A corn leaf area index estimation method and system based on multi-source remote sensing and disease perception dual-task deep learning

Abstract

The application provides a corn leaf area index estimation method and system based on multi-source remote sensing and disease perception double-task deep learning. The method comprises the following steps: acquiring multi-source remote sensing data in the field, including ground hyperspectral, unmanned aerial vehicle multispectral and measured LAI data; extracting disease sensitive features based on hyperspectral, and extracting LAI sensitive features based on multispectral; constructing a double-task constrained deep neural network, a disease identification branch taking disease sensitive features as input and output disease grade, and an LAI estimation branch taking the disease identification result as prior constraint and outputting LAI estimation value after fusing the LAI sensitive features; adopting a non-forgetting incremental learning strategy, training the disease identification branch first, then introducing the LAI estimation task, retaining the disease identification knowledge through distillation loss, and optimizing the LAI estimation capability through new task loss; and using the trained model to identify diseases and estimate LAI. The method integrates the disease identification result as prior constraint into LAI estimation, and improves the precision and stability of corn LAI estimation under disease stress.

Description

Technical Field

[0001] This application relates to the fields of agricultural remote sensing, crop phenotypic monitoring, crop disease identification and smart agriculture, and in particular to a method and system for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception. Background Technology

[0002] Maize is an important food crop in my country. The leaf area index (LAI) is a core phenotypic parameter characterizing its canopy structure, canopy growth, light interception capacity, and biomass formation. Accurate LAI acquisition is a crucial foundation for growth monitoring, yield prediction, variety selection, and precision field management. Traditional destructive sampling and canopy analyzer measurements suffer from low efficiency and difficulty in achieving continuous high-throughput monitoring. Remote sensing inversion methods, with their advantages of wide coverage and non-contact operation, have become the mainstream technical approach. However, existing models still have significant shortcomings in accuracy and stability under disease stress conditions.

[0003] Leaf spot diseases in maize cause chlorophyll degradation, cell structure damage, leaf water loss, and premature senescence of lower leaves. These diseases not only alter the hyperspectral reflectance characteristics of leaves but also disrupt the inherent mapping relationship between the canopy spectrum and the leaf area index (LAI). Existing LAI estimation models are mostly single-task regression models that do not consider disease severity as a priori constraint information, neglecting the intrinsic physiological connection between disease identification and LAI estimation. This makes it difficult to effectively eliminate estimation errors caused by factors such as lesions and spectral saturation. Furthermore, a single sensor cannot simultaneously capture subtle physiological signals of the disease and cover a large spatial area. Continuous modeling across multiple growth stages is also prone to catastrophic forgetting problems, resulting in insufficient model generalization ability and failing to meet the practical needs of high-throughput phenotypic monitoring of maize under disease stress.

[0004] Therefore, how to actively introduce disease information into the LAI estimation process under disease stress conditions and achieve collaborative learning between disease identification and LAI estimation is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides a method and system for estimating maize leaf area index (LAI) based on dual-task deep learning of multi-source remote sensing and disease perception. This method incorporates disease identification results as prior constraint information into the LAI estimation branch, which significantly improves the accuracy and stability of maize LAI estimation under disease stress conditions.

[0006] Firstly, a method for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception is provided, the method comprising: S1: Acquire multi-source remote sensing data of corn fields, including: ground hyperspectral data, UAV multispectral images, measured leaf area index data and corresponding disease level data; S2: Extract disease-sensitive spectral features based on the ground hyperspectral data, and extract leaf area index (LAI)-sensitive spectral features based on the UAV multispectral imagery; S3: Construct a dual-task constrained deep neural network, including a disease identification branch and a LAI estimation branch. The disease identification branch takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The LAI estimation branch takes the output of the disease identification branch as the disease prior constraint information, fuses it with the LAI-sensitive multispectral features, and outputs the estimated value of maize leaf area index. S4: A non-forgetting incremental learning strategy is used to train a dual-task deep neural network. First, the disease identification branch is trained, and then the LAI estimation task is introduced. During the training process, the original disease identification knowledge is retained through distillation loss, and the LAI estimation ability is optimized through the new task loss. S5: Use the trained model to identify diseases and estimate LAI in the target area of ​​corn, and output the disease level, LAI estimate and model accuracy evaluation results.

