A method for the joint inversion of leaf area index and chlorophyll concentration based on multi-task learning

Abstract

This invention discloses a collaborative inversion method for leaf area index (LAI) and chlorophyll concentration (LCC) based on multi-task learning (PLE). The method includes: selecting several random sample points and uniformly distributing them within a target area; measuring the LAI and LCC values ​​of each sample point and recording their WGS84 coordinates; acquiring high-resolution remote sensing images of the target area based on the measured latitude and longitude information of the sample points; constructing multiple vegetation indices (VIs) based on the preprocessed remote sensing images and using mutual information to select sensitive VIs; dividing the measured dataset into training, validation, and test sets; optimizing the hyperparameters of the progressive hierarchical extraction network model using a tree estimator to establish the optimal PLE collaborative inversion model; evaluating the inversion accuracy of the PLE model using the test set and comparing it with the inversion results of a conventional single-task model. This invention effectively solves the problems of low efficiency and relatively limited accuracy of traditional single inversion methods, exhibiting good accuracy and universality.

Description

Technical Field

[0001] This invention relates to the field of vegetation biophysical parameter inversion technology, and in particular to a method for the coordinated inversion of leaf area index and chlorophyll concentration based on multi-task learning. Background Technology

[0002] LAI (Leaf Area Intake) is defined as half of the total green leaf area per unit land surface area. It is one of the important vegetation structure parameters in terrestrial ecosystem models for simulating gross primary productivity (GPP) and evapotranspiration (ET). LAI reflects the degree of leaf cover on the ground, determines the effective area available for photosynthesis, and directly relates to the efficiency of photosynthesis. LCC (Leaf Cover Capacity) is the total content of chlorophyll a and chlorophyll b per unit leaf area. It is the core pigment in photosynthesis that captures light energy and converts it into chemical energy, and its content directly affects the photosynthetic rate. Traditional methods for measuring LAI and LCC, whether destructive sampling or indirect measurement based on specialized instruments (such as LAI-2200C and SPAD-502), generally suffer from low efficiency, susceptibility to environmental interference, and difficulty in large-scale application. The development of remote sensing technology provides an effective solution for rapidly and non-destructively acquiring vegetation parameters.

[0003] Currently, the inversion of LAI and LCC is mainly based on physical or empirical models, establishing quantitative relationships between feature factors and measured data. These models typically employ a multi-input, single-output architecture, using multiple vegetation indices (VIs) as input variables and LAI or LCC as the output variable. The feature factors used in LAI and LCC inversion primarily originate from spectral information. However, different wavelengths exhibit significant differences in their sensitivity to LAI and LCC: the near-infrared (NIR) band, due to the multiple scattering effect within the canopy, has a more direct correlation with LAI, while the red-edge (RE) band is more sensitive to changes in LCC. Therefore, NIR VIs are typically introduced as key features in LAI inversion, while LCC inversion relies more heavily on RE VIs. It is noteworthy that RE VIs and NIR VIs usually require the integration of visible light band information (such as red light) in their mathematical construction. Furthermore, research reveals that increases in both LAI and LCC lead to a significant decrease in canopy reflectivity in the visible light band. This physical phenomenon indicates that the spectral contribution from another parameter (LCC or LAI) is not negligible in vegetation indices constructed for a single parameter (LAI or LCC). Therefore, this coupling effect may be a key bottleneck limiting further improvements in inversion accuracy. To overcome this limitation, considering the mutual influence in LAI or LCC inversion and achieving synergistic inversion of LAI and LCC is an effective way to improve inversion accuracy in the future.

[0004] In recent years, machine learning models with stronger nonlinear modeling capabilities (such as Random Forest (RF) and Support Vector Regression (SVR)) have been widely used in LAI and LCC inversion and have been applied to collaborative inversion of multi-output regression. For example, applying machine learning to the collaborative inversion of rice LAI and LCC has a physiological basis (both are affected by common factors such as nitrogen and growth period). However, the differences in sensitive features between LAI and LCC cause multi-output machine learning models to easily confuse their respective specific spectral information in the shared feature space while utilizing inter-task correlations, making it difficult to effectively extract their respective specific spectral features, thus limiting further improvement in collaborative inversion accuracy. To solve the above problems, the Progressive Layered Extraction (PLE) model, which has gradually emerged in recent years, provides a more flexible architecture for task feature separation and sharing. PLE can collaboratively optimize the learning process of common and specific features through hierarchical extraction, task experts, and gated routing mechanisms. This architecture can fully explore the potential common information between tasks while effectively preserving, separating, and strengthening the specific feature representation of each task, thereby significantly reducing the interference caused by the spectral coupling effect between parameters and providing new possibilities for improving the accuracy of two-parameter collaborative inversion. Summary of the Invention

