A neural network model and system for corn lai estimation

Abstract

This application discloses a neural network model and system for estimating maize LAI (Laboratory Area Index), relating to the field of deep learning technology. A multi-source feature extraction network performs deep representation of input feature information, enhancing the fusion and expressive ability between features from different sources and at different scales. Inversion prediction values ​​and trend prediction values ​​are then obtained through physical inversion and phenological trend branches, respectively, enabling the model to maintain the physical interpretability of LAI estimation while reflecting the temporal pattern of continuous changes in maize LAI with phenological processes. Simultaneously, an adaptive gating fusion network dynamically generates weights using accumulated temperature as a condition, automatically adjusting the contribution ratio of physical inversion results and trend prediction results according to different growth stages. This reduces the problem of large estimation bias in specific growth stages by a single model, improving the stability, continuity, and accuracy of maize LAI estimation, and making it suitable for field maize growth monitoring and intelligent farmland management.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, specifically to a neural network model and system for estimating the LAI of corn. Background Technology

[0002] Maize is an important crop for food, feed, and industrial raw materials. Its canopy growth status directly affects light capture, dry matter accumulation, water and fertilizer use efficiency, and ultimately, yield. Leaf Area Index (LAI) is a crucial biophysical parameter characterizing maize canopy structure and population growth, reflecting leaf spread, canopy coverage, and photosynthetic potential. Therefore, rapidly and accurately obtaining maize LAI is of great significance for monitoring maize growth, precision fertilization and irrigation, yield prediction, and intelligent farmland management.

[0003] Traditional methods for obtaining LAI (Local Area Index) mainly include length and width coefficient methods, destructive sampling methods, and canopy analyzer measurement methods. Although they have a certain degree of accuracy under small sample conditions, they generally suffer from problems such as strong reliance on manual labor, low operational efficiency, limited spatial coverage, and difficulty in continuous monitoring, making it difficult to meet the needs of high-frequency, continuous, and automated monitoring of maize growth processes in the field.

[0004] With the development of remote sensing platforms and intelligent algorithms, LAI estimation methods based on satellite remote sensing, UAV multispectral / hyperspectral remote sensing, and machine learning models have gradually become a research focus. Existing technologies include schemes for LAI inversion using multispectral reflectance, vegetation indices, red-edge indices, crop growth cycle information, or radiative transfer models. Chinese patent CN109115725A discloses a joint inversion method for maize canopy LAI and chlorophyll content. This method obtains the multispectral reflectance of maize canopy at various growth stages, establishes a lookup table based on the PROSAIL model, and achieves joint inversion of LAI and chlorophyll content through cost function matching; this scheme emphasizes the physical inversion path based on radiative transfer models and lookup tables. Another example is Chinese patent CN116501925A, which discloses a method for predicting leaf area index (LAI) from multi-source remote sensing data. This method constructs a reflectance lookup table showing the correspondence between growth parameters and canopy reflectance spectra at different growth cycles, and combines this with a red-edge vegetation index lookup table for matching, thus addressing the issues of band differences and multiple solutions in multi-source remote sensing data.

[0005] Existing methods have improved the physical interpretability and multi-source remote sensing adaptability of LAI inversion to some extent, but overall they still mainly rely on lookup table matching, vegetation index screening or single-path inversion, which makes it difficult to fully characterize the temporal trend of continuous changes in maize LAI with the growth process. Summary of the Invention

[0006] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a neural network model for estimating maize LAI (Label Air Quality), comprising: a multi-source feature extraction network, a physical inversion branch, a phenological trend branch, and an adaptive gated fusion network, wherein: the multi-source feature extraction network receives input feature information and outputs extracted deep features, the deep features serving as inputs to the physical inversion branch and the phenological trend branch, respectively outputting inversion prediction values ​​and trend prediction values; the adaptive gated fusion network generates dynamic weights based on accumulated temperature, and performs weighted fusion of the inversion prediction values ​​and trend prediction values ​​to obtain the final maize LAI.

[0007] In one possible implementation, the multi-source feature extraction network includes a spectral-texture feature attention module, a phenological feature attention module, a gating network, and a residual feature extraction network. The input of the gating network is connected to the output of the phenological feature attention module, and the output of the spectral-texture feature attention module is fused with the output of the gating network and then input into the residual feature extraction network.

