An inversion method, system, device and medium for the growth trend of crops

Abstract

The present invention discloses a method, system, device and medium for inverting the growth condition of crops, relating to the field of remote sensing technology, including: obtaining the red edge inflection point (REIP) and the integrated red edge chlorophyll index (IRECI) according to the reflectance of the spectral image of the crop planting area, and combining the REIP and the IRECI in different proportions; training a first deep neural network (DNN) with the combined index of the first proportion to predict the leaf area index, and optimizing the learning rate and regularization parameters of the first DNN by the particle swarm method during the training process to construct a leaf area index inversion model; training a second DNN with the combined index of the second proportion to predict the plant height data to obtain a plant height inversion model; obtaining the leaf area index and the plant height data according to the leaf area index inversion model and the plant height inversion model and fitting them by the entropy weight method to obtain the index weight result for evaluating the growth condition of crops. The present invention can accurately predict the growth condition of crops.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and particularly to a method, system, device and medium for inverting the growth condition of crops. Background Art

[0002] Remote sensing technology has the characteristics of periodic observation and large-area coverage in obtaining ground information. It plays an important role in agricultural resource monitoring. It can capture the distribution information of key biophysical parameters of terrestrial vegetation in space and time, which is the basis for the study of the growth condition of corn.

[0003] In the prior art, the growth prediction of corn mostly analyzes from a single factor, such as soil factors, meteorological factors or factors of the crop itself, or analyzes from the small-scale direction of data collected by drones.

[0004] However, a single factor is difficult to reflect the overall growth condition of crops, resulting in the inability to accurately predict the growth condition of crops. Summary of the Invention

[0005] Embodiments of the present invention provide a method, system, device and medium for inverting the growth condition of crops, which can solve the problem in the prior art that the growth condition of crops cannot be accurately predicted.

[0006] Embodiments of the present invention provide a method for inverting the growth condition of crops, including the following steps: obtaining the red edge index REIP and the infrared red edge index IRECI according to the reflectance of the spectral image of the crop planting area; combining the red edge index REIP and the infrared red edge index IRECI in different proportions to obtain a combined index of the first proportion and a combined index of the second proportion; training a first deep neural network DNN with the combined index of the first proportion to predict the leaf area index, and optimizing the learning rate and regularization parameters of the first deep neural network DNN by a particle swarm method during the training process to construct a leaf area index inversion model; training a second deep neural network DNN with the combined index of the second proportion to predict the plant height data to obtain a plant height inversion model; combining the red edge index REIP and the infrared red edge index IRECI of the spectral image of the crop planting area to be measured according to the first proportion and the second proportion respectively; substituting the combined index of the first proportion to be measured into the leaf area index inversion model to obtain the leaf area index, and substituting the combined index of the second proportion to be measured into the plant height inversion model to obtain the plant height data; fitting the leaf area index and the plant height data by the entropy weight method to obtain an index weight result for evaluating the growth condition of crops.

[0007] Further, the steps for obtaining the leaf area index inversion model specifically include: obtaining the red-edge index RE IP and the red-edge index I REC I combined according to the first ratio; setting initial hyperparameters in the first deep neural network DNN with 5 linear layers, and using the particle swarm optimization algorithm to optimize the hyperparameters to obtain the learning rate and the L2 regularization strength. Then, training the optimized first deep neural network DNN with the combined index of the first ratio to predict the leaf area index, so as to construct the leaf area index inversion model.

[0008] Further, the steps for obtaining the plant height inversion model specifically include: obtaining the red-edge index RE IP and the red-edge index I REC I combined according to the second ratio; setting hyperparameters in the second deep neural network DNN with 5 linear layers, setting the learning rate to 0.0001, adjusting the regularization strength to 0.2, and training the second deep neural network DNN with the combined index of the second ratio to predict the plant height data, so as to obtain the plant height inversion model.

[0009] Further, the steps for combining the red-edge index RE IP and the red-edge index I REC I according to different ratios specifically include: combining the red-edge index RE IP and the red-edge index I REC I according to the first ratio of 0.3:0.7; combining the red-edge index RE IP and the red-edge index I REC I according to the second ratio of 0.2:0.8.

