A corn leaf area index remote sensing inversion system and method, and application

Abstract

The application provides a corn leaf area index remote sensing inversion system and method and application, comprising: a multi-model operation layer configured to receive external input data, which is internally provided with: a kernel operation unit, which performs second processing on the external input data based on a kernel operation model to generate second intermediate data; and a feedforward network operation unit, which performs third processing on the external input data based on a feedforward neural network model to generate third intermediate data; the application adopts integral method to calculate LAD of each processing at different growth stages, and combines variance analysis to quantize main effects and stage influences of tillage mode, straw returning and nitrogen application management. Deep ploughing combined with straw returning and nitrogen application treatment performs best in crown layer construction in the early stage and maintenance in the middle and late stages, total LAD is at a leading level, and influences of each factor are relatively independent, so that the comprehensive effects of tillage measures on photosynthetic interception area extension and growth potential accumulation in the whole growth period can be comprehensively reflected.

Description

Technical Field

[0001] This invention belongs to the field of computational models, and more specifically, it relates to a remote sensing inversion system and method for maize leaf area index, and its application. Background Technology

[0002] Leaf area index (LAI) is a core physiological parameter characterizing crop canopy structure, photosynthetic capacity, and productivity. It directly determines the crop's interception efficiency of photosynthetically active radiation and effectively indicates the crop's growth and development stage, water and nutrient stress status, and final yield potential. In the context of global precision agriculture and food security strategies, real-time and accurate acquisition of large-scale farmland LAI dynamics has become a crucial technical support for optimizing cropping systems, implementing variable fertilization and irrigation decisions, and assessing the long-term effects of farming practices. However, traditional ground sampling methods have significant shortcomings: they require destructive plant harvesting, are time-consuming and labor-intensive, have poor spatial representativeness, and cannot meet the needs of high-frequency, large-area continuous monitoring throughout the entire growth cycle.

[0003] To overcome the limitations of ground sampling, UAV multispectral remote sensing has rapidly become a primary means of agricultural monitoring due to its advantages such as high spatiotemporal resolution, flexibility, and cost control. Current technologies typically utilize UAVs equipped with multispectral cameras to acquire images in the green, red, red-edge, and near-infrared bands, constructing vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and Soil-Adjusted Vegetation Index (SDI) for LAI (Local Area Index) retrieval. These indices, through band ratio calculations, suppress soil background interference to some extent and exhibit good sensitivity to low- to medium canopy cover.

[0004] However, as crops enter the mid-to-late growth stages, from jointing to grain filling, canopy closure rapidly increases, and the multiple scattering effect caused by overlapping leaves intensifies. This makes a single vegetation index prone to "spectral saturation," meaning the index value tends to level off or even decrease as LAI increases, failing to accurately reflect the true extent of canopy expansion. Furthermore, in actual farmland production, different tillage practices (deep plowing, deep loosening, no-till, shallow rotary tillage, etc.) and their combinations with straw return and nitrogen fertilizer management significantly alter topsoil compaction, micro-topographic undulation, and water and fertilizer redistribution patterns, leading to strong spatial heterogeneity in canopy structure and leaf area dynamics. This environmental heterogeneity driven by tillage practices results in a significant nonlinear and non-stationary shift in the mapping relationship between features and LAI across different sub-regions.

[0005] Existing inversion methods mostly rely on single or a few vegetation indices as input features, failing to effectively integrate texture features reflecting the heterogeneity of canopy spatial structure and topographic features regulating water and fertilizer distribution. At the same time, existing single machine learning or deep learning models have inherent limitations in handling high-dimensional heterogeneous features, making it difficult to compensate for feature distribution shifts and regional prediction biases caused by different farming practices through adaptive mechanisms. This results in low LAI inversion accuracy, poor robustness, and poor spatial continuity in complex farmland environments, making it unreliable to support dynamic monitoring of the growth period and accurate evaluation of farming effects. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a remote sensing inversion system and method for maize leaf area index, and the purpose and effectiveness of its application are achieved through the following specific technical means:

[0007] A remote sensing inversion system for maize leaf area index includes:

[0008] The multi-model computation layer is configured to receive external input data, and its internal settings include:

[0009] An integrated computing unit performs a first processing on the external input data based on an integrated computing model to generate first intermediate data;

[0010] The kernel processing unit performs a second processing on the external input data based on the kernel processing model to generate second intermediate data;

[0011] The feedforward network processing unit performs third processing on the external input data based on the feedforward neural network model to generate third intermediate data.

[0012] The convolutional network operation unit performs a fourth processing on the external input data based on the convolutional neural network model to generate fourth intermediate data;

[0013] The integrated computing layer has its input terminals connected to the output terminals of the multi-model computing layer. It has a built-in kernel combination model for performing composite operations on the received first intermediate data, second intermediate data, third intermediate data, and fourth intermediate data, and outputting the result data.

[0014] The integrated computing unit, the kernel operation unit, the feedforward network operation unit, and the convolutional network operation unit process the external input data from different perspectives, including ensemble learning, kernel methods, global mapping, and local pattern extraction. The comprehensive operation layer fuses the first to fourth intermediate data through an adaptive weighting mechanism to improve the accuracy of the result data.

[0015] As a further improvement, the ensemble computing unit is a random forest regressor, which generates the first intermediate data based on the ensemble learning method; the kernel operation unit is a kernel support vector regressor, which generates the second intermediate data based on the kernel method; the feedforward network operation unit is a multilayer perceptron network, which generates the third intermediate data based on the feedforward neural network architecture; and the convolutional network operation unit is a one-dimensional convolutional neural network, which generates the fourth intermediate data based on the convolutional neural network architecture.

[0016] The kernel combination model in the integrated computing layer is a radial basis support vector regression model, which is configured to use the first intermediate data, the second intermediate data, the third intermediate data, and the fourth intermediate data as input features, perform secondary training, and achieve integration through an adaptive weighting mechanism to output the result data;

[0017] The external input data is a set of preprocessed and feature-filtered remote sensing features, and also includes:

[0018] The multi-source feature input and preprocessing module is used to collect UAV multispectral image data during the key growth period of maize, perform radiometric calibration, reflectance correction, and generate vegetation masks based on soil-adjusted vegetation index to eliminate soil background interference.

[0019] The multi-source feature extraction and multi-criteria screening module has its input end connected to the output end of the multi-source feature input and preprocessing module. It is used to extract vegetation index features such as normalized vegetation index, soil-adjusted vegetation index, and optimized soil-adjusted vegetation index, texture features based on gray-level co-occurrence matrix, and digital terrain model features. It constructs key feature combinations that are highly correlated with leaf area index through weighted fusion screening using Pearson correlation analysis, mutual information method, and Lasso regression, and provides the key feature combinations as external input data to the multi-model operation layer.

[0020] The prediction and application layer receives the result data output by the comprehensive computing layer at its input end, and is used to perform pixel-level leaf area index regression prediction on the target farmland area, generate spatial distribution maps of each growth stage and calculate the duration of leaf area, so as to assess the impact of different tillage practices on maize canopy development and growth potential.

[0021] The multi-model computation layer and the integrated computation layer constitute the Stacking ensemble learning architecture. By combining heterogeneous base learners and meta-learners, the mapping relationship between the key features and leaf area index is learned from the perspectives of ensemble learning, kernel methods, global mapping and local pattern extraction, thereby achieving high-precision leaf area index inversion.

[0022] A remote sensing inversion method for maize leaf area index (LAI) utilizes the aforementioned maize LAI remote sensing inversion system. By fusing multi-source feature information and employing a multi-criteria feature selection strategy to construct a key feature set, and leveraging a Stacking ensemble learning architecture comprised of heterogeneous base learners and meta-learners, including random forests, support vector machines, multilayer perceptrons, and one-dimensional convolutional neural networks, the method achieves high-precision remote sensing inversion of farmland LAI. The method includes the following steps:

[0023] S1: Collect multi-source input feature data and corresponding target variable label data;

[0024] S2: Preprocess the input feature data to extract effective feature regions and eliminate background interference;

[0025] S3: Extract statistical features, texture features, and structural features from the preprocessed data to construct a multi-source feature candidate set;

[0026] S4: A multi-criteria comprehensive evaluation method is used to screen the multi-source feature candidate set to obtain the key feature combination for model training;

[0027] S5: Construct multiple heterogeneous base learners, including random forest, support vector machine, multilayer perceptron and one-dimensional convolutional neural network, and use stacking ensemble learning strategy to train meta-learner to adaptively fuse the prediction outputs of each base learner to form a maize leaf area index remote sensing inversion system.