[0008] In one possible implementation, the input of the gated network is connected to the output of the phenological feature attention module, and the output of the spectral-texture feature attention module is fused with the output of the gated network and then input into the residual feature extraction network, including: The original feature input is processed by the spectral-texture feature attention module. Attention module for phenological features Generate attention score matrices for two features. and : Subsequently, the attention score matrix is ​​applied to the original feature input to obtain weighted feature data: Where: ⊙ represents the Hadamard product, For spectral-texture features, Phenological characteristics; Attention scores are generated via the gated network Gate: The fused feature vector after feature self-attention and multi-source gating mechanism Represented as: Simultaneously extract deep nonlinear mappings. The data is fed into the residual feature extraction network and mapped to a high-dimensional latent space. Where: Sigmoid and ReLU are activation functions, and MLP is a multilayer perceptron; Extracted deep features It will be distributed to the subsequent physical parameter prediction head, and the phenological characteristics input phenological trend parameter head.

[0009] In one possible implementation, the physical parameter prediction head includes a structural parameter head, a strength parameter head, and a phenological correction parameter head. Three description vectors are obtained from the three parameter headers and fed into the physics inversion branch: Inputting phenological characteristics into the phenological trend parameter header yields: , This will be sent to the subsequent phenological trend branch.

[0010] In one possible implementation, the physical inversion branch includes: a structural parameter estimator, an MNCO parameter estimator, and an LAD reconstruction module, wherein the outputs of the structural parameter estimator and the MNCO parameter estimator are both input to the LAD reconstruction module. The structural parameter estimator predicts the parameters that determine the vertical distribution of the canopy: Firstly, based on accumulated temperature... Perform baseline drift correction: Secondly, the learnable canopy dispersion is obtained: Finally, we obtain a physically meaningful learnable vertical skewness: The control of the canopy should be either dense at the top and sparse at the bottom, or dense at the bottom and sparse at the top. This indicates a truncation function that will exceed... Set the value to lower than Set the value to ; The MNCO parameter estimator is designed with a multi-stage nonlinear constraint operator: Among them: basic trends scaling factor Adaptive constraint factor Basic bias These are learnable parameters; The LAD reconstruction module, according to , , and Construct a learnable LAD function: By performing trapezoidal numerical integration on the spatial profile, the physical inversion branch prediction values ​​are obtained: .

[0011] In one possible implementation, the phenological trend branch includes: a phenological parameter estimator and an enhanced Logistic model, wherein the phenological parameter estimator receives the... The maximum leaf area was obtained respectively. growth rate and peak time : in: and To truncate the threshold, as well as To Learnable baseline adjustment parameters; Subsequently, the augmented logistic model is subjected to a learnable log scaling factor. The difference between the cumulative effective accumulated temperature Acc_GDD is normalized and truncated: in: To set a truncation threshold, in order to ensure the sensitivity of the numerical values ​​to the gradient of the Logistic function; For learnable parameters, an exponential term is used to ensure that the scaling factor is non-negative; The final output trend prediction value is obtained using the following formula: in: It is a natural exponential function.

[0012] In one possible implementation, the adaptive gated fusion network includes an adaptive fusion unit, which obtains the output fusion weights by accumulating accumulated temperature. : Combined with the above and The final LAI output is generated by fusion: in: The weighting coefficients are used for fusion.

[0013] In one possible implementation, an adaptive weighting strategy and a multi-constraint uncertainty weighted loss function are used to achieve bi-branch fusion and model optimization, including: Introducing learnable uncertainty parameters and Construct the total loss function: We employ weighted Huber loss and highly LAI-sensitive weights to mitigate spectral saturation and outlier interference. Where: δ is the Huber loss threshold, and γ is the high LAI sensitivity coefficient. = For the dynamic weights of the samples.

[0014] Secondly, embodiments of this application provide a system for estimating the LAI of maize, which is equipped with a neural network model as described in any possible implementation of the first aspect, for estimating the LAI of maize.