[0010] Further, the steps for obtaining the red-edge index RE IP and the red-edge index I RECI specifically include: obtaining multiple vegetation indices based on the reflectance included in the spectral image of the crop planting area, and using the feature recursive elimination algorithm to perform feature selection on the multiple vegetation indices, and selecting the red-edge index RE IP and the red-edge index I REC I.

[0011] Further, the steps for obtaining the index weight result for evaluating the growth of crops specifically include: performing standardization processing on the leaf area index and the plant height data to eliminate the influence of different dimensions; obtaining the information entropy of each index for the data after standardization processing, and using the information redundancy degree to obtain the weight result.

[0012] An embodiment of the present invention provides an inversion system for the growth of crops, including:

[0013] A data acquisition module, configured to obtain the red-edge index RE IP and the red-edge index I REC I according to the reflectance of the spectral image of the crop planting area; and combine the red-edge index RE IP and the red-edge index I RECI according to different ratios to obtain the combined index of the first ratio and the combined index of the second ratio;

[0014] A model training module, which is used to train a first deep neural network (DNN) with a first proportion of combined indices to predict the leaf area index, and optimize the learning rate and regularization parameters of the first DNN through a particle swarm method during the training process to construct an inversion model for the leaf area index; train a second DNN with a second proportion of combined indices to predict plant height data, and obtain an inversion model for plant height.

[0015] A growth inversion module, which is used to combine the red-edge index RE IP and the red-edge index I REC I of the spectral image of the crop planting area to be measured according to the first proportion and the second proportion respectively; substitute the combined indices of the first proportion to be measured into the inversion model of the leaf area index to obtain the leaf area index, and substitute the combined indices of the second proportion to be measured into the inversion model of plant height to obtain plant height data; fit the leaf area index and the plant height data through the entropy weight method to obtain the index weight result for evaluating the crop growth.

[0016] An embodiment of the present invention provides a computer device, including: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the inversion method for the growth of a crop as described above is implemented.

[0017] An embodiment of the present invention provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the inversion method for the growth of a crop as described above is implemented.

[0018] An embodiment of the present invention provides an inversion method, system, device and medium for the growth of a crop. Compared with the prior art, the beneficial effects are as follows:

[0019] Obtain the red-edge index RE IP and the red-edge index I RECI according to the reflectance of the spectral image of the crop planting area; combine the red-edge index RE IP and the red-edge index I REC I in different proportions to obtain the combined indices of the first proportion and the combined indices of the second proportion; train a first DNN with the combined indices of the first proportion to predict the leaf area index, and optimize the learning rate and regularization parameters of the first DNN through a particle swarm method during the training process to construct an inversion model for the leaf area index; train a second DNN with the combined indices of the second proportion to predict plant height data, and obtain an inversion model for plant height; combine the red-edge index RE IP and the red-edge index I RECI of the spectral image of the crop planting area to be measured according to the first proportion and the second proportion respectively; substitute the combined indices of the first proportion to be measured into the inversion model of the leaf area index to obtain the leaf area index, and substitute the combined indices of the second proportion to be measured into the inversion model of plant height to obtain plant height data; fit the leaf area index and the plant height data through the entropy weight method to obtain the index weight result for evaluating the crop growth.

[0020] Among them, the red edge index RE_IP and the inverted red edge chlorophyll index I_RECI that reflect the growth trend of crops are combined and input into the leaf area index inversion model in different proportions to achieve accurate prediction of the leaf area index. Then, the entropy weight method is used to fit the predicted plant height data. The leaf area index and the plant height data are used as indicators to evaluate the growth trend of crops, and the weight results of the indicators for evaluating the growth trend of crops can be accurately obtained. Finally, accurate prediction of the growth trend of crops is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart of an inversion method for the growth trend of crops provided by an embodiment of the present invention;

[0022] Figure 2 Flowchart of the main stages of the application of the DNN algorithm provided by an embodiment of the present invention;

[0023] Figure 3 Flowchart of the particle swarm optimization algorithm provided by an embodiment of the present invention;

