has the following beneficial effects:

[0055] Firstly, this invention constructs a Physiological Structure Resilience Index (PSRI), overcoming the limitations of existing technologies that rely on single geometric or physiological indicators. By introducing an entropy weight method with mechanistic constraints, this index can simultaneously consider and quantify the synchronous damage caused by lodging to the physiological functions of rice plants (such as chlorophyll and water content) and canopy structure (such as coverage and height). Verification shows that the PSRI has a very strong correlation with aboveground biomass, thus more accurately and objectively reflecting the actual biomass and yield loss caused by lodging.

[0056] Secondly, this invention proposes a novel interactive stacked ensemble framework (OEIS), which effectively overcomes the performance bottleneck of traditional ensemble models in complex scenarios with small sample sizes. Addressing the overfitting and data leakage problems easily caused by small sample sizes, the OEIS framework utilizes nested cross-validation and explicitly constructs a high-order interaction term of "base model out-of-range predicted value × highly correlated original feature," fully exploring the nonlinear correlations between multi-source features. In tasks such as landslide severity regression prediction and recovery level classification, OEIS not only significantly improves prediction accuracy but also substantially reduces model variance and enhances generalization robustness.

[0057] Thirdly, this invention achieves for the first time the accurate differentiation of complex rice lodging types that are "phenotypically similar but mechanistically different": existing technologies struggle to distinguish between stem lodging and stem bending at the same lodging angle, while this invention, leveraging the powerful decision boundary optimization capabilities of the OEIS framework, significantly reduces the confusion between these two types of severe lodging. Experimental results show that this invention significantly reduces the average misclassification rate of these two easily confused lodging types by 17.38% and improves the overall accuracy of lodging severity classification by 4.9%, providing a reliable technical basis for precise agricultural loss assessment and insurance claims. Attached Figure Description

[0058] Figure 1 A flowchart of a remote sensing assessment method for rice lodging based on interactive stacking integration provided in an embodiment of the present invention;

[0059] Figure 2 A linear relationship diagram between a single indicator and aboveground biomass provided for embodiments of the present invention;

[0060] Figure 3A linear relationship diagram between the comprehensive index PSRI and aboveground biomass provided for embodiments of the present invention;

[0061] Figure 4 A radar histogram comparing the determination coefficients R² of PSRI prediction models under different modeling strategies provided in this embodiment of the invention;

[0062] Figure 5 A radar histogram comparing the root mean square error (RMSE) of PSRI prediction models under different modeling strategies provided in this embodiment of the invention.

[0063] Figure 6 This is a radar histogram comparing the mean absolute error (MAE) of PSRI prediction models under different modeling strategies provided in this embodiment of the invention.

[0064] Figure 7 Scatter plot of PSRI predicted and measured values ​​for the single base model GBDT provided in this embodiment of the invention;

[0065] Figure 8 Scatter plot of PSRI predicted and measured values ​​for the Stacking_LassoCV stacking ensemble model provided in this embodiment of the invention;

[0066] Figure 9 Scatter plot of PSRI predicted and measured values ​​for the interactive stacked ensemble model OEIS_GBDT provided in this embodiment of the invention;

[0067] Figure 10 The heat map shows the performance evaluation index of lodging classification under different modeling strategies provided in the embodiments of the present invention.

[0068] Figure 11 This is a comparison chart of the average misclassification rates of different models provided in the embodiments of the present invention for phenotypically similar severe lodging types. Detailed Implementation

[0069] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0070] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0071] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0072] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0073] This disclosure aims to address the problems in existing rice lodging remote sensing assessments, such as the reliance on single evaluation indicators (making it difficult to consider both structural and physiological damage), the difficulty in accurately distinguishing phenotypic lodging types (e.g., stem lodging versus stem breakage), and the tendency of traditional ensemble learning models to overfit and lack interpretability in small-sample scenarios. Therefore, this disclosure proposes a rice lodging remote sensing assessment method and system based on interactive stacked ensemble (OEIS) to achieve timely and accurate monitoring of rice lodging severity and assessment of recovery potential. This method employs UAV multi-source remote sensing and machine learning techniques. By constructing a comprehensive resilience index (PSRI) that considers both physiological and structural damage, and introducing an interactive stacked ensemble (OEIS) framework, it explicitly mines the high-order nonlinear interactions between multi-source features and base model prediction results. This improves prediction accuracy, model interpretability, and reduces misclassification rates for complex lodging types in small-sample, complex scenarios.