[0015] In this embodiment, a multi-source feature extraction network performs deep characterization of input feature information, enhancing the fusion and expressive capabilities of features from different sources and at different scales. Inversion prediction values ​​and trend prediction values ​​are then obtained through physical inversion and phenological trend branches, respectively, ensuring that the model maintains the physical interpretability of LAI estimation while reflecting the temporal pattern of continuous changes in maize LAI with phenological progress. Simultaneously, an adaptive gating fusion network dynamically generates weights using accumulated temperature as a condition, automatically adjusting the contribution ratio of physical inversion results and trend prediction results according to different growth stages. This reduces the problem of large estimation biases by a single model at specific growth stages, improving the stability, continuity, and accuracy of maize LAI estimation, making it suitable for field maize growth monitoring and intelligent farmland management. Attached Figure Description

[0016] Figure 1 A schematic diagram of the framework of a neural network model for corn LAI estimation provided in an embodiment of this application; Figure 2 A schematic diagram of the framework of the multi-source feature extraction network provided in the embodiments of this application; Figure 3 A schematic diagram of the framework for the physical inversion branch provided in the embodiments of this application; Figure 4 A schematic diagram of the framework of phenological trend branches provided in the embodiments of this application; Figure 5A schematic diagram of the spatial distribution of the research area and experimental field provided for embodiments of this application; Figure 6 A comparison chart of the actual and predicted LAI values ​​of each model provided in the embodiments of this application on the training and validation sets in 2025; Figure 7 A comparison chart of the actual and predicted LAI values ​​of each model provided in the embodiments of this application on the 2025 test set A; Figure 8 A comparison chart of the actual and predicted LAI values ​​of each model provided in the embodiments of this application on the 2023 generalization test set B; Figure 9 A schematic diagram of a system framework for corn LAI estimation provided in this application embodiment. Detailed Implementation The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0017] See Figure 1 The neural network model for estimating maize LAI provided in this embodiment includes: a multi-source feature extraction network, a physical inversion branch, a phenological trend branch, and an adaptive gated fusion network. The multi-source feature extraction network receives input feature information and outputs extracted deep features. These deep features serve as inputs to the physical inversion branch and the phenological trend branch, respectively, and output inversion prediction values ​​and trend prediction values. The adaptive gated fusion network generates dynamic weights based on accumulated temperature, and performs weighted fusion of the inversion prediction values ​​and trend prediction values ​​to obtain the final maize LAI.

[0018] See Figure 2 The multi-source feature extraction network in this embodiment includes a spectral-texture feature attention module, a phenological feature attention module, a gating network, and a residual feature extraction network. The input of the gating network is connected to the output of the phenological feature attention module, and the output of the spectral-texture feature attention module is fused with the output of the gating network and then input into the residual feature extraction network.

[0019] In this embodiment, the original feature input is processed by the spectral-texture feature attention module. Attention module for phenological features Generate attention score matrices for two features. and : Subsequently, the attention score matrix is ​​applied to the original feature input to obtain weighted feature data: Where: ⊙ represents the Hadamard product, For spectral-texture features, These are phenological characteristics.

[0020] Introducing a multi-source gating mechanism to modulate spectral-texture features using phenological characteristics. Attention scores are generated via the gated network Gate: The fused feature vector after feature self-attention and multi-source gating mechanism Represented as: Simultaneously extract deep nonlinear mappings. The data is fed into the residual feature extraction network and mapped to a high-dimensional latent space. Where: Sigmoid and ReLU are activation functions, and MLP is a multilayer perceptron that extracts deep features. It will be distributed to the subsequent physical parameter prediction head, and the phenological characteristics input phenological trend parameter head.

[0021] In this embodiment, the physical parameter prediction head includes a structural parameter head, a strength parameter head, and a phenological correction parameter head. Three description vectors are obtained from the three parameter headers and fed into the physics inversion branch: Inputting phenological characteristics into the phenological trend parameter header yields: , This will be sent to the subsequent phenological trend branch.

[0022] See Figure 3 The physical inversion branch includes a structural parameter estimator, an MNCO parameter estimator, and an LAD reconstruction module. The outputs of the structural parameter estimator and the MNCO parameter estimator are both input to the LAD reconstruction module.