[0005] To overcome the limitations in inversion accuracy caused by neglecting each other's spectral contributions in traditional LAI and LCC inversion methods, and the inefficiency resulting from constructing separate inversion models, this invention, based on the concept of multi-task learning, fully considers the inherent correlation between LAI and LCC. By introducing a PLE multi-task learning model for collaborative inversion of LAI and LCC, accurate inversion of both LAI and LCC is achieved. This method not only simplifies the modeling process but also breaks through the accuracy bottleneck of single-task inversion, improving inversion accuracy and providing a novel approach for LAI and LCC monitoring.

[0006] Therefore, this invention provides a method for the coordinated inversion of leaf area index and chlorophyll concentration based on multi-task learning, in order to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] Step 1: Distribute sample points evenly within the target area. Use the LAI2200C canopy analyzer and SPAD502PLUS chlorophyll meter to measure the LAI and LCC values ​​of the sample points respectively, and record the WGS84 coordinate latitude and longitude information of the sample points.

[0009] Step 2: Obtain high-resolution remote sensing images of the target area based on the measured latitude and longitude information of the sample points. The images shall contain at least red light, green light, blue light, at least one near-infrared band and at least one red edge band, and the images shall be preprocessed such as radiometric calibration and atmospheric correction.

[0010] Step 3: Construct multiple vegetation indices (VIs) based on the preprocessed remote sensing images, and use the mutual information (MI) method to screen out the sensitive VIs for LAI and LCC;

[0011] Step 4: Divide the actual dataset into training set, validation set and test set. In the process of partitioning, a hierarchical random partitioning method based on correlation constraints is adopted to ensure that the difference between the Pearson correlation coefficient of LAI and LCC in each subset and the complete dataset does not exceed ±0.02.

[0012] Step 5: Use the tree-based Parzen estimator (TPE) to optimize the hyperparameters of the progressive hierarchical extraction network (PLE) model and establish the optimal PLE collaborative inversion model;

[0013] Step 6: Evaluate the inversion accuracy of the PLE model using the test set and compare it with the inversion results of the conventional single-task model.

[0014] As a preferred embodiment of the present invention, the measured points in step one should not be too dense, but should be distributed as evenly as possible. The measured data mainly includes: WGS84 coordinate information of the points, LAI and LCC.

[0015] The measurement specifications are as follows:

[0016] LAI measurements were conducted strictly following instrument operating procedures, with the drone aerial survey performed at dusk to avoid interference from strong midday radiation and rapidly changing cloud cover. A 270° field-of-view limiting cap was used during the measurement process to effectively eliminate the influence of operator input on the sensor. The measurement plan for each sample point included: three reference light intensity (A-value) measurements were taken in an adjacent open area (generally, an unobstructed sky above the sample point is sufficient. If there is an obstruction, an area within a 5m radius of the sample point can be selected; if no suitable area is found within the 5m buffer zone, another sample point is selected). The average value was then taken. Subsequently, canopy transmitted light intensity (B-value) was measured once each in the east, west, south, and north directions within the sample point, for a total of four measurements. All measurement data were processed using FV2200 software to obtain the LAI value for each sample point.

[0017] LCC was determined using a SPAD-502 Plus chlorophyll meter. This instrument, based on the principle of dual-wavelength optical concentration difference at 650 nm (chlorophyll absorption peak) and 940 nm (reference wavelength), enables non-destructive and rapid determination of the relative chlorophyll content in leaves. To reduce damage to rice plants and improve measurement efficiency, this study abandoned the traditional solvent extraction method and adopted an in-situ non-destructive measurement scheme. Specifically, four representative plants (requiring samples to truly represent the population; operationally, healthy, non-deformed plants with growth at the population average level were selected, and they were evenly distributed across different locations within the sample point to ensure that the measured values ​​are reliable estimates for that sample point) were selected at each sampling point. Leaves from three leaf age groups (young, middle-aged, and old) were collected, and three SPAD readings were obtained from each leaf, for a total of 12 measurements. The arithmetic mean of these 12 measurements was taken as the final LCC value for that sample point. This method effectively improved the representativeness of the samples while ensuring the accuracy and reliability of the measurement data.