[0074] Example 1: Please refer to Figure 1 , Figure 1 A flowchart of a remote sensing assessment method for rice lodging based on interactive stacking integration, as described in an embodiment of this disclosure, is shown. The overall process mainly includes the following five steps:

[0075] S1. Acquire field measurement data and UAV multi-source images of the target rice area.

[0076] In this embodiment, to comprehensively characterize the lodging state, an artificial lodging control experiment was conducted and ground-to-air data was acquired simultaneously:

[0077] S1.1 A design combining a non-lodging control and factor treatments was established, covering different lodging stages (12 days and 24 days after heading) and different degrees of lodging. Lodging degrees included: stem lodging at 45° (D1), stem lodging at 90° (D2, stems leaning towards the ground to avoid breakage), and stem bending at 90° (D3, the plant was manually bent at the third internode, causing it to lie at approximately 90° to the ground with its base bent). All treatments did not disturb the root system, simulating the natural lodging process.

[0078] S1.2. The relative chlorophyll content (SPAD) of the plants was determined using a soil and plant analyzer. The total fresh weight (FW) of the spike, stem sheath, and leaves, and the dry weight (DW) after drying to constant weight were measured. The plant moisture content (PWC) was calculated using the following formula: .

[0079] Leaf area index (LAI) was measured using a portable leaf area meter. Nitrogen concentration at different leaf positions was determined using the Dumas combustion nitrogen determination method, and the green leaf nitrogen concentration (LNC) of the plant was then calculated.

[0080] S1.3. Remote sensing images are acquired synchronously using a UAV equipped with an RTK module. The UAV is equipped with a visible light (RGB) camera and a multispectral camera covering five bands: blue (475±20nm), green (560±20nm), red (668±10nm), red edge (717±10nm), and near-infrared (842±40nm). Optimal flight parameters are: flight altitude 30m, forward overlap 85%, lateral overlap 85%, and flight speed 2m / s. The images are preprocessed to generate a digital surface model (DSM), orthophotos, and spectral reflectance for each band.

[0081] S2. Extract multi-dimensional feature parameter sets.

[0082] Based on UAV multi-source imagery, a total of 35 feature parameters were extracted, including:

[0083] S2.1 Extract core vegetation indices that are highly sensitive to greenness, moisture status, and canopy structure, including:

[0084] Normalized Difference Vegetation Index (NDVI), Normalized Difference Vegetation Index (GNDVI), Normalized Difference Vegetation Index with Red Edge (NDRE), Difference Chlorophyll Index with Red Edge (CIre), Difference Chlorophyll Index with Green Light (CIgreen), Terrestrial Chlorophyll Index (MTCI), Modified Chlorophyll Reflectance Index (MCARI), Transformed Chlorophyll Reflectance Index (TCARI), Soil-Optimized Vegetation Index (OSAVI), Two-Band Enhanced Vegetation Index (EVI2), Near-Infrared Vegetation Index (NIRv), Normalized Difference Water Index (NDVI) ), Red-edge normalized infrared index ( ), Red-edge water stress index ( The structural non-sensitive pigment index (SIPI), vegetation senescence reflectance index (PSRI), carotenoid reflectance index 2 (ARI2), visible light atmospheric impedance vegetation index (VARI), and normalized index 45 (NDI45) are also included.

[0085] S2.2. The Gray-Level Co-occurrence Matrix (GLCM) algorithm is used to extract texture features from the red, green, and blue channels of the RGB image of the rice canopy. The GLCM calculation parameters are set to distance 1 and angle 45°. Eight feature indicators are selected, namely, second moment of angle, entropy, contrast, correlation, inverse moment, variance, mean, and homogeneity, which respectively characterize the texture uniformity, complexity, and brightness difference, so as to achieve a quantitative description of the details of the canopy surface.