[0023] The structural parameter estimator predicts the parameters that determine the vertical distribution of the canopy: Firstly, based on accumulated temperature... Perform baseline drift correction: Secondly, the learnable canopy dispersion is obtained: Finally, we obtain a physically meaningful learnable vertical skewness: The control of the canopy should be either dense at the top and sparse at the bottom, or dense at the bottom and sparse at the top. This indicates a truncation function that will exceed... Set the value to lower than Set the value to ; The MNCO parameter estimator is designed with a multi-stage nonlinear constraint operator: Among them: basic trends scaling factor Adaptive constraint factor Basic bias These are learnable parameters; The LAD reconstruction module, according to , , and Construct a learnable LAD function: By performing trapezoidal numerical integration on the spatial profile, the physical inversion branch prediction values ​​are obtained: .

[0024] See Figure 4 The phenological trend branch includes: a phenological parameter estimator and an enhanced Logistic model, wherein the phenological parameter estimator receives the... The maximum leaf area was obtained respectively. growth rate and peak time : in: and To truncate the threshold, as well as To Learnable baseline adjustment parameters.

[0025] Subsequently, the augmented logistic model is subjected to a learnable log scaling factor. The difference between the cumulative effective accumulated temperature Acc_GDD is normalized and truncated: in: To set a truncation threshold, in order to ensure the sensitivity of the numerical values ​​to the gradient of the Logistic function; For learnable parameters, an exponential term is used to ensure that the scaling factor is non-negative; The final output trend prediction value is obtained using the following formula: in: It is a natural exponential function.

[0026] The adaptive gated fusion network in this embodiment includes an adaptive fusion unit, which obtains the output fusion weights by accumulating accumulated temperature. : Combined with the above and The final LAI output is generated by fusion: in: The weighting coefficients are used for fusion.

[0027] In order to optimize the neural network model, this embodiment adopts an adaptive weighting strategy and a multi-constraint uncertainty weighted loss function to achieve dual-branch fusion and model optimization.

[0028] Specifically, learnable uncertainty parameters are introduced. and Construct the total loss function: We employ weighted Huber loss to suppress outliers and apply dynamic weights to high LAI samples to mitigate the saturation effect. Where: δ is the Huber loss threshold, and γ is the high LAI sensitivity coefficient. = For the dynamic weights of the samples.

[0029] To validate the neural network model for maize LAI estimation provided in the above embodiments, this embodiment uses summer maize as the research object. The experimental site is located in Tai'an City, Shandong Province, China, which has a warm temperate semi-humid continental monsoon climate with an average annual temperature of 12.9℃ and an annual precipitation of 697mm, suitable for maize growth observation. The experiment includes Experimental Area A in 2025 and Experimental Area B in 2023, covering different years, varieties, farming methods, water and fertilizer management, UAV flight altitude, and image resolution, such as... Figure 5 As shown.

[0030] Experimental Area A (2025): Variety Zhengdan 958, sowing density 67,500 plants / hm² -2Multiple tillage treatments, including plowing, deep loosening, rotary tillage, and no-till, were implemented, divided into 10 rows, with 8 quadrats of 3m × 3m per row, for a total of 80 sampling units. Experimental Area B (2023): Variety Denghai 605, sowing density 67,500 plants / hm². -2 Unified cultivation management was implemented, with 3 rows, each containing 10 2.5m×1m quadrats, for a total of 30 sampling units. Data collection covered the seedling stage, jointing stage, large trumpet stage, tasseling stage, silking stage, grain filling stage, and milk stage.

[0031] The DJI M300 RTK drone was used as the flight platform. Test area A was equipped with a RedEdge-P multispectral camera; test area B was equipped with a RedEdge-MX multispectral camera; both cameras acquired data in five bands: blue, green, red, red edge, and near-infrared.

[0032] Test area A had a flight altitude of 15m and a forward / lateral overlap of 80%; test area B had a flight altitude of 50m and a forward / lateral overlap of 75%. Data acquisition took place on clear days from 10:00 to 14:00. Before flight, a radiometric calibration board was photographed, and ground control points were established to improve geometric accuracy. After acquisition, Agisoft Metashape was used for stitching, radiometric correction, and geometric correction to generate orthophotos. Quadratic areas of interest (ROIs) were established using ArcMap, and standard image patches were obtained through batch cropping using Python scripts: 320px × 320px for test area A and 145px × 55px for test area B.

[0033] LAI was measured using an AccuPAR LP-80 canopy analyzer. Each quadrat was measured four times, and the average value was taken as the true value. A total of 730 LAI samples were obtained, including 640 samples in experimental area A and 90 samples in experimental area B.