[0018] As a preferred embodiment of the present invention, the high-resolution remote sensing image data required in step two shall include at least red light, green light, blue light, at least one near-infrared band, and at least one red-edge band.

[0019] In the LAI and LCC inversion process, the characteristic factors used are mainly derived from spectral information. However, different bands exhibit significant differences in their sensitivity to LAI and LCC: the near-infrared (NIR) band, due to the influence of multiple scattering effects within the canopy, has a more direct correlation with LAI, while the red-edge (RE) band is more sensitive to changes in LCC. Therefore, the high-resolution remote sensing images used in this invention should include as many near-infrared and red-edge bands as possible to provide a data foundation for LAI and LCC inversion.

[0020] In a preferred embodiment of this invention, the VIs in step three should include red-edge VIs and near-infrared VIs. To improve the effectiveness and robustness of VIs in LAI and LCC modeling, this invention employs the Mutual Information (MI) method to evaluate the correlation between multiple VIs and LAI and LCC. Mutual information is a non-parametric method that measures the amount of information shared between two random variables. It can identify linear and non-linear relationships and is suitable for the quantitative analysis of variable correlations in complex ecological environments. To avoid the instability of variable importance caused by the randomness of a single calculation, this invention designs a multiple randomized repeated evaluation mechanism. Specifically, the mutual information calculation between each VI and LAI is repeated 100 times, with a different random seed used for initialization each time to enhance the stability and representativeness of the evaluation results. Finally, the mean mutual information value (Mean MI) and standard deviation (Standard Deviation) of each vegetation index in the 100 evaluations are statistically analyzed to reflect its stable correlation with LAI and LCC and its degree of fluctuation.

[0021]

[0022]

[0023] in Indicates the number of experiments. Indicates the target variable as LAI or LCC. Represents the i-th vegetation index. This represents the average mutual information value of the i-th VI with respect to the target variable. Let represent the k-th mutual information experiment of the i-th VI with respect to the target variable. Let represent the standard deviation of the i-th VI with respect to the target variable. The mean of the obtained mutual information is used to rank the vegetation indices by importance, while the standard deviation is used to evaluate the consistency of the results of each VI across multiple samplings.

[0024] Based on the mutual information results, the first four LAI sensitive features and LCC sensitive features were selected as model inputs.

[0025] As a preferred embodiment of the present invention, the specific reasons and methods for ensuring that the correlation between LAI and LCC remains unchanged during the data partitioning process in step four are as follows:

[0026] The effectiveness of the PLE model highly depends on the true correlation structure between learning tasks. However, in model construction, traditional random data partitioning strategies may unintentionally disrupt the inherent joint distribution characteristics of LAI and LCC in the entire test area sample, causing the statistical associations (such as Pearson correlation) between them in the training, validation, and test sets to deviate from the actual situation. This inconsistency makes it difficult for the PLE model to accurately capture the true interdependence and independent contributions of the two during the learning, tuning, and evaluation stages, thus severely limiting its ability to decouple spectral coupling effects and ultimately restricting the potential for improving inversion accuracy. To ensure that the correlation structure between LAI and LCC is preserved during data partitioning, this invention adopts a data partitioning method based on correlation constraints. First, the Pearson correlation coefficient between LAI and LCC in the complete dataset is calculated as the partitioning benchmark. Second, the two variables are standardized and jointly used as feature inputs, and the samples are divided into several categories using the agglomerative clustering method. Subsequently, based on the clustering results, hierarchical random partitioning is performed to generate training, validation, and test sets in a 6:2:2 ratio. Finally, in each partition, the correlation coefficient between LAI and LCC within each subset is calculated. A partition is considered valid only if the correlation difference between all subsets and the original dataset does not exceed ±0.02. This process is repeated up to 1000 times to obtain a partition scheme that meets the conditions.