[0086] S2.3. Based on the digital surface model (DSM) and digital elevation model (DEM), the height data of the rice canopy is obtained by pixel-by-pixel difference calculation. The formula is: CH=DSM-DEM.

[0087] S2.4. An excess green index (EXG) model was constructed to separate rice plants from the soil background. The calculation formula is as follows: Where G is the green channel grayscale value of a pixel in the UAV RGB image; R is the red channel grayscale value; and B is the blue channel grayscale value. Binarization separation is achieved using the strong response characteristics of EXG to green vegetation, and canopy coverage is calculated according to the formula: ,in This represents the number of pixels representing vegetation. This represents the total number of pixels in the cell.

[0088] S3. Construct the physiological structural resilience index (PSRI) with mechanistic constraints.

[0089] To quantify the synchronous damage caused by lodging to plant physiological functions and canopy structure, an entropy weight method (EWM) was used, and mechanistic constraints were introduced to construct a comprehensive evaluation index. The specific steps are as follows:

[0090] S3.1, Directional Consistency and Standardization: Given the number of samples n and the number of indicators m, let the original matrix be... Since CC increases after lodging while other indicators decrease, CC is set as a negative indicator, and the remaining indicators—leaf area index (LAI), canopy height (CH), relative chlorophyll content (SPAD), green leaf nitrogen concentration (LNC), and plant water content (PWC)—are normalized to their ranges.

[0091] The standardized formula for positive indicators is: ;

[0092] The standardized formula for negative indicators is: ;

[0093] in, For the i-th sample and the j-th indicator, the original observation value is denoted as . It is the minimum value of the j-th indicator among all n samples; It is the maximum value of the j-th indicator among all n samples.

[0094] S3.2 Map each index column to a "probability distribution", the calculation formula is as follows:

[0095]

[0096] in, Let be the probability mapping value of the i-th sample under the j-th indicator, and n be the total number of samples.

[0097] S3.3, Define the dimensionless coefficient of entropy. The information entropy of the j-th indicator With "effective information content" Defined as:

[0098] ;

[0099] ;

[0100] in, Let be the natural logarithm to the base e.

[0101] S3.4, The basic entropy weight formula is: ;

[0102] in, The effective information content of the j-th indicator; For the first The effective information content of each indicator; m is the number of indicators.

[0103] To prevent excessive fluctuations in the CC index from causing it to acquire abnormally high weights, an upper limit is imposed on the entropy weight of CC.

[0104] The specific operation is as follows: with a step size of Δ=0.005, the upper limit of CC weight (cap) is densely searched in the interval [0,0.50], and the weights of the structural subsets are recalculated under each cap. The gradient decay method is used to find the inflection point of the "performance-stability" trade-off, and finally the upper limit of CC weight cap≈0.465 is determined. The weights of the remaining structural indices (LAI and CH) are redistributed proportionally.

[0105] S3.5. Calculate the Physiological Function Subset Score (PFI) and the Canopy Structure Subset Score (CSI) under the above weighting. Then construct the functional structural resilience index: ;

[0106] in, For a specific combination coefficient The functional structure toughness index is calculated below; For the combination coefficients, to maximize Pearson correlation coefficient with measured ear biomass Y To achieve the objective, the optimal coefficients are determined through a fine-grid search (step size 0.01). :

[0107] ;

[0108] in, The parameter value operator is used to maximize the objective function. The Pearson correlation coefficient measures the strength of the linear correlation between the two factors.

[0109] The optimal index model is ultimately determined using the above process and used as the label value for evaluation.

[0110] S4. Build and train the Interactive Stacked Integration (OEIS) model.

[0111] For complex scenarios with small sample sizes, a three-level modeling framework of "base model optimization - stacking ensemble - interactive stacking ensemble" is adopted to improve prediction accuracy:

[0112] S4.1. Eight classic machine learning algorithms were selected as base models, including: Ridge Regression, Lasso Regression, Elastic Network, KNN, SVR-RBF, Random Forest, GBDT, and XGBoost. Bayesian optimization was used to optimize the hyperparameters, with the root mean square error (RMSE) as the objective function.