[0024] Figure 4 Deep neural network DNN optimized by the particle swarm optimization algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0026] See Figures 1 to 4 , an embodiment of the present invention provides an inversion method for the growth trend of crops, including the following steps:

[0027] Step 1: Obtain a plurality of vegetation indices based on the reflectance included in the spectral image of the crop planting area, and use the feature recursive elimination algorithm to perform feature selection on the plurality of vegetation indices, and select the red edge index RE_IP and the inverted red edge chlorophyll index I_RECI. Among them, RE_IP is the red edge position index, and I_RECI is the inverted red edge chlorophyll index. The red edge index RE_IP and the inverted red edge chlorophyll index I_RECI are combined in different proportions. Specifically: The red edge index RE_IP and the inverted red edge chlorophyll index I_RECI are combined in a first ratio of 0.3:0.7. The red edge index RE_IP and the inverted red edge chlorophyll index I_RECI are combined in a second ratio of 0.2:0.8.

[0028] Step 2: Train the first deep neural network DNN with the combined index of the first ratio to predict the leaf area index, and optimize the learning rate and regularization parameters of the first deep neural network DNN by the particle swarm method during the training process to construct a leaf area index inversion model; train the second deep neural network DNN with the combined index of the second ratio to predict the plant height data and obtain a plant height inversion model.

[0029] Step 3: Combine the red edge index RE IP and the red edge index I RECI of the spectral image of the crop planting area to be measured according to the first ratio and the second ratio respectively; substitute the combined index of the first ratio to be measured into the leaf area index inversion model to obtain the leaf area index, and substitute the combined index of the second ratio to be measured into the plant height inversion model to obtain the plant height data. Fit the leaf area index and the plant height data by the entropy weight method to obtain the index weight result for evaluating the growth of the crop.

[0030] Descriptive statistical analysis and visualization display were carried out on the field measurement data set, and 9 vegetation indices were selected to analyze their relationships with LAI and plant height. The NDVI index can quantify the amount of vegetation, with high values indicating dense tree canopies and low or negative values indicating urban and water feature areas. The PVI index is the perpendicular distance from a vegetation pixel to the soil brightness line in the two-dimensional coordinate system of R-N I R. The SAVI index is a hybrid between a ratio-based index and a perpendicular index. The RE IP index was developed for applications in heterogeneous farmland biomass and nitrogen (N) uptake measurement / management, the inflection point of strong red light absorption in the near-infrared reflection, which contains information on the growth status, and the reflectance near the red edge is very sensitive to large-scale changes in crop LAI. The IRECI index was developed for applications in heterogeneous farmland biomass and nitrogen (N) uptake measurement / management, the inflection point of strong red light absorption in the near-infrared reflection, which contains information on the growth status, and the reflectance near the red edge is very sensitive to large-scale changes in crop LAI. The PSSRa index is used to estimate chlorophyll content from MERI S (Medium Resolution Imaging Spectrometer) data, with the goal of estimating RE IP. The MTCI index aims to estimate the red edge position (REP), which is the maximum inflection point in the red and near-infrared regions of the plant spectral reflectance, and it is of great significance for observing chlorophyll content, vegetation senescence, water, and nutrient deficiency stress. The MCARI index responds to changes in chlorophyll, and the index algorithm is responsive to leaf chlorophyll concentration and ground reflectance. The calculation methods of the above indices are shown in Table 1 Index Calculation Formulas.

[0031] Table 1 Index Calculation Formulas

[0032]

[0033]

[0034] To ensure the accuracy of LAI estimation, it is crucial to establish a temporal correspondence between Sentinel-2 satellite reflectance data and field-measured LAI.