[0027] As a preferred embodiment of the present invention, the PLE model architecture in step five is implemented as follows:

[0028] The PLE model employs a hierarchical feature extraction scheme, dividing the model into multiple feature extraction layers. Each layer contains two types of experts and a gated network: Shared Experts: extract common features across tasks; Task Experts: experts specific to each task, used to learn task-specific features; Gate network: dynamically fuses the outputs of Shared Experts and Task Experts at this layer for each task. The previous layer distinguishes between shared and private features through feature extraction, and its gated output serves as the input for the experts in the next layer. This model uses a progressive separation method to gradually distinguish between shared and private features across multiple layers, focusing on the sharing of features across different tasks while emphasizing the inherent differences between tasks. This allows for better explanation of the relationships between multiple tasks and enhances the model's expressive power. The core formula is as follows:

[0029]

[0030]

[0031]

[0032] in Indicates the model input, Indicates PLE number Tasks in the extraction layer The output of the gated network, Indicates the first Tasks in the extraction layer The weighting function, As the first The input to the extraction layer represents the first layer. The extraction layer is the gated output of the previous layer. Indicates the first Extract all available tasks in the layer A vector composed of the outputs of the selected shared expert and the task expert.

[0033] Final Mission The final output is:

[0034]

[0035] in It's the last floor. The gated output is the final representation after all the features extracted, separated and fused from the previous layers. Indicates task Tower networks, commonly in the form of multilayer perceptrons, fuse features Convert to task Required final prediction output .

[0036] In multi-task learning, the joint loss is often represented as a weighted sum of the losses from each task:

[0037]

[0038] in For joint losses, For task weight parameters, For example, cross-entropy loss is used for classification tasks, and mean squared error is used for regression tasks. For the number of tasks, The task number is used. In this study, the LAI and LCC tasks are of equal importance, so their weights are both set to 0.5.

[0039] This invention extends and optimizes the early stopping mechanism, proposing a multi-task early stopping mechanism. It innovatively employs a triple monitoring system (independent sub-task loss and joint loss) to achieve precise convergence determination by synchronously tracking each optimization path. This mechanism terminates training only when all task metrics and total loss have not improved within a preset tolerance period, and saves the model weights at the optimal joint loss. This effectively solves the premature convergence problem caused by asynchronous optimization in multi-task learning, significantly improving the model's generalization ability in complex tasks. "Improvement" refers to the current loss value being at least one minimum tolerance change (min_delta) lower than the historical best loss value, which is set to 0.001 here.

[0040] The tree-based Parzen estimator is used to optimize the structural parameters of the PLE model.

[0041] As a preferred embodiment of the present invention, the conventional LAI and LCC inversion models in step six are: Random Forest Regression (RFR), Support Vector Machine Regression (SVR), Gradient Boosting Regression (GBRT), etc. The method for evaluating the inversion accuracy is as follows:

[0042] Using root mean square error (RMSE) and coefficient of determination (R²) 2 The formula for using accuracy evaluation indicators is as follows:

[0043] Root Mean Square Error (RMSE):

[0044]

[0045] In the formula, n is the total number of samples. These are the model's predicted values. The true value is given; the ranges for root mean square error and mean absolute percentage error are both [missing information]. The value equals 0 when the predicted value perfectly matches the actual value, which is a perfect prediction.

[0046] Coefficient of determination (R) 2 The calculation formula is as follows:

[0047]

[0048] in:

[0049] It is the sum of squared residuals, representing the total difference between the model's predicted values ​​and the actual observed values.

[0050] It is the total sum of squares, representing the sum of the differences between the actual observed values ​​and the observed mean.

[0051] R 2This represents the proportion of data variance that the model can explain. When RV... 2 When R approaches 1, it indicates that the model fits the data well; when R... 2 When the value is close to 0, it indicates that the model cannot fit the data well.