[0007] It should be understood that the dual-task constrained deep neural network constructed in this application includes a disease identification branch and a leaf area index (LAI) estimation branch. The disease identification branch takes disease-sensitive spectral features as input and outputs disease level or probability; while the LAI estimation branch does not operate independently, but rather uses the output of the disease identification branch as prior disease constraint information, fuses it with the LAI-sensitive multispectral features, and then uses both as input to finally output the LAI estimate. This end-to-end bidirectional design, where "identification results directly constrain the estimation process," enables the model to actively perceive the perturbation of the canopy spectral-structural relationship by disease stress, thereby maintaining the stability and interpretability of the estimation under disease conditions.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes a corn disease inoculation and growth period observation step performed before step S1: artificial inoculation is carried out during the key growth period of corn, and the disease level is divided according to the proportion of lesion area, the average disease index of the plot is calculated, and the disease level is divided into mild, moderate, severe and highly susceptible.

[0009] It should be understood that in agronomic field trials, a "plot" is the smallest experimental unit artificially divided in a field trial for arranging a single treatment (such as a specific variety, planting density, or disease inoculation treatment). In the embodiments of this application, four types of plots were set up: maize density test, variety material, disease inoculation treatment, and healthy control. Protective rows were set between each plot to prevent cross-infection of diseases. All hyperspectral measurements, LAI measurements, and UAV image feature extraction were organized and observed using the plot as the basic unit to ensure an accurate correspondence between each sample data and the field treatment conditions.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, step S2 includes: First, correlation analysis is used to retain candidate features whose absolute values ​​of correlation coefficients with measured leaf area index or disease severity are higher than a set threshold. The set threshold is determined experimentally based on specific data. Then, a recursive feature elimination method based on random forest is used to screen key features for disease identification and measured leaf area index estimation at different growth stages. The key features refer to spectral features related to disease severity or leaf area index.

[0011] It should be understood that the above screening strategy can effectively reduce redundant information in hyperspectral and multispectral data, retain the most representative features of sensitivity to diseases and LAI, thereby improving the training efficiency and generalization stability of the dual-task model under limited sample conditions.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, in step S3, the specific method by which the measured leaf area index estimation branch uses the output result of the disease identification branch as the prior constraint information of the disease, and fuses it with the measured leaf area index sensitive multispectral features as input, is as follows: the disease level probability vector output by the disease identification branch is concatenated with the measured leaf area index sensitive multispectral feature vector to form a joint feature vector, which is then input into the fully connected network of the measured leaf area index estimation branch.

[0013] It should be understood that the disease level probability vector output by the disease identification branch is fused with the LAI-sensitive multispectral feature vector using a channel concatenation method: the four probability values ​​corresponding to the disease level are used as additional features and directly concatenated to the end of the LAI-sensitive feature vector to form a joint feature vector with expanded dimensions, which is then input into the fully connected network of the LAI estimation branch. This direct concatenation method at the feature level allows the fusion of prior disease information and canopy spectral structure information to be completed at the network input, ensuring that the disease level serves as a constraint throughout the entire forward computation process of LAI estimation, achieving end-to-end joint learning.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, in the dual-task constrained deep neural network, the disease identification branch and the measured leaf area index estimation branch share some hidden layer features, and the output of the disease identification branch is directly connected to the input of the measured leaf area index estimation branch, forming an end-to-end unidirectional information flow structure, wherein disease information participates in the entire process of measured leaf area index estimation as a constraint condition.

[0015] It should be understood that in the neural network architecture of this application, the disease identification branch and the LAI estimation branch share some hidden layer features. This means that the two branches have a common representational basis in low-level feature extraction, which can mutually enhance their ability to understand spectral information. Simultaneously, the output of the disease identification branch is directly connected to the input of the LAI estimation branch, forming an end-to-end unidirectional information flow structure: disease level or probability, as prior constraint information, is unidirectionally transmitted from the identification branch to the estimation branch, without reverse flow. This design ensures that disease information continuously participates in every layer of LAI estimation calculation throughout the entire forward propagation process, enabling the model to dynamically adjust the mapping relationship between the spectrum and canopy structure under disease stress conditions, thereby significantly improving the robustness and interpretability of LAI estimation.