[0028] S6: Utilize the trained Stacking ensemble learning architecture to perform sample-level target variable prediction for the target region, generate prediction results, and further calculate a comprehensive evaluation index to assess the response under different conditions.

[0029] As a further improvement, step S1 specifically includes: during key growth stages such as the jointing stage, tasseling stage, silking stage, and grain-filling stage of maize, randomly collecting maize plant samples using a fixed quadrat method and measuring leaf length and width parameters.

[0030] Based on the geometric parameters of the sampled leaves, planting density, and leaf area correction coefficient, the measured leaf area index value on the ground is determined using the leaf area index calculation formula and used as the training label data for the inversion model; simultaneously, UAV multispectral remote sensing image data corresponding to the ground sampling points are collected.

[0031] As a further improvement, step S2 specifically includes: using a multispectral UAV to acquire raw image data containing green light, red light, red edge and near-infrared bands;

[0032] Radiometric calibration, reflectance correction, and image stitching are performed on the raw image data to generate high-resolution orthophotos.

[0033] Based on the soil-adjusted vegetation index combined with the Otsu threshold segmentation algorithm, a vegetation mask is automatically generated and soil background pixel interference is removed.

[0034] As a further improvement, step S3 specifically includes: extracting multiple vegetation index features such as normalized vegetation index, soil-adjusted vegetation index, and optimized soil-adjusted vegetation index to characterize crop spectral response; performing texture analysis on red-edge bands and typical vegetation indices based on gray-level co-occurrence matrix, extracting texture indicators such as mean and information entropy, and taking multi-directional average values ​​to describe canopy spatial heterogeneity; and extracting digital terrain models as terrain features to reflect the regulatory role of micro-topography on field water and fertilizer distribution.

[0035] As a further improvement, step S4 specifically includes: S41: evaluating the linear correlation strength between features and leaf area index through Pearson correlation analysis, quantifying nonlinear dependencies through mutual information method, and achieving feature sparsity selection through Lasso regression; normalizing the feature scores of the three evaluation methods and then weighting and fusing them to construct a comprehensive feature importance evaluation index; and selecting core feature combinations covering the three dimensions of spectrum, texture, and terrain based on the comprehensive evaluation results for subsequent model training.

[0036] As a further improvement, step S5 specifically includes: selecting random forest, support vector machine, multilayer perceptron, and one-dimensional convolutional neural network as base learners to characterize the mapping relationship between features and leaf area index from different perspectives; dividing the sample data into training set and test set according to a preset ratio; using the prediction results of each base learner on the validation samples as new input features and feeding them into the radial basis function support vector regression meta-learner for secondary training to achieve adaptive weighted fusion of the prediction advantages of each base learner; and using the coefficient of determination and root mean square error as evaluation indicators to verify and optimize the generalization performance of the Stacking ensemble learning architecture.

[0037] As a further improvement, step S6 specifically includes: inputting the selected key features into the trained Stacking ensemble learning architecture to predict the leaf area index at the pixel level for the entire target area image; generating a spatial distribution map of the leaf area index for each growth stage of maize from the jointing stage to the grain-filling stage, quantitatively characterizing the spatiotemporal differences in canopy development under different tillage treatments; and using the integral method to calculate the duration of leaf area at different growth stages for each treatment, comprehensively evaluating the impact of tillage measures on the growth potential of maize.

[0038] As a further improvement, the Stacking ensemble learning strategy complements and fuses the residual characteristics and prediction bias of each base learner in a high-dimensional heterogeneous feature space, and then uses a meta-learner to achieve adaptive weighting.

[0039] The multi-criteria comprehensive evaluation method integrates the linear synergistic information captured by Pearson correlation analysis, the nonlinear dependency patterns extracted by mutual information method, and the sparse feature subsets achieved by Lasso regression in a weighted manner. The selected core feature combination simultaneously takes into account the effects of soil background suppression during the seedling stage, canopy saturation mitigation in the middle and late stages, and micro-topographic water and fertilizer regulation.

[0040] By combining the statistical analysis of pixel frequency distribution of LAI spatial distribution maps at different growth stages with the variance analysis of tillage factors, the combination of tillage measures with the strongest canopy expansion rate and anti-aging ability during the tasseling to grain filling stage was identified. By ensuring the continuous extension of the area intercepted by photosynthetically effective radiation throughout the entire growth period, the optimal accumulation of maize growth potential was achieved.

[0041] As a further improvement, an application of a remote sensing inversion method for maize leaf area index (LAI) is proposed. This method involves predicting LAI at the pixel level during the key growth stages from jointing to grain filling, generating spatial distribution maps of LAI at each growth stage, calculating the duration of LAI using an integral method, and combining pixel frequency distribution statistics with variance analysis of tillage factors to assess the impact of different tillage practices on maize canopy development and growth potential.

[0042] Beneficial effects

[0043] This invention integrates multi-source remote sensing information from vegetation index, texture features, and topographic features, employs a multi-criteria feature selection strategy to construct a key feature set, and utilizes a Stacking ensemble learning architecture built with heterogeneous base learners and meta-learners to achieve high-precision remote sensing inversion of maize leaf area index in complex farmland environments. Specific beneficial effects are as follows:

[0044] 1. It effectively overcomes the spectral saturation problem of a single vegetation index under high canopy closure, and improves the ability of the inversion model to capture the dynamics of LAI throughout the entire growth period.

[0045] By fusing spectral features such as the Normalized Difference Vegetation Index (NDVI) and Soil-Adjusted Vegetation Index (SDI) with texture features based on the gray-level co-occurrence matrix (GLCM) and a digital topography model, the model can simultaneously acquire information from three dimensions: spectral response, canopy spatial structural heterogeneity, and micro-topographic water and fertilizer regulation. In the mid-to-late growth stages, when canopy closure increases and multiple scattering effects intensify, texture and topographic features effectively compensate for spectral saturation, ensuring that the inversion results maintain good linear response and spatial continuity in the high LAI range. This significantly improves the monitoring accuracy of canopy development dynamics throughout the entire growth period.

[0046] 2. The core feature combination selected through a multi-criteria comprehensive evaluation method takes into account the explanatory power, robustness and simplicity of the model, and reduces the negative impact of feature redundancy on inversion accuracy.

[0047] This invention integrates the scores from three evaluation dimensions—Pearson correlation analysis, mutual information method, and Lasso regression—using normalized weighting to construct a comprehensive feature importance index. This screening mechanism simultaneously retains soil-modifying vegetation indices strongly linearly correlated with LAI, captures non-linearly dependent texture entropy features, and reflects environmental constraints in topographic features, avoiding the potential for missing key information or introducing redundant features by a single method. The screened feature set effectively reduces the input dimensionality while maintaining high interpretability, making subsequent model training more efficient and less prone to overfitting, laying the foundation for stable inversion in complex farmland environments.

[0048] 3. The Stacking ensemble learning architecture, constructed using heterogeneous base learners and meta-learners, achieves adaptive complementary fusion of the predictive advantages of each base learner, significantly improving the model's generalization performance and prediction accuracy under high-dimensional heterogeneous features.

[0049] Random forest base learners reduce variance through random feature selection, while support vector machine base learners enhance margin robustness through high-dimensional mapping. Multilayer perceptron and one-dimensional convolutional neural network base learners respectively mine high-order interactions between features from the perspectives of global nonlinear mapping and local pattern extraction. The meta-learner (radial basis support vector regression) uses the predicted outputs of each base learner as input features for secondary training, and can automatically learn the optimal weighted fusion strategy based on the position of different samples in the feature space. This residual complementarity and adaptive weighting mechanism effectively alleviates the systematic bias of a single model when dealing with feature distribution shifts caused by tillage measures, resulting in a test set determination coefficient of 0.786 and a root mean square error reduced to a minimum, far superior to a single base learner or a traditional linear model.

[0050] 4. It effectively adapts to field heterogeneity caused by different farming practices, and significantly improves the spatial continuity and regional adaptability of the inversion model in complex farmland contexts.