[0034] The phenological data comes from NASA’s publicly available reanalysis dataset, which includes temperature, precipitation, surface horizontal radiation, and diffuse radiation. Four climatic features were calculated and combined with the number of days after sowing to form a five-dimensional phenological feature.

[0035] The calculation formula is as follows: in: , These are the daily maximum and minimum temperatures, respectively. =10℃, which is the baseline temperature for corn growth; start is the sowing date, and current is the collection date.

[0036] The NDRE is calculated based on multispectral data, and five texture features are extracted, which together with the spectral index constitute a 10-dimensional spectral-texture feature.

[0037] The calculation formula is as follows: Where: RE is the red edge channel, NIR is the near-infrared channel; μ is the image mean, and σ is the standard deviation; Where N is the pixel value and N is the total number of pixels; Let K be the probability of the k-th gray level, and K be the number of histogram intervals.

[0038] For the target acquisition date, 15-dimensional features of the current period (10-dimensional spectral-texture + 5-dimensional phenology) are taken, and 15-dimensional features of the previous acquisition date are taken as lag features. These are then concatenated to obtain 30-dimensional dual-temporal multi-source features, which are then decoupled into dual-temporal spectral-texture features. Biphasic phenological characteristics As the standard input for neural network models.

[0039] In 2025, test area A was divided into training, validation, and test sets A in an 8:1:1 ratio; in 2023, all of test area B was used as the cross-conditional generalization test set B. The evaluation metric used was R0. 2 Accuracy of RMSE assessment: in: Represents the total number of samples. For the first The true LAI value of each sample. Indicates the first LAI prediction values ​​for each sample, This is the mean LAI of the true values ​​of all samples.

[0040] The comparison models include: traditional machine learning models PLSR, SVR, RFR, XGBoost, and NN; and deep learning models MS-1DCNN and MS-VCNN.

[0041] See the results for dataset A Figures 6-7Experimental results show that among traditional machine learning methods, RFR performs best on both the training and validation sets, due to its random sampling mechanism's strong resistance to overfitting when handling high-dimensional features. SVR performs best on the test set, mainly thanks to its structural risk minimization criterion for finding the optimal hyperplane in high dimensions, which gives it good extrapolation capabilities when dealing with nonlinear inversion problems. Comparing deep learning methods, MS-VCNN based on multispectral images performs best on both the training and validation sets, while PROSAIL-DNN has the best accuracy on the test set. This indicates that although visual CNNs have a strong fitting ability to training data, they are prone to overfitting when the sample size is insufficient, contrasting with the PROSAIL-based pre-trained model.

[0042] Among all models, PhysLAINet achieved the best inversion accuracy, with Ri on test set A. 2 The RSI reached 0.8870, with an RMSE as low as 0.6137. Compared to PROSAIL-DNN, the best-performing comparative model on test set A, PhysLAINet's RSI was significantly lower. 2 The performance improved by 3.95%, and the RMSE decreased significantly by 11.52%. More importantly, when LAI > 3, most contrastive models exhibited a "lazy prediction" phenomenon, where predicted values ​​clustered around a certain fixed value, failing to effectively increase with the increase of the true value. This phenomenon stems from the spectral saturation effect under high-density canopy, where the sensitivity of the spectral response to the increase of leaf area decreases drastically. Most contrastive models, lacking additional physical constraints, tend to output the statistical average of high-value regions in the training set. Both MS-VCNN and PhysLAINet significantly alleviated this bottleneck. However, their mitigation mechanisms differ; MS-VCNN relies on its strong visual representation capabilities, while PhysLAINet depends on a physically constrained network structure. Although both have considerable mitigation capabilities for the spectral saturation problem, PhysLAINet's R² on the test set... 2 It outperforms MS-VCNN by 10.57% and reduces RMSE by 23.26%. This demonstrates that the physically constrained structure, while alleviating spectral saturation, has stronger predictive robustness.