[0035] Feature selection (FS) is a technique widely used in pattern recognition applications. By removing irrelevant, noisy, and redundant features from the original feature space, FS alleviates the problem of overfitting, improves the performance of the model, and can also reduce the time and space costs of the learning algorithm. More importantly, we can gain a deeper understanding of the data by analyzing the importance of features. The Recursive Feature Elimination (RFE) algorithm is a wrapper method for feature selection. It gradually optimizes the feature subset by repeatedly building models and eliminating the least important features, thereby improving the prediction performance of the model. The following is a detailed introduction to the principle of the RFE algorithm. The purpose of the RFE algorithm is to eliminate redundancy among features, select the optimal feature combination, and reduce the feature dimension. It aims to find the feature subset with the best performance by repeatedly creating models. Its basic principle can be summarized in the following steps:

[0036] 1. Initial feature subset selection: First, select K features from the original feature set as the initial feature subset. This value of K can be set according to the actual situation or experience, or it can start from 1 and gradually increase to explore the impact of different numbers of features on the model performance. In this experiment, the original features start from 1 and increase one by one until 8. Each case is traversed.

[0037] 2. Model training and evaluation: Train a machine learning model using the initial feature subset and evaluate the performance of the model. The evaluation method usually includes cross-validation to ensure the reliability of the model performance. In this experiment, the cross-validation method is adopted.

[0038] 3. Feature importance evaluation: After the model training is completed, evaluate the importance of each feature. The importance of features can be determined by different methods, such as feature weights, coefficients, information gain, etc., depending on the machine learning algorithm used.

[0039] 4. Feature elimination and subset update: According to the importance ranking of features, eliminate the least important feature (or retain the most important feature, depending on the algorithm implementation), and use the remaining features to construct a new feature subset. Realize the functions of dimensionality reduction and optimization of the feature subset.

[0040] 5. Iterative process: Repeat the above steps, that is, train the model using the new feature subset, evaluate the model performance, evaluate the feature importance, eliminate (or retain) features, until the predetermined number of features or other stopping conditions (such as the model performance no longer improves significantly) are reached.

[0041] 6. Select the optimal feature subset: During all iterations, multiple feature subsets with different numbers of features are generated. Finally, the feature subset with the highest accuracy is selected as the final feature subset. The results of RFE analysis show that the optimal number of features is 2, and the corresponding features are REIP and IRECI.

[0042] Build a plant height inversion model based on DNN:

[0043] A deep neural network (DNN) is a multi-layer neural network that forms a deep structure by connecting multiple neurons together. The DNN developed from ANN has a deeper architecture and can establish a complex mapping function from input data to output data. The basic structure of a feed-forward DNN usually consists of an input layer, multiple hidden layers, and an output layer. There is one or more neurons in each layer, and each neuron is connected to other neurons in the upper and lower layers. Neurons receive the output of neurons in the previous layer, transfer it through an activation function, and then pass it to other neurons in the next layer during the forward propagation process. The activation function of a neuron is usually a non-linear function, which enables the DNN to handle highly non-linear relationships between inputs and outputs. During the training process, the weights and biases of each neuron are adjusted iteratively until the optimal values are reached. It is the backward propagation of errors (losses), which transfers the error between the predicted value and the true value to the upper layers. The weights and biases can be optimized by many algorithms, such as the gradient descent method. DNN can be used to process various types of data, including images, text, and speech, etc. In a deep neural network, the number of hidden layers and the number of neurons in each layer can be determined according to the specific task and data type. The training process of DNN usually adopts the backpropagation algorithm and the gradient descent optimization method, and continuously adjusts the parameters of the neural network to minimize the prediction error and the loss function.

[0044] The deep neural network has multiple non-linear mapping feature transformations and can fit highly complex functions. Compared with the shallow modeling method, the deep modeling can represent actual complex non-linear problems more carefully and efficiently.

[0045]

[0046] where x i represents the i-th input of the neuron (the output of the neuron in the previous layer), w i is the weight between two neurons, b is the bias of the neuron, and f represents the activation function.

[0047] Taking the vegetation index as the input and the plant height as the output, it contains 5 linear layers. Set the hyperparameters, set the learning rate to 0.0001, adjust the regularization strength to 0.2, train with small-area corn crops, obtain the corn plant height model, and verify it in a large area.