[0051] Different tillage methods, such as deep plowing, deep loosening, and no-till, and their combinations with straw return and nitrogen topdressing, lead to differences in soil compaction, micro-topography, and water and fertilizer distribution, resulting in strong spatial variability in canopy structure and LAI development. This invention, through multi-source feature fusion and a stacking adaptive fusion mechanism, enables the model to simultaneously capture spectral response variations, textural differences, and terrain-driven environmental constraints, maintaining stable predictive performance across different treatment areas. This avoids the problem of significant bias or discontinuous spatial inversion results in single-model implementations under specific tillage conditions.

[0052] 5. It achieves the combination of pixel-level high-resolution LAI spatial inversion and pixel frequency distribution statistics, which can accurately depict the spatiotemporal evolution of canopy development at different growth stages.

[0053] By using a trained Stacking model to perform pixel-level inference on full-region imagery, spatial distribution maps of the canopy index (LAI) at each growth stage from jointing to grain filling can be generated. Combined with the smooth rightward shift of the pixel frequency distribution curve, the dynamic transition process of LAI from low to high value ranges can be quantitatively revealed, as well as the evolutionary pattern of inter-treatment growth differentiation, which begins to appear at the tasseling stage and peaks at the grain filling stage. This dual verification system of "spatial pattern + statistical distribution" provides an intuitive and quantifiable basis for accurately evaluating the impact of cultivation practices on canopy expansion rate and anti-aging ability.

[0054] 6. By combining the integral calculation of leaf area duration (LAD) with multi-factor ANOVA, a multi-dimensional and precise analysis of the growth effects of different tillage measures was achieved, providing scientific decision support for the selection of the optimal tillage mode in cold and arid regions.

[0055] This invention uses an integral method to calculate the photosynthetic intercept area (LAD) of each treatment at different growth stages, and combines it with analysis of variance to quantify the main effects and stage-specific influences of tillage methods, straw return, and nitrogen topdressing. The results show that the deep plowing combined with straw return and nitrogen topdressing treatment performed best in early canopy formation and mid-to-late stage maintenance, with the total LAD leading the way, and the influence of each factor was relatively independent. This multidimensional analytical method avoids the one-sidedness of single-index evaluation and can comprehensively reflect the combined effects of tillage measures on the extension of photosynthetic intercept area and the accumulation of growth potential throughout the entire growth period. Attached Figure Description

[0056] Figure 1 This is an overview map of the research area for the remote sensing inversion system for maize leaf area index of this invention.

[0057] Figure 2 This is a technical roadmap for a remote sensing inversion system for maize leaf area index according to the present invention.

[0058] Figure 3 This is a ranking chart of the importance of comprehensive features.

[0059] Figure 4 This is a graph showing the results of GLM inversion.

[0060] Figure 5 A comparison chart of corn LAI inversion models based on RF, SVM, MLP and 1D-CNN.

[0061] Figure 6 This is a graph showing the inversion results of the ensemble learning model.

[0062] Figure 7 This is a spatiotemporal distribution map of maize LAI based on the Stacking model.

[0063] Figure 8 The pixel proportion distribution characteristics of LAI retrieved at different reproductive stages.

[0064] Figure 9 This is a graph showing the response of leaf area index at different growth stages to different tillage practices.

[0065] Figure 10 This is a flowchart illustrating the steps of a remote sensing inversion method for maize leaf area index according to the present invention.

[0066] Figure 11 This is a diagram showing the connection of the secondary modules of a remote sensing inversion system for maize leaf area index according to the present invention.

[0067] Figure 12 This is a diagram showing the connection of the first-level modules of a remote sensing inversion system for maize leaf area index according to the present invention. Detailed Implementation

[0068] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0069] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0070] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0071] Example

[0072] As attached Figure 1 To be continued Figure 12 As shown:

[0073] This invention provides a remote sensing inversion system for maize leaf area index, comprising:

[0074] The multi-model computation layer is configured to receive external input data, and its internal settings include:

[0075] An integrated computing unit performs a first processing on the external input data based on an integrated computing model to generate first intermediate data;

[0076] The kernel processing unit performs a second processing on the external input data based on the kernel processing model to generate second intermediate data;

[0077] The feedforward network processing unit performs third processing on the external input data based on the feedforward neural network model to generate third intermediate data.

[0078] The convolutional network operation unit performs a fourth processing on the external input data based on the convolutional neural network model to generate fourth intermediate data;

[0079] The integrated computing layer has its input terminals connected to the output terminals of the multi-model computing layer. It has a built-in kernel combination model for performing composite operations on the received first intermediate data, second intermediate data, third intermediate data, and fourth intermediate data, and outputting the result data.

[0080] The integrated computing unit, the kernel operation unit, the feedforward network operation unit, and the convolutional network operation unit process the external input data from different perspectives, including ensemble learning, kernel methods, global mapping, and local pattern extraction. The comprehensive operation layer fuses the first to fourth intermediate data through an adaptive weighting mechanism to improve the accuracy of the result data.

[0081] As a further improvement, the ensemble computing unit is a random forest regressor, which generates the first intermediate data based on the ensemble learning method; the kernel operation unit is a kernel support vector regressor, which generates the second intermediate data based on the kernel method; the feedforward network operation unit is a multilayer perceptron network, which generates the third intermediate data based on the feedforward neural network architecture; and the convolutional network operation unit is a one-dimensional convolutional neural network, which generates the fourth intermediate data based on the convolutional neural network architecture.

[0082] The kernel combination model in the integrated computing layer is a radial basis support vector regression model, which is configured to use the first intermediate data, the second intermediate data, the third intermediate data, and the fourth intermediate data as input features, perform secondary training, and achieve integration through an adaptive weighting mechanism to output the result data;

[0083] The external input data is a set of preprocessed and feature-filtered remote sensing features, and also includes:

[0084] The multi-source feature input and preprocessing module is used to collect UAV multispectral image data during the key growth period of maize, perform radiometric calibration, reflectance correction, and generate vegetation masks based on soil-adjusted vegetation index to eliminate soil background interference.

[0085] The multi-source feature extraction and multi-criteria screening module has its input end connected to the output end of the multi-source feature input and preprocessing module. It is used to extract vegetation index features such as normalized vegetation index, soil-adjusted vegetation index, and optimized soil-adjusted vegetation index, texture features based on gray-level co-occurrence matrix, and digital terrain model features. It constructs key feature combinations that are highly correlated with leaf area index through weighted fusion screening using Pearson correlation analysis, mutual information method, and Lasso regression, and provides the key feature combinations as external input data to the multi-model operation layer.

[0086] The prediction and application layer receives the result data output by the comprehensive computing layer at its input end, and is used to perform pixel-level leaf area index regression prediction on the target farmland area, generate spatial distribution maps of each growth stage and calculate the duration of leaf area, so as to assess the impact of different tillage practices on maize canopy development and growth potential.

[0087] The multi-model computation layer and the integrated computation layer constitute the Stacking ensemble learning architecture. By combining heterogeneous base learners and meta-learners, the mapping relationship between the key features and leaf area index is learned from the perspectives of ensemble learning, kernel methods, global mapping and local pattern extraction, thereby achieving high-precision leaf area index inversion.

[0088] To construct a computational model capable of integrating multi-source heterogeneous features and effectively overcoming the systemic bias of a single model through a complementary and adaptive fusion mechanism of heterogeneous base learners, thereby achieving high-precision, robust, and spatially continuous remote sensing inversion of maize leaf area index throughout the entire growth period, especially in the middle and late stages, the multi-model computation layer incorporates integrated computation units, kernel computation units, feedforward network computation units, and convolutional network computation units. These units employ heterogeneous base learners such as ensemble learning, kernel methods, feedforward neural networks, and convolutional neural networks to perform parallel first to fourth processing on external input data from different perspectives, including ensemble learning, kernel methods, global mapping, and local pattern extraction, generating multi-angle intermediate data. Finally, the integrated computation layer incorporates a kernel combination model (radial basis support vector regression) with an adaptive weighting mechanism to perform composite computation and fusion on the four intermediate data streams, outputting the final result data. This architecture effectively overcomes the inherent limitations of single models in characterizing extreme nonlinear boundaries and adapting to feature distribution shifts driven by tillage measures under high-dimensional heterogeneous remote sensing features by leveraging the complementary advantages of heterogeneous base learners and the adaptive residual hedging of meta-learners. This significantly improves the accuracy, robustness, and spatial continuity of leaf area index inversion, and solves the technical challenges of spectral saturation of single vegetation indices, insufficient utilization of multi-source features, and insufficient generalization performance of single models.