[0016] In conjunction with the first aspect, in some implementations of the first aspect, step S4 includes: First, the disease identification branch is trained to convergence using only disease grade label data, which consists of four grades: mild, moderate, severe, and highly susceptible, obtained through field observation and disease index calculation. Then, while keeping the disease identification branch network structure unchanged, a measured leaf area index estimation task is introduced. The output of the disease identification branch is used as a fixed or differentiable prior feature, and the measured leaf area index estimation branch is trained together with the measured leaf area index data. During the training process, distillation loss is used to constrain the output distribution of the disease identification branch to prevent significant shifts.

[0017] In conjunction with the first aspect, in some implementations of the first aspect, the non-forgetting incremental learning strategy employs a knowledge distillation mechanism. The trained disease identification branch is used as the teacher network, and the difference between its output soft label and the corresponding output of the student network is used as the distillation loss. The student network is the disease identification branch in the dual-task deep neural network model. The distillation loss is weighted and summed with the mean squared error regression loss of the measured leaf area index estimation task, and used as the total loss function for joint optimization. This ensures that the dual-task deep neural network model retains its ability to discriminate against old tasks while learning new tasks, where the old task is disease identification.

[0018] It should be understood that this application first trains the disease identification branch to convergence using only disease label data, enabling it to possess stable disease discrimination capabilities. Subsequently, while maintaining the network structure of the disease identification branch, a LAI estimation task is introduced. The output of the disease identification branch is used as a fixed or differentiable prior feature, and the LAI estimation branch is trained jointly with the measured LAI data. To prevent the learning of new tasks from corrupting the acquired disease identification knowledge, distillation loss is introduced during training. The trained disease identification branch is used as the teacher network, and the difference between its output soft label and the corresponding output of the student network (i.e., the disease identification branch in the dual-task model) is used as the distillation loss, which is then weighted and jointly optimized with the regression loss of the LAI estimation task. In this way, while optimizing the LAI estimation capability, the model can constrain the output distribution of the disease identification branch from significant shifts, thereby effectively preserving the discrimination capability for the old task (disease identification) and achieving knowledge consolidation and task expansion in multi-fertility continuous modeling.

[0019] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes a UAV multispectral image preprocessing step: performing image stitching, radiometric correction, reflectance conversion and orthophoto generation on the original UAV images, performing spatial registration using ground control points, outputting GeoTIFF format images, and extracting cell-scale multispectral reflectance and vegetation index based on the boundaries of the experimental cell.

[0020] Secondly, a maize leaf area index estimation system based on dual-task deep learning of multi-source remote sensing and disease perception is provided, the system comprising: The data acquisition module is used to acquire multi-source remote sensing data in the corn field. The multi-source remote sensing data includes ground hyperspectral data, UAV multispectral images, measured leaf area index data, and corresponding disease level data. The feature extraction module is used to extract disease-sensitive spectral features based on the ground hyperspectral data and to extract leaf area index-sensitive spectral features based on the UAV multispectral image. The dual-task deep neural network module includes a disease identification submodule and a leaf area index estimation submodule. The disease identification submodule takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The leaf area index estimation submodule takes the disease level or disease probability output by the disease identification submodule as prior constraint information, fuses it with the leaf area index-sensitive spectral features, and outputs the estimated maize leaf area index value. The incremental learning training module is used to train the dual-task deep neural network module using a non-forgetting incremental learning strategy. First, the disease identification sub-module is trained, and then the leaf area index estimation task is introduced. The original disease identification knowledge is retained through distillation loss, and the leaf area index estimation ability is optimized through the new task loss. The output module is used to output the disease severity level, estimated leaf area index, and model accuracy evaluation results.