[0113] S4.2. Generate the out-of-fold (OOF) prediction results of each base model under the outer 5-fold cross-validation. Calculate the Pearson correlation coefficient between the OOF of each base model and the original 35 features, and select the top-3 highly correlated features. Construct an interaction term of "OOF predicted value × original feature" to form an enhanced feature matrix.

[0114] S4.3 Input the enhanced feature matrix into the meta-model (GBDT is used for regression tasks) to complete the fitting. Perform SHAP (SHapley Additive ex Planations) analysis based on the trained model to quantify the contribution and direction of influence of each feature and interaction term on the PSRI prediction results.

[0115] S5. Output the rice lodging assessment results based on the trained model.

[0116] The model was validated using the following modified evaluation metrics, and the image of the region to be tested was input into the trained model to output the evaluation results:

[0117] S5.1, Using the coefficient of determination ( The root mean square error (RMSE), relative root mean square error (nRMSE), and mean absolute error (MAE) are used for evaluation.

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] in, The measured PSRI values ​​for the sample are as follows. These are the model's predicted values. is the average of the measured values, and n is the number of validation samples.

[0123] S5.3, Restoration Level and Lodging Severity Measurement: Output the continuous PSRI prediction values ​​of the measured pixels to characterize the comprehensive damage, and use the equidistant interval division strategy to map them into three levels: low recovery ability, medium recovery ability, and high recovery ability.

[0124] S5.4 Accurate Lodging Type Identification: Using the PSRI value (PSRI_CK) of the non-lodging control at the same time as a benchmark, the ratio (PSRI / PSRI_CK) is calculated as a quantitative input feature for lodging severity. The above-described system is applied to a classification task (the meta-model is replaced by a classifier), outputting the target pixels as D1 (stem lodging type 45°), D2 (stem lodging type 90°), or D3 (stem breakage type 90°). The OEIS framework significantly uncovers higher-order associations across models, reducing the average misclassification rate of D2 and D3 (phenotypically similar severe lodging) by 17.38%.

[0125] Example 2: A remote sensing assessment system for rice lodging based on interactive stacking integration, comprising:

[0126] Data acquisition and feature extraction module: used to acquire multispectral and visible light images and extract a set of multidimensional feature parameters including vegetation index, texture features, canopy height and canopy coverage.

[0127] The comprehensive index construction module includes a built-in entropy weight algorithm (EWM) with upper limit constraints and a supervised optimization engine to perform standardization, weight allocation, and finding the optimal coefficients. The calculation generates the Physiological Structure Resilience Index (PSRI) as a label.

[0128] The interactive stacked integrated model building module includes a built-in basic machine learning algorithm library. It performs nested cross-validation to generate out-of-flight (OOF) predictions, selects relevant features to construct higher-order interaction terms to form an enhanced feature matrix, and uses a meta-learner for secondary fitting to train and generate the OEIS model. The lodging status assessment module receives remote sensing feature streams from the target area and uses the trained OEIS model for inference. It outputs the quantitative PSRI prediction value, recovery level, and precise lodging type for the target plot.

[0129] Example 3: Model Validation and Technical Effect Analysis.

[0130] To verify the effectiveness of the rice lodging assessment method based on multi-source remote sensing and interactive stacked ensemble (OEIS) proposed in this invention, this embodiment verifies and analyzes the prediction and classification performance of the constructed physiological structural resilience index (PSRI) and the OEIS model:

[0131] For example, the validity verification of the PSRI index: In order to clarify the practical application value of the physiological structural resilience index (PSRI) constructed in this invention, this embodiment compares the linear correlation between PSRI and each individual index (including leaf area index LAI, canopy coverage CC, canopy height CH, leaf nitrogen content LNC, relative chlorophyll content SPAD, plant water content PWC) and aboveground biomass.