[0089] First, the original UAV multispectral imagery is standardized using a multi-source feature input and preprocessing module to obtain high-quality vegetation area data.

[0090] Subsequently, the multi-source feature extraction and multi-criteria screening module extracts multi-dimensional features from the preprocessed image and selects the most representative key feature combination through multi-criteria weighted fusion, effectively reducing feature redundancy and retaining information sensitive to LAI.

[0091] Next, the four base learners in the heterogeneous base learner layer learn the selected features in parallel, capturing the mapping relationship between the features and LAI from different perspectives;

[0092] Then, the meta-learner fusion layer uses the prediction results of each base learner as new input features for secondary training. Through an adaptive weighted fusion mechanism, it achieves the complementarity of the advantages of each base learner and error hedging, significantly improving the robustness and accuracy of the overall regression prediction. Finally, the prediction and application layer uses the trained architecture to perform pixel-level inference on the entire target area image, generating a LAI spatial distribution map. Furthermore, through leaf area duration calculation and tillage factor variance analysis, it achieves a quantitative assessment of the spatiotemporal differences in canopy development under different tillage measures.

[0093] During key growth stages of maize, such as jointing, tasseling, silking, and grain filling, ground-based measured LAI tag data and corresponding UAV multispectral images were collected simultaneously.

[0094] The above architecture is used to train and validate paired samples. After training, newly acquired images of target farmland areas are input into the architecture, which can quickly output pixel-level LAI prediction results and spatial distribution maps. Then, the leaf area duration is calculated and the effects of different farming measures are evaluated by combining variance analysis, providing a scientific basis for precision agriculture decision-making.

[0095] It effectively overcomes the spectral saturation problem of a single vegetation index under high canopy closure. By fusing multi-source features of spectrum, texture and topography, and with the help of the complementary learning ability of heterogeneous base learners, the model maintains good inversion accuracy and linear response throughout the entire growth period, especially in the middle and late stages.

[0096] It significantly improves the generalization performance of the model under high-dimensional heterogeneous features. The meta-learner effectively alleviates the system bias of the single model when the feature distribution shift is caused by different tillage treatments through the adaptive weighted fusion mechanism, and achieves higher regression prediction accuracy.

[0097] It better adapts to the spatial heterogeneity of complex farmland backgrounds, ensuring stable spatial continuity of inversion results across different cultivation zones. It achieves a combination of pixel-level high-resolution LAI spatial inversion and multidimensional evaluation, accurately characterizing the spatiotemporal evolution of canopy development at different growth stages. Furthermore, by combining leaf area duration with variance analysis, it quantitatively reveals the impact of different cultivation practices on maize growth potential, providing reliable technical support for the selection of optimal cultivation models in cold and arid regions.

[0098] This invention provides a remote sensing inversion method for maize leaf area index (LAI). Utilizing the aforementioned maize LAI remote sensing inversion system, it fuses multi-source feature information, employs a multi-criteria feature selection strategy to construct a key feature set, and leverages a Stacking ensemble learning architecture—comprising heterogeneous base learners and meta-learners including random forests, support vector machines, multilayer perceptrons, and one-dimensional convolutional neural networks—to achieve high-precision regression prediction of the target variable under complex environments. The method includes the following steps:

[0099] S1: Collect multi-source input feature data and corresponding target variable label data;

[0100] S2: Preprocess the input feature data to extract effective feature regions and eliminate background interference;

[0101] S3: Extract statistical features, texture features, and structural features from the preprocessed data to construct a multi-source feature candidate set;

[0102] S4: A multi-criteria comprehensive evaluation method is used to screen the multi-source feature candidate set to obtain the key feature combination for model training;

[0103] S5: Construct multiple heterogeneous base learners, including random forest, support vector machine, multilayer perceptron and one-dimensional convolutional neural network. Use the stacking ensemble learning strategy to train the meta-learner to adaptively fuse the prediction outputs of each base learner to form a target variable regression prediction model. The target variable regression prediction model is specifically the maize leaf area index inversion model.

[0104] S6: Use the trained prediction model to predict the target variable at the sample level in the target area, generate prediction results, and further calculate the comprehensive evaluation index to evaluate the response under different conditions.

[0105] The target area is the target farmland area. The sample-level target variable is predicted as the pixel-level spatial inversion of maize leaf area index, generating spatial distribution maps of LAI at different growth stages, and further calculating the duration of leaf area to assess the growth response under different farming practices.

[0106] In this embodiment, the soil was sandy loam, the temperature range was controlled at 6-8℃, the annual precipitation was controlled at approximately 400mm, the frost-free period at approximately 140 days, and the altitude at approximately 1015m. The experiment was conducted on a long-term soil fertility improvement experimental platform, employing a multi-factor randomized block design with 13 treatments covering three dimensions: tillage method (deep plowing DT, deep loosening SB, strip deep rotary tillage SDR, no-till NT, and farmer shallow rotary tillage CK), straw return to the field (return to the field / no return to the field), and nitrogen fertilizer management (topdressing / no topdressing). The specific treatment codes were: DT1, DT2, DT2-N, SB1, SB2, SB2-N, SDR1, SDR2, NT1, NT2, NT2-N, CK1, and CK1-N. The tested maize variety was 'Xianyu 696', planted with equal row spacing (0.6m) and a planting density of 8.25×10⁶. 4 Plants / hm². The total fertilizer application was consistent across all treatments (N 225 kg / hm², P₂O₅ 210 kg / hm², K₂O₂ 202.5 kg / hm²). Phosphorus and potassium fertilizers were applied as basal fertilizer, while nitrogen fertilizer (urea) was applied via drip irrigation at a ratio of 3:6:1 (jointing stage: large trumpet stage: grain-filling stage). Irrigation was conducted four times throughout the growth period (750 m³ / hm² each time).

[0107] In this embodiment, ground sampling and UAV imagery were carried out simultaneously during four key growth stages of maize: jointing, tasseling, silking, and grain filling, to provide paired datasets for subsequent model training and validation.

[0108] This embodiment sets up 13 treatments. In each treatment plot, three 1m×1m fixed quadrats are randomly deployed at each growth stage, totaling 156 quadrats across four key growth stages. Through precise registration using handheld GPS positioning and UAV orthophotos (DOM), approximately 25–35 effective pixels (after removing edge effects and anomalous pixels) are extracted from each quadrat within the vegetation mask, resulting in approximately 4,800 pixel-level multi-source feature LAI label paired samples. All samples are randomly divided into a training set (approximately 3,360 samples) and a test set (approximately 1,440 samples) at a 70%:30% ratio for base learner training, secondary training of the Stacking meta-learner, and final generalization performance evaluation. Although this sample size falls within the typical range of small-sample farmland datasets, the overfitting risk is effectively mitigated through multi-source feature fusion, multi-criteria selection, and heterogeneous base learner Stacking integration, achieving a high accuracy of 0.786 R² on the test set.

[0109] This embodiment conducts the experiment according to steps S1–S6. The specific process is as follows: First, ground truth and UAV imagery are collected, followed by preprocessing, feature engineering, model building and spatial inversion, and finally, the effects of tillage measures are evaluated through LAD and ANOVA.

[0110] Step S1: Collect ground-based measured leaf area index data and corresponding UAV multispectral image data during the key growth stages of maize;

[0111] S11: During the four key growth stages mentioned above, fixed quadrats (1m × 1m) were used as sampling units, and multiple quadrats were randomly distributed within each treatment plot. Three representative maize plants were randomly selected, and the length L of all unfolded leaves of each plant was measured with a measuring tape. i (cm) and width W i (cm). Avoid field edge effects during sampling to ensure the samples are representative.

[0112] S12: Calculate the measured ground-based LAI value based on the sampled data, which serves as the label data for the inversion model. Specifically, the length-width method is used, and the calculation formula is as follows:

[0113]

[0114] In the formula, LAI is the leaf area index (m²). 2 / m 2 )

[0115] m represents the number of plants sampled;

[0116] n is the total number of leaves of the sampled plant;

[0117] L i This represents the length (cm) of the i-th leaf.