[0052] The present invention has the following advantages:

[0053] This invention provides a method for the collaborative inversion of leaf area index (LAI) and chlorophyll concentration (LCC) based on multi-task learning. It fully considers the correlation between LAI and LCC inversion tasks and their coupling effect on spectral reflectance. It introduces the PLE multi-task learning framework combined with Bayesian optimization algorithm to adaptively construct the optimal collaborative inversion model for regional LAI and LCC. This effectively solves the problems of low efficiency and relatively limited accuracy caused by the need for separate model construction in traditional individual inversion methods. It has good accuracy and universality. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the inversion method described in this invention;

[0056] Figure 2 This is a schematic diagram of the sample point distribution used in this invention;

[0057] Figure 3 This is a diagram showing the screening results based on mutual information features in this invention;

[0058] Figure 4 This is a diagram showing the data partitioning results based on correlation constraints in this invention;

[0059] Figure 5 This is a schematic diagram of the PLE model structure determined in this invention;

[0060] Figure 6 Box plots of LAI derived from different models of this invention;

[0061] Figure 7 The LCC box plots are generated by different models of this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0063] This invention provides a method for the coordinated inversion of leaf area index and chlorophyll concentration based on multi-task learning. For detailed procedures, please refer to [link / reference]. Figure 1 The specific steps are as follows:

[0064] Step 1: Distribute sample points evenly within the target area. Use the LAI2200C canopy analyzer and SPAD502PLUS chlorophyll meter to measure the LAI and LCC values ​​of the sample points respectively, and record the WGS84 coordinate latitude and longitude information of the sample points.

[0065] This implementation method uses high-standard farmland in Yucai Ecological Zone, Sanya City, Hainan Province, China as the target area and rice as the research object (growing stage, with LAI and LCC in a rapid growth phase). It details the implementation process of the LAI and LCC collaborative inversion method based on multi-task learning, verifying the feasibility and superiority of the method. The target area is located at 18°25′–18°30′ N, 109°20′–109°25′ E, belonging to a tropical monsoon climate. The farmland is contiguous, the soil type is lateritic red soil, and the rice planting density is uniform (approximately 25 plants / m²). 2 It has no obvious pests, diseases, or stress, and is suitable as a typical area for the inversion of LAI and LCC.

[0066] A uniform grid layout method was used to establish 70 sampling points within the target farmland, with a spacing of 50m between points (to avoid spatial autocorrelation), covering both the edge and central areas of the farmland to ensure sample representativeness. The latitude and longitude of each sampling point in the WGS84 coordinate system were recorded using a handheld GPS device. A total of 70 sets of measured data for LAI and LCC were ultimately selected. The data distribution is shown in Table 1, and the sampling point distribution is as follows: Figure 2 As shown.

[0067] Table 1. Statistical Table of Measured Data for LAI and LCC

[0068]

[0069] The LAI ranges from 2.57 to 6.21, and the LCC ranges from 42.5 to 50.8, covering the typical parameter ranges during the booting stage of rice, making them suitable as training labels for the model.

[0070] Step 2: Obtain high-resolution remote sensing images of the target area based on the measured latitude and longitude information of the sample points. The images should contain at least red light, green light, blue light, at least one near-infrared band and at least one red edge band. Perform preprocessing such as radiometric calibration and atmospheric correction on the images.

[0071] This invention utilizes a DJI Phantom 4 multispectral drone to acquire high-resolution remote sensing images of the target area. The acquisition was conducted on the same day as the ground-based experiment. To ensure image quality, the drone was used on a clear day with no wind or light wind (wind speed <5m / s) to avoid cloud cover and strong winds that could cause image blurring. The preferred time period was 10:00 AM to 2:00 PM (higher solar altitude angle, less shadow interference). The drone's forward overlap was 80%, and the lateral overlap was 70%. The flight altitude was 120m, corresponding to a spatial resolution of 0.067m. An automatically planned flight path was used. Multispectral image data stitching, geometric correction, radiometric calibration, and band registration were performed using DJI Terra software.

[0072] Step 3: Construct multiple vegetation indices (VIs) based on the preprocessed remote sensing images, and use the mutual information (MI) method to screen out the sensitive VIs for LAI and LCC;

[0073] Considering the specific responses of plants to different electromagnetic spectra, this invention performs specific nonlinear combination operations on different bands to enhance the signal characteristics of sensitive bands and suppress the influence and interference of non-sensitive bands. Finally, 14 vegetation indices such as NDVI, EVI, and NDRE were constructed, as shown in Table 2. In Table 2, B, G, R, RE, and NIR represent the blue, green, red, red-edge, and near-infrared bands of remote sensing images, respectively.