[0021] In conjunction with the second aspect, in some implementations of the second aspect, in the dual-task deep neural network module, the output of the disease identification branch is directly connected to the input of the LAI estimation branch. The disease level probability vector and the LAI sensitive spectral feature vector are concatenated and then input into the LAI estimation branch to form an end-to-end unidirectional information constraint structure. Attached Figure Description

[0022] Figure 1 The flowchart illustrates a method for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception, as provided in this application embodiment.

[0023] Figure 2 This is a schematic diagram of a dual-task constrained deep neural network structure provided in an embodiment of this application. Detailed Implementation

[0024] The terminology used in the following embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to also include expressions such as “one or more,” unless the context clearly indicates otherwise. It should also be understood that in the following embodiments of this application, “at least one” and “one or more” refer to one, two, or more than two. The term “and / or” is used to describe the relationship between related objects, indicating that three relationships may exist; for example, A and / or B can indicate: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character “ / ” generally indicates that the preceding and following related objects are in an “or” relationship.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] Maize is an important food crop in my country, and its leaf area index (LAI) is a core phenotypic parameter characterizing its canopy structure and canopy growth, serving as a crucial foundation for growth monitoring, yield prediction, and precision field management. Traditional destructive sampling and canopy analyzer measurements are inefficient and struggle to achieve high-throughput continuous monitoring. While remote sensing inversion has become the mainstream technology, its accuracy and stability under disease stress are significantly insufficient. Maize leaf spot diseases disrupt the inherent mapping relationship between canopy spectra and LAI. Existing single-task models do not incorporate prior constraints on diseases and suffer from sensor limitations and catastrophic forgetting problems in multi-growth-stage modeling. Therefore, achieving collaborative learning between disease identification and LAI estimation is a pressing technical challenge in this field.

[0027] This application provides a method and system for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception, which can effectively overcome the above-mentioned problems.

[0028] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings.

[0029] Figure 1 The flowchart illustrates a method for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception, as provided in this application embodiment.

[0030] refer to Figure 1 In some examples, the method includes: S1: Acquire multi-source remote sensing data of corn fields, including: ground hyperspectral data, UAV multispectral images, measured leaf area index data and corresponding disease level data; S2: Extract disease-sensitive spectral features based on the ground hyperspectral data, and extract leaf area index-sensitive spectral features based on the UAV multispectral imagery; S3: Construct a dual-task constrained deep neural network, including a disease identification branch and a LAI estimation branch. The disease identification branch takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The LAI estimation branch takes the output of the disease identification branch as the disease prior constraint information, fuses it with the LAI-sensitive multispectral features, and outputs the estimated value of maize leaf area index. S4: A non-forgetting incremental learning strategy is used to train a dual-task deep neural network. First, the disease identification branch is trained, and then the LAI estimation task is introduced. During the training process, the original disease identification knowledge is retained through distillation loss, and the LAI estimation ability is optimized through the new task loss. S5: Use the trained model to identify diseases and estimate LAI in the target area of ​​corn, and output the disease level, LAI estimate and model accuracy evaluation results.

[0031] In one possible implementation, this embodiment of the application was conducted at the Xinxiang Comprehensive Experimental Base of the Chinese Academy of Agricultural Sciences. The experimental plot was 52m long from east to west and 81.75m long from north to south. Four types of plots were set up: maize density test, variety material, disease inoculation, and healthy control. The disease treatment was maize Curvularialunata leaf spot. Protective rows were set up between different treatment replicates to prevent cross-infection. Artificial inoculation was carried out using motorized sprayers during the maize growing season. Continuous observation was conducted at four key growth stages: VT, R1, R2, and R3. The severity of the disease was divided into 1-9 levels according to the proportion of lesion area. After calculating the average disease index of each plot, the samples were further divided into four categories: mild, moderate, severe, and highly susceptible.