[0118] W i This represents the width (cm) of the i-th leaf.

[0119] k is the leaf area correction coefficient, with an empirical value of 0.75 used to correct the deviation between the actual shape of corn leaves and the assumed rectangular shape.

[0120] D represents plant density (plants / m²) 2 )

[0121] S is the conversion factor, 10000 (cm²) 2 / m 2 The calculation results will be converted to (cm) 2 / m 2 LAI is expressed in units of ).

[0122] For each 1m×1m quadrat, all leaves of the 3 plants were sampled. i and W iThe leaf area of ​​each plant was measured and summed to obtain the total leaf area per plant; this sum was then multiplied by a correction factor k=0.75; multiplied by the planting density D and divided by 10000 to obtain the representative LAI value of that quadrat. The average LAI of multiple quadrats was taken as the measured value of the corresponding treatment plot. All calculations were automated using Python scripts to ensure data consistency and repeatability. This formula directly supports S12, providing high-precision, low-destructive ground truth labels for model training.

[0123] A detailed performance comparison of each model on the validation set is shown in Table 1 below.

[0124] S13: Simultaneously, a DJI Movik 3 multispectral drone was used to collect multispectral images corresponding to the ground sampling points during the weather-stable period (11:00-14:00). The flight parameters were: flight altitude 50m, overlap rate 80%, and raw image data in four bands: green (G), red (R), red edge (RE), and near-infrared (NIR). The images covered all 13 processing cells to ensure spatiotemporal registration.

[0125] Step S2: Preprocess the UAV multispectral image data to extract vegetation areas and remove soil background interference;

[0126] S21: Use the DJI Movic 3 multispectral drone to acquire raw images containing G, R, RE, and NIR bands according to parameters S13.

[0127] S22: Using DJITerra software, the original images are sequentially radiometrically calibrated, reflectance corrected, and image stitched to generate a high-resolution orthophoto map (DOM). This step eliminates sensor response differences, atmospheric effects, and geometric distortions, providing a radiometrically consistent and geometrically accurate image foundation for subsequent feature extraction.

[0128] S23: To eliminate soil background interference, an automated vegetation extraction workflow based on the Soil Adjusted Vegetation Index (SAVI) and the Otsu algorithm is constructed. This is implemented in a Python environment:

[0129] First, calculate the SAVI index (SAVI=(NIR-R) / (NIR+R+L)×(1+L, where L=0.5 is the soil conditioning parameter).

[0130] Then, the Otsu threshold segmentation algorithm is used to binarize the SAVI image and automatically determine the optimal threshold to generate a vegetation mask.

[0131] Finally, a mask was used to perform pixel-level cleaning on the orthophoto, retaining only vegetation pixels for feature extraction. This preprocessing effectively suppressed soil noise during the seedling stage and under sparse canopy conditions, supported S23, and provided high-purity data for subsequent multi-source feature extraction.

[0132] Step S3: Extract vegetation index features, texture features, and terrain features from the preprocessed image data to construct a multi-source feature candidate set;

[0133] Within the vegetation mask area generated by S2, three types of complementary features are extracted to overcome the limitations of single spectral features:

[0134] S31: Thirteen vegetation index features were extracted, including but not limited to Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Optimized Soil-Adjusted Vegetation Index (OSAVI), Improved Soil-Adjusted Vegetation Index (MSAVI2), Normalized Red Edge Index (NDRE), Normalized Water Index (NDWI), and Red Edge Vegetation Index (RERVI). These indices characterize crop biomass, chlorophyll content, water status, and soil background inhibition capacity from different perspectives, particularly alleviating the spectral saturation problem under high canopy closure.

[0135] S32: Texture features are extracted based on the Gray-Level Co-occurrence Matrix (GLCM). The red-edge band and typical indices such as NDVI and SAVI are linearly stretched to 32 gray levels. A 3×3 sliding window is used to calculate eight texture indices: mean (MEA), information entropy (ENT), variance, contrast, uniformity, correlation, and second moment of angle. The average values ​​at 0°, 45°, 90°, and 135° are calculated as the final texture feature output, ensuring rotation invariance. Texture features supplement the lack of spectral information from the perspective of spatial heterogeneity, effectively characterizing the complexity of canopy structure, especially in the later stages of canopy closure.

[0136] S33: Extract the digital terrain model (DTM) as terrain features. The DTM generated from UAV imagery reflects the regulatory effect of micro-topographic undulations on the redistribution of water and fertilizer in the field, introducing environmental background constraints into the inversion model.

[0137] By eliminating collinearity interference through variance inflation factor (VIF) and stepwise regression analysis, three core significant features were selected to construct a generalized linear model (GLM). Specific feature statistics are shown in Table 1. The mean VIF of the selected features was only 1.697, far below the conventional collinearity threshold (VIF < 10), and all were significant at the 0.01 level, indicating that the selected variables possess a high degree of statistical independence and representativeness.

[0138] Experimental results show that ( Figure 4The initial baseline model demonstrated basic predictive ability on the validation set, with a determination coefficient R² of 0.522, RMSE of 0.4807, and MAE of 0.4035. From the scatter distribution, while the GLM model could initially characterize the growth trend in the low-to-medium value range of LAI, its linear mapping mechanism exhibited significant estimation bias when dealing with the strong nonlinear fluctuations caused by high-value LAI and complex farmland backgrounds. The results indicate that conventional linear inversion methods are insufficient to fully resolve the deep interaction information between multi-source remote sensing features, thus confirming the necessity of subsequently introducing nonlinear deep learning algorithms and ensemble learning architectures to improve monitoring accuracy.

[0139]

[0140] Table 1. Results of the multicollinearity test for variables.

[0141] Step S4: Use a multi-criteria comprehensive evaluation method to screen the multi-source feature candidate set and obtain the key feature combination for model training;

[0142] To avoid the limitations of a single screening method, this embodiment adopts a multi-criteria comprehensive evaluation scheme:

[0143] S41: Calculate three types of evaluation indicators respectively: Pearson correlation coefficient (evaluates the linear synergy between the feature and the measured LAI), mutual information value (quantifies the nonlinear dependence between the feature and the LAI), and Lasso regression coefficient (achieves feature sparsity through L1 regularization and identifies substantial contributing factors).

[0144] S42: The raw scores of the three methods are respectively processed by Min-Max normalization, and then weighted and fused according to empirical weights (e.g., 0.4 linear + 0.3 nonlinear + 0.3 sparsity) to construct a comprehensive feature importance evaluation index. The weighted fusion mechanism is as follows:

[0145] A single Pearson algorithm may miss nonlinear relationships, mutual information captures nonlinearity but is insensitive to redundant features, and while Lasso can achieve sparsity, it may incorrectly remove weakly correlated texture features that are important in specific farming contexts. By using a multi-criteria synergy, robust feature ranking can be obtained in scenarios such as the seedling stage (high soil disturbance), the middle and late stages (high risk of saturation), and different farming treatments (micro-topographical differences).

[0146] In this embodiment, the weights are determined by grid search on the validation set with the goal of maximizing R², and in practice, 0.4 (Pearson) + 0.3 (mutual information) + 0.3 (Lasso) are used; the weighted fusion can be further optimized through cross-validation.

[0147] S43: Based on the comprehensive score, nine core features were selected, including top-ranked soil regulation indices such as SAVI, MSAVI2, and OSAVI, supplemented by NDVI, NDRE, NDWI, RERVI, DTM, and the NDVI information entropy texture feature NDVI_ENT_3m based on a 3×3 window. These features simultaneously cover three dimensions: spectral response, spatial structure, and environmental constraints, providing concise yet information-rich input for the subsequent Stacking model, effectively supporting stable generalization performance under 13 differentiated tillage treatment scenarios.

[0148] The Stacking ensemble learning architecture of this invention is implemented through a Python pipeline (scikit-learn, TensorFlow / PyTorch, rasterio / joblib). Specifically, the multi-source feature input and preprocessing module corresponds to a preprocessing function, the multi-source feature extraction and multi-criteria selection module corresponds to a feature engineering function, the heterogeneous base learner layer consists of four independently trained model objects, the meta-learner fusion layer is an RBF-SVR meta-model, and the prediction and application layer is a post-processing script combined with inference. The modules are chained together via a pipeline or a custom class, and input and output are passed via pandas.DataFrame or numpy.ndarray.