[0132] like Figure 2-3 As shown, the horizontal axis represents the values ​​of each individual indicator or the PSRI index constructed in this invention, and the vertical axis represents aboveground biomass. The yellow dots in the figure represent the actual observed data points of each rice experimental sample (i.e., the correspondence between the calculated value of the indicator and the measured biomass of each sample); the yellow solid line in the figure represents the linear fitting trend line between the horizontal axis variable and the vertical axis variable (aboveground biomass).

[0133] The specific meanings of the English letters and symbols marked in the image are as follows:

[0134] Pearson's coefficient: Represents the Pearson correlation coefficient, used to measure the strength and direction of the linear correlation between the horizontal axis indicators and the vertical axis aboveground biomass. The closer the value is to 1 or -1, the stronger the correlation.

[0135] R²: Represents the coefficient of determination (or goodness of fit), indicating the extent to which the linear model explains biomass variation. The closer the R² value is to 1, the stronger the model's ability to explain biomass changes and the better the fit.

[0136] p<0.05: This represents a statistically significant result, indicating that the linear correlation between this indicator and aboveground biomass is statistically significant.

[0137] The comparative data in the figure show that, except for canopy height (CH), all other individual indicators exhibit a significant linear correlation with aboveground biomass (p<0.05). However, the correlation between the comprehensive indicator PSRI constructed in this invention and aboveground biomass (Pearson's r=0.790) is significantly higher than that of all individual indicators, and a significant linear correlation exists between the two (p<0.05). The coefficient of determination R² of its linear fitting model reaches 0.623, meaning that the model can explain approximately 62.3% of the biomass variation. This confirms that the PSRI constructed in this invention not only integrates multidimensional physiological and structural characteristics but also has superior and more stable characterization power for biomass loss caused by lodging compared to single indicators.

[0138] For example, the performance of the interactive stacked ensemble (OEIS) model in regression prediction: In the PSRI regression prediction task (i.e., quantitative assessment of lodging severity), the interactive stacked ensemble (OEIS) framework proposed in this invention exhibits significant performance advantages. To more clearly illustrate the predictive effects of different modeling strategies, this embodiment provides a visual representation and comparison of the evaluation metrics and prediction scatter plots.

[0139] like Figure 4-6 As shown, a radar bar chart compares the average performance metrics (such as coefficient of determination R², root mean square error RMSE, and mean absolute error MAE) of the PSRI prediction model under different modeling strategies. In this radar chart, the circumferential direction (angle) represents the names of the various machine learning models and their variants (e.g., SVR, Ridge, RF, Lasso, KNN, GBDT, ENet, etc.) involved in the comparison; the radial direction (radius length) represents the specific numerical values ​​of the corresponding evaluation metrics (R², RMSE, MAE). Different colored sector bars are used in the chart to distinguish the modeling strategies: red sector bars represent the test results using a single base model (Base); blue sector bars represent the test results using a conventional stacking ensemble model (Stacking); and orange-yellow sector bars represent the test results using the interactive stacking ensemble model (OEIS) of this invention. The comparison of the radii of the bars in the figure clearly shows that the orange-yellow bar representing the OEIS strategy extends the furthest outward in terms of the R² index, while contracting the deepest in terms of the RMSE and MAE (error index), proving that the OEIS strategy achieves globally optimal prediction accuracy.

[0140] like Figure 7-9As shown, the scatter plots of PSRI predicted and measured values ​​for the best models among the three modeling strategies (i.e., the single base model GBDT, the stacked ensemble model Stacking_LassoCV, and the interactive stacked ensemble model OEIS_GBDT of this invention) are displayed. In each scatter plot, the horizontal axis (X-axis) represents the field measured value of PSRI, and the vertical axis (Y-axis) represents the PSRI predicted value output by the corresponding model. The gradient dots in the figure (transitioning from dark purple to bright yellow) represent the data points of each rice sample participating in the validation, and the gradient of color maps the feature numerical density or specific time-series / structural distribution characteristics of the sample. The thick black solid line in the figure is the linear regression fitting line obtained based on the fitting of all scatter samples; the corresponding light gray shaded area represents the 95% confidence interval of the linear fitting line. The thick black dashed line (1:1 line) in the figure is the absolute accuracy reference line under ideal conditions, that is, if the predicted value is exactly equal to the measured value, all dots should fall strictly on this dashed line. The smaller the angle between the solid black line and the dashed black line, and the higher the degree of overlap, the more accurate the model prediction. The evaluation metrics (R², RMSE, nRMSE, MAE) of the corresponding models are marked in the upper left corner of the figure.