[0032] Ground data collection and field observation were carried out simultaneously. The reflectance at the leaf scale of maize was collected using an ASDFieldSpecPRO4 hyperspectral analyzer. Fixed plants were selected in each plot, and measurements were taken at the 1 / 3, 1 / 2, and 2 / 3 positions of the leaves. Calibration was completed using leaf clips with built-in light sources and standard white boards. The obtained data were used to extract disease-sensitive spectral characteristics. The leaf area index (LAI) was measured using the SunScan canopy analysis system. Each plot was observed three times along the 0°, 45°, 90°, and 135° directions of the sowing row, and the average value of 12 sets of data was taken as the measured LAI value of the plot.

[0033] Multispectral image acquisition by UAV was conducted simultaneously with ground data collection. A DJI Motrice M600 Pro hexacopter UAV equipped with a MicaSense RedEdge-MX five-band multispectral camera was used to acquire images in the blue, green, red, red-edge, and near-infrared bands. Flight operations were conducted between 10:00 and 14:00 on clear, cloudless days, at an altitude of 50m and a speed of 2m / s, with a forward and lateral overlap of 85%. Multispectral calibration plates were photographed before and after each flight for radiometric correction and reflectivity conversion, ultimately outputting raw multispectral images of the corn canopy.

[0034] In some examples, the method further includes maize disease inoculation and growth period observation steps prior to step S1: artificial inoculation is carried out during the critical growth period of maize, and the disease level is divided according to the proportion of lesion area, the average disease index of the plot is calculated, and the disease level is divided into mild, moderate, severe and highly susceptible.

[0035] In some examples, step S2 includes: First, correlation analysis is used to retain candidate features whose absolute values ​​of correlation coefficients with measured leaf area index or disease severity are higher than a set threshold. The set threshold is determined experimentally based on specific data. Then, a recursive feature elimination method based on random forest is used to screen key features for disease identification and measured leaf area index estimation at different growth stages. The key features refer to spectral features related to disease severity or leaf area index.

[0036] In one possible implementation, during the feature construction stage, based on preprocessed ground hyperspectral data, the original band reflectance, first derivative spectrum, second derivative spectrum, spectral angle mapping features, and disease-sensitive spectral features such as difference vegetation index (DVI), ratio vegetation index (RVI), and normalized vegetation index (NDVI) are calculated. Based on UAV multispectral imagery, RGB color index, visible light vegetation index, red edge index, near-infrared index, and LAI-sensitive spectral features such as modified soil-adjusted vegetation index (MSAVI), modified soil-adjusted vegetation index 2 (MSAVI2), optimized soil-adjusted vegetation index (OSAVI), normalized red edge index (NDRE), greenness normalized vegetation index (GNDVI), and enhanced vegetation index (EVI) are calculated.

[0037] In the feature selection stage, the correlation coefficient between each feature and the measured LAI or disease severity was first calculated by Pearson correlation analysis, and candidate features with high correlation were retained. Then, a recursive feature elimination method based on random forest was used to select the top 10 optimal key features for LAI estimation and disease identification based on the differences in canopy structure and disease performance at the four key growth stages of maize: VT, R1, R2, and R3.

[0038] In some examples, in step S3, the specific method by which the measured leaf area index estimation branch uses the output of the disease identification branch as the prior constraint information of the disease, and fuses it with the measured leaf area index sensitive multispectral features as input, is as follows: the disease level probability vector output by the disease identification branch is concatenated with the measured leaf area index sensitive multispectral feature vector to form a joint feature vector, which is then input into the fully connected network of the measured leaf area index estimation branch.

[0039] Figure 2 This is a schematic diagram of a dual-task constrained deep neural network structure provided in an embodiment of this application.

[0040] refer to Figure 2 ,refer to Figure 2In some examples, the LAI estimation branch uses the output of the disease identification branch as prior constraint information for the disease. The specific method for fusing this information with the LAI-sensitive multispectral features is as follows: the disease level probability vector output by the disease identification branch is concatenated with the LAI-sensitive multispectral feature vector through channels to form a joint feature vector, which is then input into the fully connected network of the LAI estimation branch, thereby achieving LAI inversion under disease perception constraints. (The appendix is ​​not provided in the original text.) Figure 2 The "disease perception feature representation" shown refers to the disease level probability distribution vector output by the disease identification branch; (See attached image) Figure 2 The “LAI inversion” shown refers to the complete forward process by which the LAI estimation branch calculates and outputs continuous LAI estimates based on the fused joint feature vector through a regression network.