[0149] Step S5: Construct multiple heterogeneous base learners, including random forest, support vector machine, multilayer perceptron and one-dimensional convolutional neural network, and use the stacking ensemble learning strategy to train the meta-learner to adaptively fuse the prediction outputs of each base learner to form a maize leaf area index inversion model.

[0150] Based on the nine core features selected by S4, a Stacking ensemble model was constructed and trained. By complementaring and fusing heterogeneous base learners and adaptively weighting meta-learners, the limitations of a single model in handling complex, high-dimensional, heterogeneous farmland features were overcome.

[0151] S51: Four heterogeneous base learners are selected to characterize the mapping relationship between features and LAI from different mechanisms:

[0152] Random Forest (RF): Employs a Bagging strategy and random feature selection to reduce model variance, is robust to feature importance ranking, and is suitable for handling redundancy and noise in multi-source features;

[0153] Support Vector Machine (SVM): It maps features to a high-dimensional space through kernel functions, maximizes the classification / regression margin, and has good generalization ability in high-dimensional scenarios with small samples;

[0154] Multilayer Perceptron (MLP): Composed of stacked fully connected layers, it can fit complex global nonlinear relationships and capture high-order interactions between features;

[0155] One-dimensional convolutional neural networks (1D-CNNs) apply one-dimensional convolution and pooling to feature vectors to automatically extract local patterns and translation-invariant features, supplementing the shortcomings of the global model.

[0156] The four base learners are implemented using Python's scikit-learn (RF, SVM, SVR) and TensorFlow / PyTorch (MLP, 1D-CNN), and the hyperparameters are determined on the training set through grid search or Bayesian optimization.

[0157] As a preferred embodiment, the hyperparameters of the random forest are set as follows: n_estimators=200, max_depth=12, min_samples_split=4, max_features='sqrt', bootstrap=True; the support vector machine uses the RBF kernel function, C=8.0, epsilon=0.05, gamma='scale'; the multilayer perceptron (MLP) uses three fully connected hidden layers with the number of neurons being 128, 64, and 32 respectively, the activation function being ReLU, and the dropout rate being 0.25. The first layer is a single neuron (linear activation) using the Adam optimizer (initial learning rate 0.001, with ReduceLROnPlateau decay strategy), batch size 32, maximum training epochs 200, and early stopping tolerance value 20. The second layer is a one-dimensional convolutional neural network (1D-CNN) with two one-dimensional convolutional layers (both with a kernel size of 3, padding='same', output channels 32 for the first layer and 64 for the second). Each convolutional layer is followed by ReLU activation and dropout of 0.2, followed by a global average pooling layer and two fully connected layers, specifically 64-32-1, with the output being the LAI prediction value. All model hyperparameters are determined by 5-fold cross-validation combined with Bayesian optimization (or grid search) on the training set, with the minimum RMSE on the validation set as the optimization objective.

[0158] S52: Randomly divide the paired feature-LAI dataset into training and test sets in a 70%:30% ratio. The training set is used to train the base learners and meta-learners, and the test set is used for final performance evaluation (using the coefficient of determination R² and root mean square error RMSE as indicators).

[0159] S53: The Stacking ensemble training process is as follows:

[0160] First, four base learners are trained independently on the training set;

[0161] Then, cross-validation or hold-out method is used to obtain the prediction outputs of each base learner on the validation set, and these prediction values ​​are used as new meta-feature vectors.

[0162] Finally, the meta-feature vectors are input into the Radial Basis Function Support Vector Regression (RBF-SVR) meta-learner for training. The meta-learner learns the residual characteristics and predictive advantages of each base learner on different samples, achieving adaptive weighted fusion. After training, the complete Stacking model (base learners combined with the meta-learner) is saved.

[0163] S54: Evaluate model performance using the test set, calculating R² (reflecting the proportion of explained variance) and RMSE (reflecting prediction bias). In this embodiment, the Stacking model achieved an R² of 0.786 on the test set and reduced its RMSE to the lowest level in the entire study, significantly outperforming a single base learner (the best single model has an R² of approximately 0.714) and the traditional GLM linear model (R² = 0.522).

[0164] To quantitatively compare the LAI inversion performance of different models, the prediction results of different models were compared on the validation set, and the results are shown in Table 2.

[0165]

[0166] Table 2. Performance comparison of different inversion models in LAI prediction;

[0167] From a single-model perspective, deep learning and machine learning algorithms demonstrate significantly stronger nonlinear fitting performance compared to traditional GLM methods, both outperforming traditional linear models in the inversion of corn LAI. Ensemble learning further optimizes LAI inversion accuracy by combining the predictive advantages of multiple heterogeneous models, and each ensemble method exhibits excellent robustness on the test set. In benchmark comparisons, MLP deep networks (R... 2 The accuracy (R² = 0.714) is outstanding in single-model learning, while ensemble learning strategies achieve a further leap in accuracy through algorithmic synergy. Voting strategies effectively improve prediction consistency (R² = 0.714) through simple averaging effects. 2 =0.708), while the fusion model based on the Stacking strategy, through a two-layer meta-learner, complements the advantages of RF, SVM, MLP and 1D-CNN, enhancing the stability and reliability of the model, and achieving the highest R on the test set. 2 (0.786), and its RMSE (0.3440) dropped to the lowest level in the entire study, with the inversion effect being the best among all models.

[0168] Single models have inherent limitations. While RF is robust, its ability to fit extreme nonlinear boundaries is limited; SVM is sensitive to kernel function selection; MLP is prone to overfitting on small-sample farmland data; and while 1D-CNN can extract local patterns, its global context capture is weak. Each base learner exhibits complementary error patterns under different tillage treatments leading to feature distribution shifts. For example, RF performs better in texture-dominated regions, while CNN excels in regions where spectral and texture elements are combined.

[0169] The Meta-Learner (RBF-SVR) learns the optimal fusion weights specific to each sample through nonlinear mapping, effectively hedging against errors. Particularly in the complex context of high field heterogeneity and non-stationary feature-LAI relationships caused by 13 different tillage treatments, it significantly improves the model's robustness and spatial continuity. This effectively mitigates the overfitting risk of single deep learning models and addresses the shortcomings of traditional machine learning models in characterizing complex nonlinear boundaries.

[0170] Step S6: Use the trained inversion model to perform pixel-level spatial inversion of maize leaf area index (LAI) in the target farmland area, generate spatial distribution maps of LAI at different growth stages, and further calculate the duration of leaf area to evaluate the growth response under different tillage measures.

[0171] S61: The nine core features selected from S4 are input into the trained Stacking model to perform pixel-level parallel inference on the preprocessed image of the entire target region, obtaining the LAI prediction value for each pixel. The inference process is implemented in batches using Python's rasterio and joblib to ensure efficient generation of the entire image.

[0172] S62: Generate spatial distribution maps of LAI (Laminated Area Ion) for the four growth stages: jointing stage, stamening stage, silking stage, and grain-filling stage (corresponding to...). Figure 7 The color in the image gradually changes from orange-yellow (low LAI) to dark green (high LAI), clearly demonstrating the dynamic process of the canopy from limited coverage in the seedling stage to complete closure in the grain-filling stage. At a fine spatial scale, deep plowing and deep loosening treatments show obvious high-value dark green bands in the later stages, the straw-returning plots are darker in color than the non-returning plots, and the nitrogen topdressing treatment (-N) shows more continuous and stable high-value bands in the grain-filling stage, intuitively reflecting the comprehensive regulatory effect of tillage, straw, and nitrogen fertilizer.

[0173] Further statistical analysis of pixel frequency distribution (corresponding to) Figure 8 The distribution of LAI during the jointing stage is skewed to the left, with the peak value concentrated in the low value range; during the tasseling stage, the main peak shifts to the right to the middle value range; the distribution range is widest and the peak value is highest during the silking and grouting stages, and they are close to symmetrical single peaks. This quantitatively reveals the transition law of LAI from the low value area to the high value area and the process of gradual intensification of differentiation between treatments.

[0174] S63: The integral method is used to calculate the leaf area duration (LAD) to quantify the canopy development potential throughout the entire growth period. The formula is as follows:

[0175]

[0176] In the formula:

[0177] LAD represents leaf area duration (m²·d / m²).

[0178] n represents the total number of observation intervals;

[0179] LAI n This represents the mean leaf area index (m² / m²) at the nth observation.