[0048] In theory, a neural network can approximate any function with an appropriate architecture. However, as the number of layers increases, the architecture of the neural network becomes more complex, and the parameters of the neural network also become more numerous. This makes the learning process much slower. In a neural network with multiple hidden layers, it is difficult to train using classical algorithms due to the vanishing gradient of errors during the backpropagation process. Similar to traditional neural networks, an increase in the number of DNN layers will make the network structure more complex. At this time, it is more difficult to adjust the hyperparameters of DNN (such as the number of network layers, the number of hidden units, activation functions, and optimization methods, etc.), and a considerable amount of time and effort are required for setting and adjustment. Because the setting and adjustment of DNN hyperparameters are usually done manually based on a large amount of experience and professional knowledge. To solve this problem, this paper proposes a method for automatically optimizing DNN hyperparameters based on particle swarm optimization.

[0049] The particle swarm algorithm, also known as the bird flock algorithm, is evidently inspired by the foraging behavior of bird flocks. It belongs to genetic algorithms and swarm intelligence algorithms. The particle swarm algorithm focuses on two attributes of particles: position and velocity. Each particle searches independently in space. They remember the optimal solutions they have found and also know the current optimal solution found by the entire particle swarm. Where to go next depends on the current direction of the particle, the direction of the optimal solution it has found, and the direction of the current optimal solution of the entire particle swarm. The main process is as follows: First, PSO initializes a population of individuals with a position of and a velocity of , where each individual corresponds to a random candidate solution of the objective function. Then, according to the fitness value, the single extremum pbest, that is, the local optimal solution, and the global extremum gbest, the global optimal solution, are updated using equations. What is optimized using the particle swarm optimization is the hyperparameters of the DNN model. The parameter optimization process based on the particle swarm optimization algorithm can be summarized as the solution process of a mathematical optimization problem:

[0050]

[0051]

[0052] where v id and x id respectively represent the velocity component and position component of the i t th particle in the k 1 th generation, c 2is the learning rate, which respectively controls the evolution amplitudes towards the single best particle and the global best particle. The above process will be repeated until the expected error value is reached or the maximum number of iterations is reached. Finally, PSO outputs the best position of the particle, which corresponds to the optimal solution of the problem. Due to its easy implementation and minimal tunable parameters, PSO can be a good choice for solving optimization problems in DNNs.

[0053] This study introduces the particle swarm method, whose goal is to optimize and obtain the best learning rate and weight decay of DNN. The core idea of this work is: through the particle swarm optimization algorithm, automatically search for the optimal learning rate and L2 regularization strength to optimize the performance of the deep neural network, thereby improving the accuracy and generalization ability of the model. During the iterative evolution process, it gradually converges to the optimal value as the population evolves. Then, the evaluation function connects the particle swarm optimization algorithm and model training. It first receives parameters, then uses them to train the model, and returns the loss on the validation set of the model as an evaluation metric. This loss value is used for the search process of the particle swarm optimization algorithm. According to the individual optimal positions of the particle swarm, iteratively update the positions of each particle swarm, then find the optimal combination of hyperparameters, train the final model, and finally find the optimal combination of hyperparameters to train the optimal final model. Specifically, it includes the following 4 steps:

[0054] 1) Initialize parameters and population. In the constructor of the PSO class, some parameters of the PSO algorithm are initialized, including the inertia weight w, the individual and social learning factors c1 and c2, as well as the population size pN, the dimension dim, and the maximum number of iterations max_iter. Then, the init_Population method is called to initialize the positions and velocities of the population, and initialize the individual optimal solutions and the global optimal solution. Then, randomly initialize the values of these hyperparameters within the given value range of the hyperparameters.

[0055] 2) Construct the DNN model. The DNN model is constructed layer by layer using the current values of the hyperparameters.

[0056] 3) Obtain the optimization results. Obtain the global optimal solution, that is, the combination of hyperparameters that minimizes the validation set loss found during the entire optimization process, and record the fitness value (i.e., the validation set loss) of the global optimal solution after each iteration.

[0057] 4) Train the final model using the best hyperparameters. Train the deep regression model and return the performance of the model. Create a new instance of the deep learning model, such as the value of the loss function or the accuracy on the validation set, and use the best hyperparameters found by PSO to configure the optimizer (i.e., the learning rate and the L2 regularization coefficient).