[0041] As a preferred implementation, in this embodiment, the dual-task constrained deep neural network adopts a fully connected deep multilayer perceptron (MLP) architecture, consisting of a disease identification branch and a LAI estimation branch. The two branches share some hidden layer features, forming an end-to-end structure with unidirectional information flow. The above architecture corresponds to the attached... Figure 2 The “LWF model” shown here employs a non-forgetting incremental learning strategy during training. Through a knowledge distillation mechanism, it retains the discriminative ability of the old task of disease identification while introducing new tasks such as LAI estimation.

[0042] In some examples, in the dual-task constrained deep neural network, the disease identification branch and the LAI estimation branch share some hidden layer features, and the output of the disease identification branch is directly connected to the input of the LAI estimation branch, forming an end-to-end unidirectional information flow structure, in which disease information participates in the entire LAI estimation process as a constraint condition.

[0043] In this embodiment, the dual-task constrained deep neural network adopts a fully connected deep multilayer perceptron (MLP) architecture, which consists of a disease identification branch and a LAI estimation branch. The two branches share some hidden layer features at the bottom layer, forming an end-to-end structure with unidirectional information flow.

[0044] Disease identification branch (classification network): The input layer dimension is consistent with the dimension of the sensitive spectral features of the selected diseases; there are 4 hidden layers with 256, 128, 64 and 64 neurons respectively, and the ReLU function is used uniformly; the output layer has 4 neurons, corresponding to the four disease levels of mild, moderate, severe and highly susceptible, and the Softmax activation function is used to output the probability distribution vector of each category.

[0045] LAI estimation branch (regression network): The input layer receives the joint feature vector after channel splicing (the dimension is the sum of the LAI sensitive multispectral feature dimension and the disease level probability vector dimension); there are 3 hidden layers with 128, 64 and 32 neurons respectively, and the ReLU function is used uniformly; the output layer has 1 neuron and uses a linear activation function to output continuous LAI estimation values.

[0046] Branch connection relationship: The first two hidden layers of the disease identification branch are shared feature layers, and their output features are used by both the disease identification classification output and the LAI estimation branch. The output of the disease identification branch is directly connected to the input of the LAI estimation branch, and the disease level probability vector participates in the forward calculation process of LAI estimation as prior constraint information throughout the process.

[0047] In some examples, step S4 includes: First, the disease identification branch is trained to convergence using only disease grade label data, which consists of four grades: mild, moderate, severe, and highly susceptible, obtained through field observation and disease index calculation. Then, while keeping the disease identification branch network structure unchanged, a measured leaf area index estimation task is introduced. The output of the disease identification branch is used as a fixed or differentiable prior feature, and the measured leaf area index estimation branch is trained together with the measured leaf area index data. During the training process, distillation loss is used to constrain the output distribution of the disease identification branch to prevent significant shifts.

[0048] In some examples, the non-forgetting incremental learning strategy employs a knowledge distillation mechanism, using the trained disease identification branch as the teacher network, and the difference between its output soft label and the corresponding output of the student network as the distillation loss. The student network is the disease identification branch in the dual-task deep neural network model. The distillation loss is weighted and summed with the mean squared error regression loss of the measured leaf area index estimation task, and used as the total loss function for joint optimization. This allows the dual-task deep neural network model to retain its ability to discriminate against old tasks, namely disease identification, while learning new tasks.

[0049] In one possible implementation, during model training, the disease identification branch uses the cross-entropy loss function, and the LAI estimation branch uses the mean squared error or root mean square error loss function; the optimizer is Adam, with an initial learning rate of 1×10⁻⁶. -3 Batch size 32, maximum training epochs 200, L2 weight decay coefficient 1×10 -4 An early stopping strategy was employed to monitor the validation set loss and prevent overfitting. The distillation temperature T was set to 2, and the loss balance coefficient was set to... Set to 0.5, and the final output will be the intertemporal incremental learning LAI estimation model.