[0141] Combination Figure 7-9 The data distribution shows that for the single-foundation model (GBDT), the scatter distribution is relatively scattered, with a coefficient of determination R² = 0.759 and a root mean square error (RMSE) of 0.110. For the conventional stacking ensemble model (Stacking_LassoCV), the scatter points converge towards the 1:1 dashed line, resulting in improved accuracy (R² = 0.759, RMSE = 0.110). However, for the interactive stacking ensemble model (OEIS_GBDT) used in this invention, the dots (especially samples in the mid-to-high value region) are most closely clustered around both sides of the 1:1 thick black dashed line, with the black solid line almost coinciding with the dashed line, resulting in the narrowest confidence interval. This model achieves the highest coefficient of determination (R² = 0.801) and the lowest error level (RMSE = 0.099, nRMSE = 21.48%, MAE = 0.077). This fully demonstrates that this invention, by constructing an explicit interaction term of "outside predicted value × original feature," successfully corrects the bias of conventional models in complex nonlinear mappings, achieving a dual breakthrough in accuracy and robustness.

[0142] For example, the performance of accurate identification of lodging recovery level and type: In the lodging recovery level classification task, the interaction items of the OEIS framework effectively improve the separability of classification boundaries. To intuitively verify the superiority of the present invention in classification tasks, this embodiment provides a visual comparison of the classification performance indicators and misclassification rates of different modeling strategies.

[0143] like Figure 10As shown, a heatmap of the performance evaluation metrics for the classification of landslides under different modeling strategies is presented: In this heatmap, the horizontal axis (X-axis) represents multiple core evaluation metrics for the classification task, including accuracy, precision, recall, F1 score, and area under the curve (AUC).

[0144] The vertical axis (Y-axis) represents the various machine learning models participating in the comparison, and is divided into three major sections from bottom to top according to different modeling strategies: the bottom section is a single base classification model (such as GBDT, KNN, Logistic, RF, SVC, XGB); the middle section is a conventional stacking ensemble model; and the top section is the interactive stacking ensemble model (OEIS) proposed in this invention.

[0145] The rectangular color blocks in the figure use gradient color levels (such as gradually transitioning from dark red / warm to dark blue / cool) to visually map the numerical value of the model on this indicator.

[0146] The right side of the diagram shows a vertical color bar that indicates the mapping between color intensity and specific numerical values. The closer the color of the patch is to the top of the color bar (e.g., dark blue), the higher the score and the better the performance of the model on that evaluation metric.

[0147] As can be seen from the color distribution of the heatmap, the top OEIS strategy model area (especially OEIS_GBDT and OEIS_Logistic) shows a dark color that represents high scores in all indicator columns, which is significantly better than the bottom single basic model area. This proves that the present invention effectively improves the separability of classification boundaries and the overall discrimination ability.

[0148] like Figure 11 The figure shows a comparison of the average misclassification rates of different models for phenotypically similar severe lodging types (i.e., stem lodging at 90° and stem bending at 90°):

[0149] In this line graph, the horizontal axis (X-axis) also represents the various machine learning models arranged in sequence and their respective modeling strategy ranges (Base, Stacking, OEIS).

[0150] The vertical axis (Y-axis) represents the average misclassification rate (misclassification rate) of the corresponding model when distinguishing between stem inverted 90° (D2) and stem folded 90° (D3).

[0151] The dark solid dots in the figure represent the actual average misclassification rates obtained by each model during testing and validation. The black solid line connecting these dots visually illustrates the dynamic changes and decreasing trend of the misclassification rate as the ensemble strategy evolves (from a single model to conventional stacking, and then to the interactive stacking of this invention).