[0074] Table 2. Details of Vegetation Index (VIs) Construction

[0075]

[0076] The mutual information calculation between each VI and LAI was repeated 100 times, with a different random seed used for initialization each time to enhance the stability and representativeness of the evaluation results. Finally, the Mean MI and Standard Deviation of each vegetation index were statistically analyzed across the 100 evaluations to reflect its stable correlation and fluctuation with LAI and LCC. To present the evaluation results more intuitively, this invention generated a bar chart of the average mutual information and a distribution map for each evaluation, which can be further referenced. Figure 3 ,in Figure 3 (a) and Figure 3 Figure (b) shows the feature selection results for LAI and LCC, respectively, displaying the average importance ranking and stability analysis of each VI. Here, the top 4 VIs are selected as sensitive features by sorting them from largest to smallest Mean MI.

[0077] Step 4: Divide the actual dataset into training set, validation set and test set. In the process of partitioning, a hierarchical random partitioning method based on correlation constraints is adopted to ensure that the difference between the Pearson correlation coefficient of LAI and LCC in each subset and the complete dataset does not exceed ±0.02.

[0078] A raw scatter plot of rice samples was drawn with LAI on the horizontal axis and LCC on the vertical axis to visually represent the distribution characteristics of the data at different time periods. Based on the spatial distribution pattern of the samples in the scatter plot, the appropriate number of clusters for each time period was determined. To improve the stability and robustness of clustering, the LAI and LCC data were first standardized to eliminate dimensional differences. Then, hierarchical clustering was used to perform cluster analysis on the samples. Based on this, stratified random sampling was implemented according to the clustering results, dividing the data into training, validation, and test sets in a 6:2:2 ratio. The partitioning results are shown below. Figure 4 As shown in Table 3, the Pearson correlation coefficients between LAI and LCC in each subset before and after the partitioning are listed, indicating that the correlation between variables in each subset is well preserved. Figure 4 As shown in Table 3, the data partitioning method based on correlation constraints effectively ensures the uniform distribution of LAI and LCC within each subset space. This method better reflects the overall distribution characteristics of LAI and LCC data within the sample plot and maintains the correlation structure between them, thereby ensuring the applicability and generalization ability of the model in this sample plot area.

[0079] Table 3. Comparison of Pearson correlation coefficients between LAI and LCC before and after dataset splitting.

[0080] Original dataset training set Validation set test set 0.651 0.644 0.657 0.660

[0081] Step 5: Use the tree-based Parzen estimator (TPE) to optimize the hyperparameters of the progressive hierarchical extraction network (PLE) model and establish the optimal PLE collaborative inversion model;

[0082] The structure parameters of the PLE model were optimized using a tree-based Parzen estimator. The parameter ranges were set as follows: num_layers: 1~4; num_shared_experts: 1~10; num_lai_experts: 1~10; num_lcc_experts: 1~10; expert_dim: [32, 64, 128, 256, 512]; dropout_rate: 0~1; hidden_units: [[128, 64], [64, 32], [128, 64, 32], [256, 128, 64], [256, 128]]. The final optimized hyperparameters are shown in Table 4, and the network model diagram is shown below. Figure 5 .

[0083] Table 4. Results of Hyperparameter Optimization for PLE Model

[0084]

[0085] Step 6: Evaluate the inversion accuracy of the PLE model using the test set and compare it with the inversion results of the conventional single-task model.

[0086] To effectively control the impact of data partitioning randomness on model performance evaluation, the dataset was randomly partitioned 50 times according to the aforementioned data partitioning method based on correlation constraints. The model was built on the partitioned training set each time and evaluated on the corresponding test set. The final result was taken as the average statistic of the 50 repeated experiments as a robust estimate of model performance. LAI or LCC sensitive features were input into the RFR, SVR, and GBRT models respectively to individually invert LAI (LCC). LAI and LCC sensitive features were simultaneously input into the PLE model to collaboratively invert LAI and LCC. The inversion results for different models are detailed in Table 5. Figure 6 and Figure 7 ,in Figure 6 (a) Figure 6 (b) represents the R values ​​of LAI inversion using different models. 2 Box plot with RMSE Figure 7 (a) Figure 7 (b) represents the R values ​​for LCC inversion using different models. 2 Box plot with RMSE.