[0043] See the results on test set B. Figure 8 The models exhibited extremely significant performance divergence, with all comparison models showing varying degrees of generalization collapse in cross-year tests, and most models showing low R-values. 2 All values ​​decreased to negative values, failing to capture the changing trend of LAI on test set B. The optimal machine learning comparison model, XGBoost, has a lower R-value. 2 The R² value is only 0.4034, while the RMSE is as high as 0.6710. In contrast, the best deep learning comparative model, MS-1DCNN, has a lower R² value.2 The value was only 0.5176, while the RMSE was still as high as 0.6034. This indicates that when faced with extreme heterogeneous conditions, black-box models not designed for the biophysical mechanisms of maize LAI will fail to generalize effectively.

[0044] In contrast, PhysLAINet demonstrates significantly better generalization. Even when the comparison models collectively failed, PhysLAINet maintained an R-value of 0.7002. 2 With an RMSE of 0.4756, the R-value of the best comparison model is significantly lower. 2 It is 35.28% higher, and the RMSE is 21.18% lower than the best comparison model. Although its accuracy has decreased compared to 2025 due to the extreme heterogeneity of the dataset, PhysLAINet can still effectively capture the trend of LAI changes, further demonstrating its strong generalization ability under heterogeneous conditions.

[0045] Ablation experiments were then conducted, as shown in Table 1. The experimental results on test set A show that, except for the removal of adaptive branch fusion, the removal of each component resulted in R... 2 The varying degrees of decline indicate that the modules exhibit good coupling and synergy when processing similarly distributed data. This is attributed to the fact that the physical inversion branch mitigates the inversion distortion of LAI under saturation effects through structural constraints, while the phenological trend branch supplements the dynamic information of crop growth over time. Notably, although removing the adaptive branch fusion module resulted in higher performance on test set A than PhysLAINet, it failed to generalize at all on test set B (R0). 2 =-0.4195, RMSE=1.0350), demonstrating that the adaptive fusion branch based on accumulated temperature information can effectively utilize the complementary information of each branch and significantly enhance the generalization ability of the model.

[0046] On test set B, the model exhibits extremely high ablation sensitivity. After removing the critical component, R... 2 A significant degradation occurred, even resulting in negative values. On one hand, the loss of temporal information such as lag features caused the model to lose its ability to perceive the cumulative effects of maize growth. On the other hand, the removal and replacement of physical components caused the model to revert to a black-box regressor lacking biophysical constraints, unable to effectively generalize under unknown conditions based on constrained priors. This demonstrates that the components of PhysLAINet must work closely together to solve the generalization collapse problem that easily occurs in real agricultural scenarios.

[0047] Table 1 Ablation experimental results of the model in this application on test set A and test set B. Corresponding to the neural network model for corn LAI estimation provided in the above embodiments, this application also provides a system for corn LAI estimation.

[0048] join Figure 9 The system 20 for corn LAI estimation includes a processor 201, a memory 202, and a communication unit 203.

[0049] These components communicate via one or more buses 204. Those skilled in the art will understand that the system for corn LAI estimation shown in the figure does not constitute a limitation on the embodiments of this application. It can be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0050] The communication unit 203 is used to establish a communication channel, so that the system used for corn LAI estimation can communicate with other devices, such as corn image acquisition devices.

[0051] The processor 201 is the control center of the system used for corn LAI estimation. It connects various parts of the system through various interfaces and lines. It drives the internally deployed neural network model to estimate corn LAI by running or executing software programs and / or modules stored in memory 202 and calling data stored in memory.

[0052] Memory 202 is used to store the execution instructions of processor 201. Memory 202 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0053] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0054] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A neural network model for estimating the LAI (Label Artery Identification) of maize, characterized in that, include: The system comprises a multi-source feature extraction network, a physical inversion branch, a phenological trend branch, and an adaptive gated fusion network. The multi-source feature extraction network receives input feature information and outputs extracted deep features. These deep features serve as inputs to the physical inversion branch and the phenological trend branch, respectively, outputting inversion prediction values ​​and trend prediction values. The adaptive gated fusion network generates dynamic weights based on accumulated temperature, and weights and fuses the inversion prediction values ​​and trend prediction values ​​to obtain the final maize LAI (Local Area Index).

2. The neural network model for maize LAI estimation according to claim 1, characterized in that, The multi-source feature extraction network includes a spectral-texture feature attention module, a phenological feature attention module, a gating network, and a residual feature extraction network. The input of the gating network is connected to the output of the phenological feature attention module. The output of the spectral-texture feature attention module is fused with the output of the gating network and then input into the residual feature extraction network.