[0152] As shown by the scatter plots and broken lines in the figure, stem lodging and stem breakage at the same lodging angle are highly similar in phenotype, representing a fundamental challenge in remote sensing identification. The misclassification rate of a single base model is generally high. Traditional stacking ensemble models, compared to the optimal single base model, only reduce the average misclassification rate of the two types of lodging by 4.67%. However, the interactive stacking ensemble framework (OEIS) proposed in this invention, by explicitly introducing higher-order interaction terms, lowers the corresponding solid dots to the lowest point in the entire image, significantly reducing the average misclassification rate by 17.38% and improving the overall severity classification accuracy by 4.9%. This fully demonstrates that this invention has completely overcome the bottleneck of classification technology where "phenotypes converge but mechanisms differ."

[0153] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A remote sensing assessment method for rice lodging based on interactive stacking integration, characterized in that, The method includes the following steps: S1. Acquire field measured data and UAV multi-source images of the target rice area, wherein the field measured data includes physiological function parameters and canopy structure parameters, and the UAV multi-source images include visible light images and multispectral images. S2. Based on the UAV multi-source imagery, extract a multi-dimensional feature parameter set including canopy coverage to characterize the state of the rice canopy. S3. Based on the field measured data, the physiological structure resilience index is constructed using the entropy weight method with mechanistic constraints. The mechanistic constraints are upper limit threshold constraints imposed on the final weight of the canopy coverage index, and the physiological structure resilience index is used as the label value for model training. S4. Using the multidimensional feature parameter set as input and the physiological structure resilience index as label value, construct and train an interactive stacked ensemble model; wherein, during the training process of the interactive stacked ensemble model, generate the out-of-fold prediction values ​​of each base model, select the original features that are highly correlated with the out-of-fold prediction values, multiply the out-of-fold prediction values ​​with the selected original features to construct an interaction term, and concatenate the interaction term with the original features to form an enhanced feature matrix, which is then input into the meta-model for fitting; S5. Acquire multi-source UAV images of the area to be tested and extract the multi-dimensional feature parameter set to be tested. Input the set into the trained interactive stacked ensemble model and output the rice lodging assessment results of the area to be tested. The upper limit threshold value is 0.465; After determining the final weights of each indicator, the physiological function index (PFI) and the canopy structure index (CSI) were calculated separately, and then the functional structure resilience index was constructed. Constructing a physiological structural resilience index model: ； in, For a specific combination coefficient The physiological structural resilience index and combination coefficient calculated below To maximize Pearson correlation coefficient with measured ear biomass Y To achieve the objective, the optimal combination coefficients are determined through grid search. : ； in, The parameter value operator is used to maximize the objective function; the optimal combination coefficients are... Substituting these values ​​into the physiological structure resilience index model yields the label values ​​used for model training.

2. The method for remote sensing assessment of rice lodging based on interactive stacking integration according to claim 1, characterized in that, In S2, the extracted multidimensional feature parameter set includes: vegetation index, texture features, canopy height, and canopy coverage; The formula for calculating the canopy height CH is as follows: CH=DSM-DEM; Wherein, DSM is a digital surface model generated based on the visible light image, and DEM is a digital elevation model; Canopy coverage The acquisition process is as follows: A segmentation model of rice plants and soil background is constructed using the Excess Green Index (EXG), followed by binarization separation and calculation. ； ； Wherein, G, R, and B are the gray values ​​of the green, red, and blue channels of the pixels in the visible light image, respectively; This represents the number of vegetation pixels in the binarized image. This represents the total number of pixels in the target rice paddy area.

3. The method for remote sensing assessment of rice lodging based on interactive stacking integration according to claim 1, characterized in that, In S3, the process of constructing the physiological structural resilience index using the entropy weight method with mechanistic constraints includes: The field measured data were divided into a physiological function subset including plant relative chlorophyll content, plant green leaf nitrogen concentration, and plant water content, and a canopy structure subset including leaf area index, canopy height, and canopy coverage. Set the canopy coverage as a negative indicator and the other parameters as positive indicators, and perform range normalization. The standardized formula for a positive indicator is: ； The standardized formula for negative indicators is: ； Map the normalized data to a probability distribution. : ； in, Let be the original observation value of the i-th sample and the j-th indicator. and These are the minimum and maximum values ​​of the indicator across all samples, respectively, where n is the number of samples. This is the standardized data.