[0087] Table 5. Comparison of Inversion Accuracy between LAI and LCC Models

[0088]

[0089] Therefore, the PLE model has the best inversion accuracy. This implementation method has been verified in paddy fields in Sanya, demonstrating that the LAI and LCC collaborative inversion method based on multi-task learning has the following advantages:

[0090] High accuracy: PLE model inverts the R-value of LAI 2 The RMSE of LCC is reduced to 1.43, reaching 0.80, which is better than the traditional single-task model;

[0091] Standardized processes: From sample point measurement and image preprocessing to model optimization, each step is reproducible and adaptable to different vegetation and regions;

[0092] Superior efficiency: Achieves dual-parameter inversion in a single training session, improving efficiency by 50% compared to separate modeling.

[0093] This method can be effectively applied to scenarios such as agricultural growth monitoring and ecosystem assessment, providing a reliable technical solution for large-scale vegetation parameter inversion.

[0094] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for collaborative inversion of leaf area index and chlorophyll concentration based on multi-task learning, characterized in that, Comprising the following steps: Step one: uniformly arrange sample points in the target area, measure the leaf area index value LAI of the sample points using the LAI2200C canopy analyzer, measure the leaf chlorophyll concentration value LCC of the sample points using the SPAD502PLUS chlorophyll meter, and record the WGS84 coordinate system longitude and latitude information of each sample point simultaneously; Step two: according to the sample point longitude and latitude information recorded in step one, obtain a high-resolution remote sensing image of the target area, the high-resolution remote sensing image needs to at least contain a red light band, a green light band, a blue light band, one and more near-infrared bands and one and more red edge bands; the high-resolution remote sensing image is sequentially subjected to radiation calibration, atmospheric correction, geometric correction and band registration pretreatment; Step three: based on the pretreated high-resolution remote sensing image, a plurality of vegetation indices VIs including near-infrared vegetation indices and red edge vegetation indices are constructed, and the mutual information method is used to screen the sensitive VIs of LAI and the sensitive VIs of LCC as the subsequent model input features; Step four: divide the LAI and LCC measured data set obtained in step one into a training set, a validation set and a test set, and the division process adopts a hierarchical random division method based on correlation constraints to ensure that the Pearson correlation coefficients of LAI and LCC in the training set, the validation set and the test set differ by no more than ±0.02 from the Pearson correlation coefficient of LAI-LCC of the complete measured data set; Step five: use the tree Parzen estimator TPE to optimize the hyperparameters of the progressive layer extraction network PLE model, the hyperparameters include the number of feature extraction layers of the PLE model, the number of shared experts in each layer of the feature extraction layer, the number of LAI task-specific experts in each layer of the feature extraction layer, the number of LCC task-specific experts in each layer of the feature extraction layer, the output feature dimension of the shared experts and the task-specific experts, the neuron dropout rate in the model training process for preventing overfitting, and the number of hidden layer units in the task tower network in the final output layer of the PLE model, representing the number of units in each hidden layer in the form of a list; after determining the optimal hyperparameter combination, a PLE collaborative inversion model based on the optimal hyperparameters is constructed, the PLE collaborative inversion model includes a plurality of feature extraction layers and one final output layer: The multi-layer feature extraction layer comprises a shared expert, a task-specific expert and a gating network; the shared expert is used to extract cross-task common spectral features of the LAI inversion task and the LCC inversion task; the task-specific expert comprises an LAI task-specific expert and an LCC task-specific expert, and is used to learn private spectral features of the LAI inversion task and private spectral features of the LCC inversion task, respectively; the gating network is provided with an independent gating network for the LAI inversion task and the LCC inversion task, respectively, and each gating network dynamically adjusts the output weight of the shared expert and the corresponding task-specific expert according to the input features to realize feature fusion; the fused features are used as the input of the next layer of the feature extraction layer; The final output layer adopts a task tower network, and the task tower network is a multi-layer perceptron structure; the task tower network receives the fused features of the gating network of the last layer of the feature extraction layer and converts the fused features into an LAI prediction value and an LCC prediction value; A joint loss function is used in the model training process: wherein and are the mean square error loss for the LAI and LCC tasks, respectively; 0.5 is the loss weight for the LAI retrieval task and the LCC retrieval task, used to balance the optimization priority of the two tasks. A multi-task early stopping mechanism is used in the model training process: At each iteration cycle of the model training, the independent loss of the LAI inversion task is calculated and monitored synchronously , the independent loss of the LCC inversion task , and the joint loss of the model ; A preset loss improvement tolerance period, i.e., a threshold of the number of continuous iterations, is set; if 、 、 If the continuous preset tolerance periods do not appear to be improved, the model training is terminated. During the training process, the model weights of each iteration cycle are saved in real time The minimum model weights are taken as the final trained PLE model weights, and the PLE collaborative inversion model for synchronous inversion of LAI and LCC is finally obtained. Step six: input the test set divided in step four into the PLE collaborative inversion model constructed in step five, and evaluate the inversion accuracy of the model by using the root mean square error RMSE and the determination coefficient R2, and compare the accuracy with the inversion accuracy of the conventional single-task regression model.