3. The neural network model for maize LAI estimation according to claim 2, characterized in that, The input of the gated network is connected to the output of the phenological feature attention module. The output of the spectral-texture feature attention module is fused with the output of the gated network and then input into the residual feature extraction network, including: The original feature input is processed by the spectral-texture feature attention module. Attention module for phenological features Generate attention score matrices for two features. and : Subsequently, the attention score matrix is ​​applied to the original feature input to obtain weighted feature data: Where: ⊙ represents the Hadamard product, For spectral-texture features, Phenological characteristics; Attention scores are generated via the gated network Gate: The fused feature vector after feature self-attention and multi-source gating mechanism Represented as: Simultaneously extract deep nonlinear mappings. The data is fed into the residual feature extraction network and mapped to a high-dimensional latent space. Where: Sigmoid and ReLU are activation functions, and MLP is a multilayer perceptron; Extracted deep features It will be distributed to the subsequent physical parameter prediction head, and the phenological characteristics input phenological trend parameter head.

4. The neural network model for maize LAI estimation according to claim 3, characterized in that, The physical parameter prediction head includes structural parameter head, strength parameter head, and phenological correction parameter head. Three description vectors are obtained from the three parameter headers and fed into the physics inversion branch: Inputting phenological characteristics into the phenological trend parameter header yields: , This will be sent to the subsequent phenological trend branch.

5. The neural network model for maize LAI estimation according to claim 4, characterized in that, The physical inversion branch includes: a structural parameter estimator, an MNCO parameter estimator, and an LAD reconstruction module. The outputs of the structural parameter estimator and the MNCO parameter estimator are both input to the LAD reconstruction module. The structural parameter estimator predicts the parameters that determine the vertical distribution of the canopy: Firstly, based on accumulated temperature... Perform baseline drift correction: Secondly, the learnable canopy dispersion is obtained: Finally, we obtain a physically meaningful learnable vertical skewness: The control of the canopy should be either dense at the top and sparse at the bottom, or dense at the bottom and sparse at the top. This indicates a truncation function that will exceed... Set the value to lower than Set the value to ; The MNCO parameter estimator is designed with a multi-stage nonlinear constraint operator: Among them: basic trends scaling factor Adaptive constraint factor Basic bias These are learnable parameters; The LAD reconstruction module, according to , , and Construct a learnable LAD function: By performing trapezoidal numerical integration on the spatial profile, the physical inversion branch prediction values ​​are obtained: .

6. The neural network model for maize LAI estimation according to claim 5, characterized in that, The phenological trend branch includes: a phenological parameter estimator and an enhanced Logistic model, wherein the phenological parameter estimator receives the... The maximum leaf area was obtained respectively. growth rate and peak time : in: and To truncate the threshold, as well as To Learnable baseline adjustment parameters; Subsequently, the augmented logistic model is subjected to a learnable log scaling factor. The difference between the cumulative effective accumulated temperature Acc_GDD is normalized and truncated: in: To set a truncation threshold, in order to ensure the sensitivity of the numerical values ​​to the gradient of the Logistic function; For learnable parameters, an exponential term is used to ensure that the scaling factor is non-negative; The final output trend prediction value is obtained using the following formula: in: It is a natural exponential function.

7. The neural network model for maize LAI estimation according to claim 6, characterized in that, The adaptive gated fusion network includes an adaptive fusion unit, which obtains the output fusion weights by accumulating accumulated temperature. : Combined with the above and The final LAI output is generated by fusion: in: The weighting coefficients are used for fusion.

8. The neural network model for maize LAI estimation according to claim 7, characterized in that, An adaptive weighting strategy and a multi-constraint uncertainty weighted loss function are used to achieve bi-branch fusion and model optimization, including: Introducing learnable uncertainty parameters and Construct the total loss function: We employ weighted Huber loss and highly LAI-sensitive weights to mitigate spectral saturation and outlier interference. Where: δ is the Huber loss threshold, and γ is the high LAI sensitivity coefficient. = For the dynamic weights of the samples.

9. A system for estimating the LAI (Labor Area Index) of maize, characterized in that, The neural network model described in any one of claims 1-8 is deployed for estimating the LAI of corn.

Images