4. The method for remote sensing assessment of rice lodging based on interactive stacking integration according to claim 3, characterized in that, Calculate the information entropy and effective information content of each indicator, and introduce an upper limit threshold constraint in the basic entropy weight calculation: Information entropy of the j-th indicator and effective information content The calculation formula is: ； ； in, is the dimensionless coefficient of entropy; Let be the natural logarithm to the base e; The basic entropy weight formula is: ； in, The effective information content of the j-th indicator; For the first The effective information content of each indicator; m is the number of indicators; When calculating the weights of the canopy structure subset, if the weight of the canopy coverage calculated by the basic entropy weight formula exceeds the set upper limit threshold, the weight of the canopy coverage is forcibly fixed to the upper limit threshold, and the weights of the remaining indicators in the canopy structure subset are redistributed proportionally.

5. The method for remote sensing assessment of rice lodging based on interactive stacking integration according to claim 1, characterized in that, In S4, the specific process of building and training the interactive stacked ensemble model is as follows: Multiple machine learning algorithms were selected as base models for training. Generate the out-of-bounds predictions of each base model under outer-layer cross-validation; Calculate the Pearson correlation coefficient between the out-of-range predicted values ​​of each base model and the original features, and select the top three highly correlated features; The predicted values ​​from the out-of-bounds region are multiplied by the selected features to construct an interaction term, forming an enhanced feature matrix; The enhanced feature matrix is ​​input into the meta-model to complete the fitting.

6. The method for remote sensing assessment of rice lodging based on interactive stacking integration according to claim 1, characterized in that, In S5, the rice lodging assessment results include lodging severity, recovery level, and lodging type; The method for outputting the lodging type is as follows: using the physiological structural resilience index of the non-lodging control area as the benchmark value, calculate the ratio of the predicted value of the test area to the benchmark value; use the ratio as the input feature to input the classifier in the trained interactive stacked ensemble model, and output the test area as stem lodging type 45°, stem lodging type 90°, or stem break type 90°.

7. A remote sensing assessment system for rice lodging based on interactive stacking integration, characterized in that, The system is used to implement the rice lodging remote sensing assessment method based on interactive stacking integration as described in any one of claims 1 to 6, the system comprising: The data acquisition and feature extraction module is used to acquire field measured data and UAV multi-source images of the target rice area, and extract a multi-dimensional feature parameter set including canopy coverage to characterize the state of the rice canopy. The comprehensive index construction module is used to construct a physiological structure resilience index based on the field measured data using an entropy weight method that introduces mechanistic constraints. The mechanistic constraints are an upper limit threshold constraint imposed on the final weight of the canopy coverage index, and the physiological structure resilience index is used as the label value for model training. The interactive stacked ensemble model construction module is used to construct and train an interactive stacked ensemble model by taking the multidimensional feature parameter set as input and the physiological structure resilience index as the label value. During the training process, the out-of-fold prediction values ​​of each base model are generated, the original features with high correlation to the out-of-fold prediction values ​​are selected and multiplied with the out-of-fold prediction values ​​to construct an interaction term, and the interaction term is concatenated with the original features to form an enhanced feature matrix, which is then input into the meta-model for fitting. The lodging status assessment module is used to acquire multi-source UAV images of the area to be tested and extract the multi-dimensional feature parameter set to be tested. The data is then input into a trained interactive stacked ensemble model, which outputs the lodging assessment results of rice in the area to be tested.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a remote sensing assessment method for rice lodging based on interactive stacking integration as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: A processor and a memory; the memory is used to store executable instructions of the processor; The processor is configured to execute, via executing the executable instructions, any one of claims 1 to 6, a rice lodging remote sensing assessment method based on interactive stacking integration.

Images