2. The method of claim 1, wherein, In the step one: During LAI measurement, a 270-degree field-of-view limiting cap is used to eliminate the interference of the operator on the sensor on the day of unmanned aerial photography; the LAI measurement of each sample point includes: measuring the reference light intensity 3 times in the adjacent open area of the sample point, and taking the arithmetic mean of the 3 measurement results as the reference light intensity of the sample point; measuring the crown layer transmission light intensity in the east, west, south and north directions of the sample point; calculating the reference light intensity and the crown layer transmission light intensity by using the FV2200 software to obtain the LAI value of the sample point; During LCC measurement, 4 representative plants are selected at each sample point; for each plant, the leaves of the young leaf age group, the middle leaf age group and the old leaf age group are collected; the SPAD 502PLUS chlorophyll meter is used to measure the SPAD readings of each collected leaf 3 times; the arithmetic mean of the 12 SPAD readings of all the leaves of the 4 plants is taken as the LCC value of the sample point.

3. The method of claim 1, wherein, The specific process of screening sensitive VIs by using the mutual information method in the step three is as follows: (1) for each constructed vegetation index VI, the mutual information value of each VI with LAI and LCC is calculated, and different random seed initialization is used each time, and the calculation is repeated 100 times; (2) the arithmetic mean of the 100 mutual information values of each VI with LAI is obtained, and the average mutual information value of the VI with LAI is obtained; similarly, the average mutual information value of each VI with LCC is obtained; (3) Sort all VIs according to the average mutual information value with LAI from large to small, and select the top 4 VIs as the sensitive vegetation index of LAI; sort all VIs according to the average mutual information value with LCC from large to small, and select the top 4 VIs as the sensitive vegetation index of LCC.

4. The method of claim 1, wherein, The hierarchical random partition method based on correlation constraint in the fourth step specifically comprises: (1) Calculate the Pearson correlation coefficient of LAI and LCC in the complete measured data set obtained in step one, and take the coefficient as the correlation reference for data partitioning; (2) Standardize the LAI values and LCC values in the complete measured data set respectively to obtain standardized LAI data and standardized LCC data; (3) Take the standardized LAI data and standardized LCC data as joint features, and use the agglomerative hierarchical clustering algorithm to classify the samples in the complete measured data set to obtain a plurality of sample clustering clusters; (4) Based on the sample clustering clusters obtained in step (3), randomly partition the samples in each clustering cluster according to a ratio of 6:2:2, and respectively into a training set, a validation set and a test set; (5) Calculate the Pearson correlation coefficient of LAI and LCC in the partitioned training set, validation set and test set, and compare it with the correlation reference determined in step (1); if the absolute difference of the LAI-LCC Pearson correlation coefficient in all subsets from the reference is ≤0.02, the partitioning is valid; if the difference of any subset exceeds ±0.02, repeat steps (4)-(5) until the partitioning is valid or the number of repetitions reaches the upper limit of 1000 times.

5. The method of claim 1, wherein, The precision evaluation index in the step six includes: root mean square error RMSE and determination coefficient R 2 , the formula is as follows: Root mean square error RMSE: where n is the total number of samples, is the model predicted value, is the true value; the range of the root mean square error and the mean absolute percentage error are equals 0, i.e. perfect prediction, when the predicted value is exactly the same as the true value. Determination coefficient R 2 The calculation formula is as follows: wherein: is the sum of squared residuals, representing the total sum of differences between the model predicted values and the actual observed values; is the total sum of squares, representing the total sum of differences between actual observations and the observed mean; R 2 represents the proportion of the variance in the data that the model can explain; when R 2 is close to 1, the model fits the data well; when R 2 is close to 0, the model does not fit the data well.

Images