[0058] Taking the vegetation index as the input and LAI as the output, it contains 5 linear layers. Set the initial hyperparameters and use the particle swarm optimization algorithm to optimize the hyperparameters to obtain the optimal hyperparameters: learning rate and L2 regularization strength. Train the small-area corn crops to obtain the corn LAI model and verify it in the large area.

[0059] Combining plant height and LAI to fit the growth trend and grading evaluation of corn:

[0060] LAI inversion model based on particle swarm optimization DNN:

[0061] Using Python to optimize the best fit between the extracted different indices and LAI, plant height. Specifically, first use some common fitting methods, including linear and non-linear (binomial, exponential, power, and logarithmic) fitting functions to construct the correlation between different vegetation indices and LAI, plant height respectively. Then, for each index, by comparing and analyzing the correlation between the index fitted by different fitting methods and LAI, plant height, select the fitting method with the best correlation with LAI, plant height as the optimal fitting model between the two. There are 8 common weight calculation methods as follows, namely AHP analytic hierarchy process, preference ranking organization method for enrichment evaluations (PROMETHEE), entropy method, principal component analysis, factor analysis, CRITIC weight, independence weight, information content weight. The comparison and explanation of their analysis principles are shown in Table 2 Weight Calculation Principles:

[0062] Table 2 Weight Calculation Principles

[0063]

[0064] The entropy method calculates weights using the entropy value information of the data, that is, the amount of information. Information is a measure of the order degree of the system, and entropy is a measure of the disorder degree of the system; according to the definition of information entropy, for a certain index, the entropy value can be used to judge the dispersion degree of a certain index. The smaller the information entropy value, the greater the dispersion degree of the index, and the greater the impact (i.e., weight) of the index on the comprehensive evaluation. If the values of a certain index are all equal, then the index has no effect on the comprehensive evaluation. Therefore, the tool of information entropy can be used to calculate the weights of each index and provide a basis for multi-index comprehensive evaluation. This type of method is applicable to methods where there are fluctuations in the data and the data fluctuations are regarded as a kind of information. The entropy method is used very frequently in the paper. In the present invention, the leaf area index and plant height data are used as indicators.

[0065] Steps of the entropy weight method:

[0066] 1. Since the measurement units of various indicators are not unified, before calculating the comprehensive indicator using them, it is necessary to perform standardization processing first, that is, convert the absolute value of the indicator into a relative value, so as to solve the homogenization problem of various inhomogeneous indicator values. Data standardization first performs the dimensionless processing on each indicator. Suppose there are m indicators given:

[0067] X 1 ,X 2 ,……X m

[0068] Among them:

[0069] X i ={x 1 ,x 2 ,……x n}

[0070] Suppose the values after standardizing the data of each indicator are:

[0071] Y 1 ,Y 2 ,……Y m

[0072] The meanings represented by the values of positive and negative indicators are different (the higher the value of the positive indicator, the better, and the lower the value of the negative indicator, the better). Therefore, different algorithms need to be used for data standardization processing of positive and negative indicators:

[0073]

[0074] 2. Calculate the ratio of each indicator under each scheme.

[0075] Suppose there are m secondary indicators for a certain primary indicator, and the data for n years have been obtained, which is recorded as a matrix. Under the same indicator, calculate the proportion of the values in each year to the total value, and the formula is as follows:

[0076]

[0077] 3. Calculate the information entropy of each indicator.

[0078] According to the definition of information entropy in information theory, the information entropy of a set of data is:

[0079] (If p ij =0, define E j =0)

[0080] 4. Determine the weight of each indicator.

[0081] According to the information entropy calculation formula, calculate the information entropy of each indicator as E 1 ,E 2 ,…,Em

[0082] 4.1 Calculate the weights of each index through information entropy.

[0083]

[0084] Here, k refers to the number of indexes, that is, k = m.

[0085] 4.2 Calculate the weights by calculating the information redundancy.

[0086] D j = 1 - E j

[0087] Then calculate the index weights:

[0088]

[0089] 5. Finally, calculate the comprehensive score of each scheme.