[0050] In some examples, the method further includes a UAV multispectral image preprocessing step: performing image stitching, radiometric correction, reflectance conversion and orthophoto generation on the original UAV images, spatial registration using ground control points, outputting GeoTIFF format images, and extracting cell-scale multispectral reflectance and vegetation index based on the boundaries of the experimental cell.

[0051] In one possible implementation, this embodiment of the application imports the original UAV imagery, positioning attitude information, and ground control points into image processing software during the data preprocessing stage to complete image stitching, radiometric correction, reflectance conversion, and orthophoto generation. The output is saved in GeoTIFF format. Subsequently, multispectral reflectance and vegetation indices at the cell scale are extracted based on the boundaries of the experimental cell.

[0052] This application also provides a maize leaf area index estimation system based on dual-task deep learning of multi-source remote sensing and disease perception. The system includes: The data acquisition module is used to acquire multi-source remote sensing data in the corn field. The multi-source remote sensing data includes ground hyperspectral data, UAV multispectral images, measured leaf area index data, and corresponding disease level data. The feature extraction module is used to extract disease-sensitive spectral features based on the ground hyperspectral data and to extract leaf area index-sensitive spectral features based on the UAV multispectral image. The dual-task deep neural network module includes a disease identification submodule and a leaf area index estimation submodule. The disease identification submodule takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The leaf area index estimation submodule takes the disease level or disease probability output by the disease identification submodule as prior constraint information, fuses it with the leaf area index-sensitive spectral features, and outputs the estimated maize leaf area index value. The incremental learning training module is used to train the dual-task deep neural network module using a non-forgetting incremental learning strategy. First, the disease identification sub-module is trained, and then the leaf area index estimation task is introduced. The original disease identification knowledge is retained through distillation loss, and the leaf area index estimation ability is optimized through the new task loss. The output module is used to output the disease severity level, estimated leaf area index, and model accuracy evaluation results.

[0053] In some examples, in the dual-task deep neural network module, the output of the disease identification branch is directly connected to the input of the LAI estimation branch. The disease level probability vector and the LAI sensitive spectral feature vector are concatenated and then input into the LAI estimation branch, forming an end-to-end unidirectional information constraint structure.

[0054] The above are merely preferred embodiments of this application. The scope of protection of this application is not limited to the above embodiments. Any equivalent modifications or variations made by those skilled in the art based on the content disclosed in this application should be included within the scope of protection set forth in the claims.

Claims

1. A method for estimating maize leaf area index based on dual-task deep learning of multi-source remote sensing and disease perception, characterized in that, The method includes: S1: Acquire multi-source remote sensing data of corn fields, including: ground hyperspectral data, UAV multispectral images, measured leaf area index data and corresponding disease level data; S2: Extract disease-sensitive spectral features based on the ground hyperspectral data, and extract leaf area index-sensitive spectral features based on the UAV multispectral imagery; S3: Construct a dual-task constrained deep neural network, including a disease identification branch and a measured leaf area index (LAI) estimation branch. The disease identification branch takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The measured LAI estimation branch takes the output of the disease identification branch as disease prior constraint information, fuses it with the measured LAI sensitive multispectral features, and outputs an estimated maize LAI value. The measured LAI estimation branch takes the output of the disease identification branch as disease prior constraint information, fuses it with the measured LAI sensitive multispectral features, and outputs an estimated maize LAI value. The specific method for using the spectral feature fusion as input is as follows: the disease level probability vector output by the disease identification branch is concatenated with the measured leaf area index sensitive multispectral feature vector to form a joint feature vector, which is then input into the fully connected network of the measured leaf area index estimation branch; the disease identification branch and the measured leaf area index estimation branch share some hidden layer features, and the output of the disease identification branch is directly connected to the input of the measured leaf area index estimation branch, forming an end-to-end unidirectional information flow structure, in which disease information participates in the entire process of measured leaf area index estimation as a constraint condition; S4: A non-forgetting incremental learning strategy is used to train a dual-task deep neural network. First, the disease identification branch is trained, and then the measured leaf area index (LAI) estimation task is introduced. During training, the original disease identification knowledge is retained through distillation loss, and the measured LAI estimation capability is optimized through the new task loss. Specifically, this includes: First, the disease identification branch is trained until convergence using only disease grade label data, which consists of four grades: mild, moderate, severe, and highly susceptible, obtained through field observation and disease index calculation. Then, while keeping the disease identification branch network structure unchanged, a measured leaf area index estimation task is introduced. The output of the disease identification branch is used as a fixed or differentiable prior feature, and the measured leaf area index estimation branch is trained together with the measured leaf area index data. During the training process, distillation loss is used to constrain the output distribution of the disease identification branch to prevent significant shifts. S5: Use the trained model to identify diseases and estimate the measured leaf area index of corn in the target area, and output the disease level, the estimated value of the measured leaf area index, and the model accuracy evaluation results.