[0180] LAI n+1 This represents the mean leaf area index (m² / m²) at the (n+1)th observation.

[0181] T n This represents the number of days in the growing season corresponding to the nth observation (d, based on the number of days after sowing or the standard phenological period).

[0182] T n+1 This represents the number of days (d) of the reproductive period corresponding to the (n+1)th observation.

[0183] The measured or inverted LAI values ​​for the four reproductive periods are arranged in chronological order. For two adjacent observations, the trapezoidal area (average LAI multiplied by the time interval) is calculated and summed to obtain the total LAD for each treatment and the stage-specific LAD for each stage (jointing-ejaculation, ejaculation-silking, silking-filling).

[0184] Differences in LAI evolution throughout the entire growth period under different tillage practices, based on the dynamic evolution trend of leaf area index of maize throughout the entire growth period under different tillage practices ( Figure 9As can be seen, throughout the observation period, the leaf area index (LAI) of maize in all treatment groups showed a continuous upward trend without any downward inflection point, indicating that different tillage patterns could support the population to maintain canopy development for a relatively long period. Quantitative analysis based on the average growth rate table for each stage (Table 3) revealed significant stage-specific differences in growth intensity among the treatments: During the explosive growth period from jointing to tasseling, DT2-N showed the best performance with a daily average growth rate of 0.0294 days, significantly higher than the control group CK1 (0.0264 days), demonstrating the strong promoting effect of deep plowing combined with nitrogen fertilizer regulation on early canopy construction; after entering the tasseling-silking stage, the growth rate generally declined, with SDR2 showing strong mid-term growth sustainability with a growth rate of 0.0215 days; while when the growth rate dropped to its lowest point during the silking-grain-filling stage, SB2 still maintained the highest residual growth rate (0.0123 days), demonstrating its potential in delaying leaf senescence. However, due to the significant accumulated advantages in the early and mid-term stages, CK1, DT2-N, and NT1 ultimately maintained a higher LAI level at the end of the observation period.

[0185]

[0186] Table 3. Average growth rate of LAI in maize at different growth stages under different tillage practices;

[0187] The results of the analysis of variance on maize LAI at different growth stages (Table 4) show that the effects of each treatment on LAI exhibit significant stage-specific changes as the growing season progresses. In the early growth stage (0621 jointing stage), the p-values ​​for tillage method (T), straw return to the field (S), and nitrogen management (N) were all greater than 0.05, indicating no significant difference (ns). This is consistent with the observation results in section 2.5.1 above, where the initial LAI levels of each treatment were similar at the beginning of jointing. As the growth stage progressed, the effects of each main factor gradually increased: tillage method (T) reached a significant level at the tasseling stage (P<0.05) and a highly significant level at the silking stage (P<0.01), becoming the core factor driving canopy development and differentiation; the influence of nitrogen management (N) also steadily increased with the progression of phenological stages, showing a highly significant effect at the grain-filling stage (P=0.005), which explains why the nitrogen fertilizer regulation group, such as DT2-N, could maintain a high absolute level of LAI in the later stages; straw return to the field had a significant supporting effect on LAI after the silking stage (P<0.05). Throughout the entire growth period, the interaction between tillage method, straw return to the field, and nitrogen management (T×S×N) did not reach a significant level (P>0.05). This indicates that under the conditions of this experiment, the effects of the three main factors on maize LAI were relatively independent, and the combination of factors was more of a linear superposition rather than a complex nonlinear intervention.

[0188] Farming method (T) 0.521 (ns) 0.028(*) 0.007(**) 0.011(*) Straw returned to the field (S) 0.435 (ns) 0.112 (ns) 0.041(*) 0.038(*) Nitrogen supplementation management (N) 0.712 (ns) 0.048(*) 0.012(*) 0.005(**) T×S×N (Interactive) 0.885 (ns) 0.312 (ns) 0.184 (ns) 0.102 (ns)

[0189] Table 4. Analysis of variance on the effects of tillage methods, straw return to the field, and nitrogen fertilizer management on the LAI of maize at different growth stages;

[0190] Note: ns indicates no significance (P>0.05), (*) indicates significance at the 0.05 level (P<0.05), and (*) indicates significance at the 0.01 level (P<0.01).

[0191] Table 5 shows the calculation results of each growth stage and total LAD, indicating that different tillage measures have significantly different effects on improving maize growth potential. In terms of total LAD, CK1, DT2-N, and NT1 lead with high values ​​of 196.06, 194.80, and 194.31, respectively, which is highly consistent with the observed high LAD levels at the end of the growth period. Looking at the performance of each treatment group, DT2-N in the deep-plowing group, with its higher LAD in the early growth stage... 59-89d A powerful build and mid-game (LAD) with a score as high as 82.22 89-113d The stable maintenance of 80.56 LAD demonstrates optimal cumulative growth potential. Among the no-till groups, NT1 showed the most outstanding performance, with a balanced distribution across all growth stages, and its total LAD closely followed DT2-N. In contrast, although deep tillage treatment exhibited better growth intensity in the later stages, its total LAD was still lower than the dominant group due to the smaller initial base. Furthermore, comparisons among the various adjustment groups revealed that scientific nitrogen management generally increased LAD values, confirming the significant driving effect of nitrogen fertilizer on canopy development in the mid-to-late stages. In conclusion, DT2-N and NT1 demonstrate significant advantages in coordinating development at each stage and accumulating canopy size, and can be considered the superior tillage measures under the conditions of this experiment.

[0192]

[0193] Table 5. Assessment of Leaf Area Duration (LAD) at Different Growth Stages of Maize under Different Tillage Practices;

[0194] By comparing the LAD values ​​and growth rates of the 13 treatments (Tables 3 and 5), it was found that DT2-N performed outstandingly in the early stage of construction (highest LAD during the jointing-tasseling stage) and the mid-stage maintenance, with the total LAD at a leading level. Combined with the results of the analysis of variance (Table 4), the tillage method (T) reached a significant or highly significant level after the tasseling stage, nitrogen management (N) had a highly significant impact during the grain-filling stage, and straw return (S) significantly supported the maintenance of LAD after the silking stage.

[0195] This study employs a multi-dimensional joint analytical system, utilizing spatial distribution maps, pixel frequency statistics, analysis of variance (ANOVA), and LAD integrals, rather than relying on a single indicator for evaluation. Spatial maps provide an intuitive understanding of the overall pattern, pixel distribution offers statistical evidence of evolution, ANOVA quantifies the main effects and interactions of various factors, and LAD integrals extend the concept of photosynthetically intercepted area over time. Comprehensive analysis accurately identifies deep plowing combined with straw return and nitrogen topdressing (DT2-N) as exhibiting the strongest canopy expansion rate and anti-aging ability during the tasseling to grain-filling stage. By ensuring the continuous and effective extension of the photosynthetically effective radiation interception area throughout the entire growth period, it achieves optimal accumulation of growth potential. This treatment has been selected as a preferred tillage model suitable for promotion in cold and arid regions.

[0196] The Stacking architecture provided by this invention can be widely applied to other scenarios that require the fusion of multi-source heterogeneous features for regression prediction, and improves the robustness of the model through heterogeneous base learners and adaptive meta-learners.

[0197] This invention is not limited to the above embodiments. Various changes and modifications can be made without departing from the spirit and scope of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the scope defined by the claims of this invention should be included within the protection scope of this invention.