[0090]

[0091] Principal component analysis and factor analysis:

[0092] When determining the weights by the principal component analysis method and the factor analysis method, the data dimensionality reduction principle is utilized, and the weights are mainly calculated by using the eigenvalue, variance interpretation rate, and loading coefficient.

[0093] The embodiment of the present invention provides an inversion system for the growth of crops, including:

[0094] A data acquisition module, configured to obtain the red edge index RE IP and the red edge index I REC I according to the reflectance of the spectral image of the crop planting area; and combine the red edge index RE IP and the red edge index I REC I in different proportions to obtain a combined index of the first proportion and a combined index of the second proportion. A model training module, configured to train a first deep neural network DNN through the combined index of the first proportion to predict the leaf area index, and optimize the learning rate and regularization parameters of the first deep neural network DNN by the particle swarm method during the training process to construct a leaf area index inversion model; train a second deep neural network DNN through the combined index of the second proportion to predict the plant height data and obtain a plant height inversion model. A growth inversion module, configured to combine the red edge index RE IP and the red edge index I REC I of the spectral image of the crop planting area to be measured in the first proportion and the second proportion respectively; substitute the combined index of the first proportion to be measured into the leaf area index inversion model to obtain the leaf area index, and substitute the combined index of the second proportion to be measured into the plant height inversion model to obtain the plant height data; fit the leaf area index and the plant height data by the entropy weight method to obtain the index weight result for evaluating the growth of crops.

[0095] An embodiment of the present invention provides a computer device, including: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the steps of an inversion method for the growth trend of crops are implemented.

[0096] An embodiment of the present invention provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the steps of an inversion method for the growth trend of crops are implemented.

[0097] A specific embodiment is as follows:

[0098] This study uses Sentinel-2 images as the basis of remote sensing data, combined with the measured plant height and LAI data in the field.

[0099] First, extract the corn planting area from the preprocessed Sentinel remote sensing image, and then calculate eight vegetation indices, namely NDVI, PVI, SAVI, IRECI, PSSRa, MTCI, REIP, and MCARI, for the corn planting area. Obtain multiple vegetation indices based on the reflectance contained in the spectral image of the crop planting area, use the feature recursive elimination algorithm to perform feature selection on the multiple vegetation indices, select the red edge index REIP and the red edge index I RECI, determine REIP and IRECI as the preferred red edge vegetation indices, and combine the red edge index REIP and the red edge index I RECI in different proportions.

[0100] Then, establish a plant height inversion model and an LAI inversion model based on the REIP and IRECI red edge indices respectively. The plant height inversion model is based on the DNN model, and 80% of the measured data is used as training data. Since the LAI inversion model directly established using DNN has poor effects, the particle swarm method is introduced to optimize the learning rate and regularization parameters in DNN, and a DNN inversion LAI model based on particle swarm optimization is established.

[0101] Finally, the remotely sensed plant height and LAI are fitted using the entropy weight method to obtain the remote sensing inversion result of the corn growth trend combined with key indicators and conduct an artificial grading evaluation of the growth trend.

[0102] The above embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent should be subject to the appended claims.

Claims

1. A method for inverting the growth potential of crops, characterized in that: The following steps are involved: Obtain a red edge index REIP and a red edge index IRECI according to the reflectance of the spectral image of the crop planting area; and combine the red edge index REIP and the red edge index IRECI in different proportions to obtain a combination index of a first proportion and a combination index of a second proportion; The first deep neural network DNN is trained by the combination index of the first ratio to predict the leaf area index, and the learning rate and regularization parameter of the first deep neural network DNN are optimized by the particle swarm method during the training process to construct a leaf area index inversion model; the second deep neural network DNN is trained by the combination index of the second ratio to predict the plant height data and obtain the plant height inversion model; The red edge index REIP and the red edge index IRECI of the spectral image of the crop planting area to be tested are respectively combined according to a first ratio and a second ratio; the combined index of the first ratio to be tested is substituted into the leaf area index inversion model to obtain the leaf area index, and the combined index of the second ratio to be tested is substituted into the plant height inversion model to obtain the plant height data; The leaf area index and plant height data were fitted by entropy weight method to obtain the index weight results for evaluating the growth of crops; The leaf area index inversion model includes: obtaining a red edge index REIP and a red edge index IRECI combined according to a first ratio; setting initial hyperparameters in a first deep neural network DNN including 5 linear layers, optimizing the hyperparameters using a particle swarm optimization algorithm to obtain a learning rate and an L2 regularization strength, and training the optimized first deep neural network DNN through the combined index of the first ratio to predict the leaf area index, so as to construct a leaf area index inversion model; The plant height inversion model includes: obtaining a red edge index REIP and a red edge index IRECI combined according to a second ratio; setting hyperparameters in a second deep neural network DNN including 5 linear layers, setting a learning rate to 0.0001, adjusting a regularization strength to 0.2, training the second deep neural network DNN through the combination index of the second ratio to predict plant height data, and obtaining a plant height inversion model.