2. The method according to claim 1, characterized in that, The method also includes the steps of corn disease inoculation and growth period observation before step S1: artificial inoculation is carried out during the key growth period of corn, and the disease level is divided according to the proportion of lesion area, the average disease index of the plot is calculated, and the disease level is divided into mild, moderate, severe and highly susceptible.

3. The method according to claim 1, characterized in that, Step S2 includes: First, correlation analysis is used to retain candidate features whose absolute values ​​of correlation coefficients with measured leaf area index or disease severity are higher than a set threshold. The set threshold is determined experimentally based on specific data. Then, a recursive feature elimination method based on random forest is used to screen key features for disease identification and measured leaf area index estimation at different growth stages. The key features refer to spectral features related to disease severity or leaf area index.

4. The method according to claim 1, characterized in that, The non-forgetting incremental learning strategy employs a knowledge distillation mechanism. The trained disease identification branch is used as the teacher network, and the difference between its output soft label and the corresponding output of the student network is used as the distillation loss. The student network is the disease identification branch in the dual-task deep neural network model. The distillation loss and the mean squared error regression loss of the measured leaf area index estimation task are weighted and summed to form the total loss function for joint optimization. This allows the dual-task deep neural network model to retain its ability to distinguish between old tasks, namely disease identification, while learning new tasks.

5. The method according to claim 1, characterized in that, The method also includes a UAV multispectral image preprocessing step: performing image stitching, radiometric correction, reflectance conversion and orthophoto generation on the original UAV images, spatial registration using ground control points, outputting GeoTIFF format images, and extracting cell-scale multispectral reflectance and vegetation index based on the boundaries of the experimental cell.

6. A maize leaf area index estimation system based on dual-task deep learning of multi-source remote sensing and disease perception, characterized in that, The system includes: The data acquisition module is used to acquire multi-source remote sensing data in the corn field. The multi-source remote sensing data includes ground hyperspectral data, UAV multispectral images, measured leaf area index data, and corresponding disease level data. The feature extraction module is used to extract disease-sensitive spectral features based on the ground hyperspectral data and to extract leaf area index-sensitive spectral features based on the UAV multispectral image. The dual-task deep neural network module includes a disease identification submodule and a leaf area index estimation submodule. The disease identification submodule takes the disease-sensitive spectral features as input and outputs the maize disease level or disease probability. The leaf area index estimation submodule takes the disease level or disease probability output by the disease identification submodule as prior constraint information, fuses it with the leaf area index-sensitive spectral features, and outputs the estimated maize leaf area index value. The incremental learning training module is used to train the dual-task deep neural network module using a non-forgetting incremental learning strategy. First, the disease identification sub-module is trained, and then the leaf area index estimation task is introduced. The original disease identification knowledge is retained through distillation loss, and the leaf area index estimation ability is optimized through the new task loss. The output module is used to output the disease severity level, estimated leaf area index, and model accuracy evaluation results.

7. The system according to claim 6, characterized in that, In the dual-task deep neural network module, the output of the disease identification branch is directly connected to the input of the measured leaf area index estimation branch. The disease level probability vector and the measured leaf area index sensitive spectral feature vector are concatenated and then input into the measured leaf area index estimation branch, forming an end-to-end unidirectional information constraint structure.

Images