[0198] The embodiments of the present invention are given for illustrative and descriptive purposes only, and are not intended to be exhaustive or to limit the invention to the forms disclosed. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of the invention, and to enable those skilled in the art to understand the invention and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A remote sensing inversion system for maize leaf area index, characterized in that, include: A multi-model computation layer, configured to receive external input data, includes: An integrated computing unit performs a first processing on the external input data based on an integrated computing model to generate first intermediate data; The kernel processing unit performs a second processing on the external input data based on the kernel processing model to generate second intermediate data; The feedforward network processing unit performs third processing on the external input data based on the feedforward neural network model to generate third intermediate data. The convolutional network operation unit performs a fourth processing on the external input data based on the convolutional neural network model to generate fourth intermediate data; The integrated computing layer has its input terminals connected to the output terminals of the multi-model computing layer. It has a built-in kernel combination model for performing composite operations on the received first intermediate data, second intermediate data, third intermediate data, and fourth intermediate data, and outputting the result data. The integrated computing unit, the kernel operation unit, the feedforward network operation unit, and the convolutional network operation unit process the external input data from different perspectives, including integrated learning, kernel methods, global mapping, and local pattern extraction. The comprehensive operation layer fuses the first to fourth intermediate data through an adaptive weighting mechanism to improve the accuracy of the result data. The external input data is a set of remote sensing features that have undergone preprocessing and feature filtering; The multi-source feature input and preprocessing module is used to collect UAV multispectral image data during the key growth period of maize, perform radiometric calibration, reflectance correction, and generate vegetation masks based on soil-adjusted vegetation index to eliminate soil background interference. The multi-source feature extraction and multi-criteria screening module has its input end connected to the output end of the multi-source feature input and preprocessing module. It is used to extract normalized vegetation index features, soil-adjusted vegetation index features, optimized soil-adjusted vegetation index features, texture features based on gray-level co-occurrence matrix, and digital terrain model features. It constructs a key feature combination that is highly correlated with leaf area index through weighted fusion screening using Pearson correlation analysis, mutual information method, and Lasso regression. The key feature combination is then provided as the external input data to the multi-model operation layer. The prediction and application layer receives the result data output by the comprehensive computing layer at its input end, and is used to perform pixel-level leaf area index regression prediction on the target farmland area, generate spatial distribution maps of each growth stage and calculate the duration of leaf area, so as to assess the impact of different tillage practices on maize canopy development and growth potential. The multi-model computation layer and the integrated computation layer constitute the Stacking ensemble learning architecture. By combining heterogeneous base learners and meta-learners, the mapping relationship between the key features and leaf area index is learned from the perspectives of ensemble learning, kernel methods, global mapping and local pattern extraction, thereby achieving high-precision leaf area index inversion.

2. The remote sensing inversion system for maize leaf area index according to claim 1, characterized in that, The ensemble computing unit is a random forest regressor, which generates the first intermediate data based on the ensemble learning method; the kernel operation unit is a kernel support vector regressor, which generates the second intermediate data based on the kernel method; the feedforward network operation unit is a multilayer perceptron network, which generates the third intermediate data based on the feedforward neural network architecture. The convolutional network operation unit is a one-dimensional convolutional neural network, and the fourth intermediate data is generated based on the convolutional neural network architecture. The kernel combination model in the integrated computing layer is a radial basis support vector regression model, which is configured to use the first intermediate data, the second intermediate data, the third intermediate data, and the fourth intermediate data as input features, perform secondary training, and achieve integration through an adaptive weighting mechanism to output the result data.

3. A remote sensing inversion method for maize leaf area index, characterized in that, Using the maize leaf area index remote sensing inversion system described in any one of claims 1-2, a key feature set is constructed by fusing multi-source feature information and employing a multi-criteria feature selection strategy. A Stacking ensemble learning architecture, including heterogeneous base learners and meta-learners such as random forests, support vector machines, multilayer perceptrons, and one-dimensional convolutional neural networks, is used to achieve high-precision remote sensing inversion of farmland leaf area index. The method includes the following steps: S1: Collect multi-source input feature data and corresponding target variable label data; S2: Preprocess the input feature data to extract effective feature regions and eliminate background interference; S3: Extract statistical features, texture features, and structural features from the preprocessed data to construct a multi-source feature candidate set; S4: A multi-criteria comprehensive evaluation method is used to screen the multi-source feature candidate set to obtain the key feature combination for model training; S5: Construct multiple heterogeneous base learners, including random forest, support vector machine, multilayer perceptron and one-dimensional convolutional neural network, and use the stacking ensemble learning strategy to train the meta-learner to adaptively fuse the prediction outputs of each base learner to form a computational model based on the stacking ensemble learning architecture. S6: Utilize the trained Stacking ensemble learning architecture to perform sample-level target variable prediction for the target region, generate prediction results, and further calculate a comprehensive evaluation index to assess the response under different conditions.

4. The method for remote sensing inversion of maize leaf area index according to claim 3, characterized in that, Step S1 specifically includes: during the key growth stages of maize, namely the jointing stage, tasseling stage, silking stage, and grain-filling stage, randomly collecting maize plant samples using a fixed quadrat method and measuring leaf length and width parameters; based on the sampled leaf geometric parameters, planting density, and leaf area correction coefficient, determining the measured leaf area index value on the ground using the leaf area index calculation formula, and using it as training label data for the inversion model; Simultaneously acquire UAV multispectral remote sensing image data corresponding to ground sampling points.

5. The method for remote sensing inversion of maize leaf area index according to claim 3, characterized in that, Step S2 specifically includes: using a multispectral drone to collect raw image data containing green light, red light, red edge and near-infrared bands; Radiometric calibration, reflectance correction, and image stitching are performed on the raw image data to generate high-resolution orthophotos. Based on the soil-adjusted vegetation index combined with the Otsu threshold segmentation algorithm, a vegetation mask is automatically generated and soil background pixel interference is removed.

6. The method for remote sensing inversion of maize leaf area index according to claim 3, characterized in that, Step S3 specifically includes: extracting normalized vegetation index features, soil-adjusted vegetation index features, optimizing soil-adjusted vegetation index features, and multiple vegetation index features to characterize crop spectral response. Based on the gray-level co-occurrence matrix, texture analysis is performed on the red edge band and typical vegetation indices. Texture indices such as mean and information entropy are extracted and multi-directional averages are taken to describe the spatial heterogeneity of the canopy. Digital terrain models are extracted as terrain features to reflect the regulatory role of micro-topography on the distribution of water and fertilizer in the field.

7. The method for remote sensing inversion of maize leaf area index according to claim 3, characterized in that, The specific steps of S4 include: S41: evaluating the linear correlation strength between features and leaf area index through Pearson correlation analysis, quantifying nonlinear dependencies through mutual information method, and achieving feature sparse selection through Lasso regression; The feature scores of the three evaluation methods are normalized and then weighted and fused to construct a comprehensive feature importance evaluation index. Based on the comprehensive evaluation results, a core feature combination covering three dimensions—spectrum, texture, and terrain—was selected for subsequent model training.

8. The method for remote sensing inversion of maize leaf area index according to claim 3, characterized in that, Step S5 specifically includes: selecting random forest, support vector machine, multilayer perceptron and one-dimensional convolutional neural network as base learners to characterize the mapping relationship between features and leaf area index from different perspectives; The sample data is divided into training and test sets according to a preset ratio. The prediction results of each base learner on the validation samples are used as new input features and fed into the radial basis support vector regression meta-learner for secondary training to achieve adaptive weighted fusion of the prediction advantages of each base learner. The coefficient of determination and root mean square error are used as evaluation indicators to verify and optimize the generalization performance of the Stacking ensemble learning architecture.

9. The remote sensing inversion method for maize leaf area index according to claim 8, characterized in that, Step S6 specifically includes: inputting the selected key features into the trained Stacking ensemble learning architecture to perform pixel-level leaf area index prediction on the entire target region image. Spatial distribution maps of leaf area index for maize at each growth stage from jointing to grain filling were generated to quantitatively characterize the spatiotemporal differences in canopy development under different tillage treatments. The duration of leaf area at different growth stages for each treatment was calculated using the integral method to comprehensively evaluate the impact of tillage measures on maize growth potential. The Stacking ensemble learning strategy complements and fuses the residual characteristics and prediction bias of each base learner in a high-dimensional heterogeneous feature space, and then uses a meta-learner to achieve adaptive weighting. The multi-criteria comprehensive evaluation method integrates the linear synergistic information captured by Pearson correlation analysis, the nonlinear dependency patterns extracted by mutual information method, and the sparse feature subsets achieved by Lasso regression in a weighted manner. The selected core feature combination takes into account the effects of soil background suppression during the seedling stage, canopy saturation relief in the middle and late stages, and micro-topography water and fertilizer regulation.

10. The application of the remote sensing inversion method for maize leaf area index according to claim 3, characterized in that, By predicting leaf area index at the pixel level during the key growth stages from jointing to grain filling, spatial distribution maps of leaf area index for each growth stage are generated. The duration of leaf area is calculated using the integral method. Combined with pixel frequency distribution statistics and variance analysis of tillage factors, the impact of different tillage practices on maize canopy development and growth potential is evaluated.

Images