2. The inversion method for crop growth as claimed in claim 1, characterized in that: The red edge index REIP and the red edge index IRECI are combined in different proportions, and the specific steps include: The red edge index REIP and the red edge index IRECI are combined in a first ratio of 0.3:0.7; The red edge index REIP and the red edge index IRECI are combined in a second ratio of 0.2:0.

8.

3. The inversion method for crop growth as claimed in claim 1, characterized in that: The obtaining of the red edge index REIP and the red edge index IRECI specifically includes: According to the reflectance contained in the spectral image of the crop planting area, multiple vegetation indices are obtained, and the feature recursive elimination algorithm is used to perform feature selection on the multiple vegetation indices to select the red edge index REIP and the red edge index IRECI.

4. The inversion method for crop growth as claimed in claim 1, characterized in that: The specific steps of obtaining the index weight results for evaluating the growth of crops include: The leaf area index and plant height data were standardized to eliminate the effects of different dimensions; The information entropy of each indicator is obtained from the standardized data, and the weight result is obtained using information redundancy.

5. A crop growth inversion system, characterized in that: include: A data acquisition module, used for acquiring a red edge index REIP and a red edge index IRECI according to the reflectance of the spectral image of the crop planting area; and combining the red edge index REIP and the red edge index IRECI in different proportions to obtain a combination index of a first proportion and a combination index of a second proportion; A model training module is used to train a first deep neural network DNN through a combination index of a first ratio to predict a leaf area index, and optimize a learning rate and a regularization parameter of the first deep neural network DNN through a particle swarm method during the training process to construct a leaf area index inversion model; train a second deep neural network DNN through a combination index of a second ratio to predict plant height data, and obtain a plant height inversion model; The growth inversion module is used to combine the red edge index REIP and the red edge index IRECI of the spectral image of the crop planting area to be tested according to a first ratio and a second ratio respectively; substitute the combined index of the first ratio to be tested into the leaf area index inversion model to obtain the leaf area index, and substitute the combined index of the second ratio to be tested into the plant height inversion model to obtain plant height data; fit the leaf area index and the plant height data through the entropy weight method to obtain the index weight result for evaluating the growth of crops; wherein the leaf area index inversion model includes: obtaining the red edge index REIP and the red edge index IRECI combined according to the first ratio; in a first deep neural network DNN including 5 linear layers Initial hyperparameters are set in the deep neural network, the hyperparameters are optimized by using a particle swarm optimization algorithm to obtain a learning rate and an L2 regularization strength, and a first deep neural network DNN is optimized by training the combination index of the first ratio to predict the leaf area index, so as to construct a leaf area index inversion model; the plant height inversion model comprises: obtaining a red edge index REIP and a red edge index IRECI combined according to a second ratio; hyperparameters are set in a second deep neural network DNN including 5 linear layers, the learning rate is set to 0.0001, the regularization strength is adjusted to 0.2, and the second deep neural network DNN is trained by the combination index of the second ratio to predict plant height data to obtain a plant height inversion model.

6. A computer device comprising: Memory and processor; The memory stores a computer program, wherein the processor implements a crop growth inversion method according to any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, an inversion method for crop growth potential according to any one of claims 1 to 4 is implemented.

Images