Wheat scab remote sensing prediction method based on combination of multi-temporal multi-feature and interpretable machine learning algorithm

Abstract

The invention relates to a wheat scab remote sensing prediction method based on the combination of multi-temporal multi-feature and an interpretable machine learning algorithm. Compared with the prior art, the method solves the defect that the prediction precision of the wheat scab is difficult to improve based on the combination of multi-temporal multi-feature. The method comprises the following steps: acquiring and preprocessing single-temporal and multi-temporal data; constructing a multi-temporal feature; screening the optimal characteristics of the wheat scab; constructing and training a wheat scab prediction model; obtaining a remote sensing image to be predicted; and obtaining a wheat scab remote sensing prediction result. According to the method, multiple time phases and multiple features of wheat are combined, the optimal XGBoost hyper-parameter training model is searched based on the simulated annealing algorithm, a rational classification model is obtained, and accurate prediction of the wheat scab remote sensing image is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition, and in particular to a remote sensing prediction method for wheat fusarium head blight based on the combination of multi-phase and multi-feature with an explainable machine learning algorithm. Background Art

[0002] Fusarium head blight is one of the most harmful diseases in wheat production, which has a serious impact on wheat yield and quality. The disease usually occurs in humid and warm climatic conditions, and mainly infects the ears of wheat from the flowering stage to the filling stage. It spreads quickly and is difficult to prevent and control, causing huge economic losses to the wheat planting industry. Therefore, the prevention and control of fusarium head blight has become a hot topic in global wheat planting and disease prevention and control research. Traditional fusarium head blight monitoring methods mostly conduct habitat monitoring by analyzing environmental factors and action mechanisms closely related to the pathogen. These methods are effective to a certain extent, but usually need to be combined with modern technologies, such as remote sensing and disease prediction and early warning systems, to improve the efficiency of monitoring and prevention and control.

[0003] With the continuous development of earth observation technology, remote sensing technology is increasingly used in crop disease monitoring, which helps to improve monitoring accuracy. For example, Liu used wheat canopy hyperspectral data to construct a spectral index for Fusarium head blight based on the red edge band, and applied it to Sentinel-2 satellite images to monitor Fusarium head blight in the field. Zhang proposed a Fusarium classification index (FCI) based on hyperspectral microscopy imaging technology. Experiments show that FCI outperforms many common spectral vegetation indices in terms of disease recognition rate and has better performance.

[0004] In recent years, machine learning methods have been widely used in the early prediction and monitoring of Fusarium head disease. For example, Shah used random forest (RF) to construct a prediction model for Fusarium head disease, and the results showed that the prediction performance of the RF model was better than that of the simple logistic regression (LR) model. Ruan extracted multi-period vegetation indices of wheat stripe rust and constructed a support vector machine (SVM) prediction model, which was significantly more accurate than the k-nearest neighbor (KNN) and back propagation neural network (BPNN) methods. Zhang used hyperspectral images to develop a deep convolutional neural network model that can quantitatively extract Fusarium head disease and accurately distinguish the severity of the disease.

[0005] Traditional machine learning models perform well in disease detection, but they usually lack interpretability. Therefore, interpretable machine learning has become a research focus. Bernardes combined low-cost imaging technology with deep learning, and by optimizing specific hyperparameters of RGB images, he not only improved the accuracy of ergot detection, but also enhanced the interpretability and visualization capabilities of the model. Ahmed used hyperspectral imaging technology and interpretable artificial intelligence models to improve the accuracy of sweet potato quality detection while providing a reliable quality assessment method. This shows that interpretability is playing an increasingly important role in machine learning research.

[0006] Machine learning has the advantages of processing massive amounts of data, high prediction accuracy, and strong generalization capabilities, but it also faces problems such as high implementation costs, high data quality requirements, lack of transparency, and possible overfitting. Interpretable machine learning models improve the transparency of the model's decision-making process, allowing agricultural experts and decision makers to understand the model's prediction basis, thereby enhancing trust in the prediction results. When errors occur in the prediction results, such models are easier to locate problems and make adjustments, thereby improving the accuracy of the model. At present, most of the research on the prediction of wheat fusarium wilt focuses on the characteristics of a single phase, and the exploration of prediction methods combining multiple phases and multiple features is still insufficient. There is still great research potential in this area.

[0007] Therefore, how to use multiple time periods and multiple features to establish a prediction model for ergot disease and build a high-precision prediction model is an issue that needs to be addressed urgently. Summary of the invention

[0008] The purpose of the present invention is to solve the defect in the prior art that it is difficult to improve the prediction accuracy of wheat fusarium scabiei based on the combination of multi-temporal and multi-features and an interpretable machine learning algorithm to solve the above problem.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] The remote sensing prediction method of wheat fusarium head blight based on multi-phase multi-feature and interpretable machine learning algorithm includes the following steps:

[0011] Acquisition and preprocessing of single-phase and multi-phase data: Sentinel-2 images were used to extract the filling and greening period data of wheat sample points, and single-phase and multi-phase data were constructed and preprocessed;

[0012] Constructing multi-temporal features: extracting vegetation index VIs, texture index TFs and color index CIs of wheat filling period as single-temporal features, selecting the features of wheat greening period, and performing normalized difference processing on the feature data of wheat filling period and greening period to construct multi-temporal features;

[0013] Screening for optimal characteristics of wheat scab;

[0014] Construction and training of wheat fusarium scab prediction model: A wheat fusarium scab prediction model was established based on the XGBoost algorithm. The optimal features were input into the wheat fusarium scab prediction model and trained in combination with the simulated annealing algorithm.

[0015] Acquisition of remote sensing images to be predicted: Acquisition of remote sensing images to be predicted and preprocessing;

[0016] Acquisition of remote sensing prediction results of wheat fusarium scabiei: The preprocessed remote sensing image to be predicted is input into the trained wheat fusarium scabiei prediction model to obtain the remote sensing prediction results of wheat fusarium scabiei.

[0017] The acquisition and preprocessing of the single-phase and multi-phase data include the following steps:

[0018] Data from 167 winter wheat fields at the grain filling stage were collected, and the infection status of wheat was recorded by visual inspection. The disease severity of the fields was divided into two categories: healthy and diseased. Sample point data of 44 fusarium head blight infected fields and 123 healthy fields were obtained.

[0019] Upload the sampled wheat sample point vector data and the study area vector file on the GEE platform, obtain the Sentinel-2 image data of wheat in the study area during the grain filling and greening periods, filter out images with cloud cover less than 10, perform cropping and fusion processing, and obtain image data of the grain filling area and greening period of the study area;

[0020] The data of the grain filling area and the greening period of the study area were extracted on the GEE cloud platform. The wheat grain filling period data was taken as single-phase data. The wheat grain filling period and greening period data were normalized and differenced to construct multi-phase data. The formula for constructing multi-phase data T is:

[0021]

[0022] The multi-temporal feature construction comprises the following steps:

[0023] Calculate vegetation indexes VIs, texture indexes TFs and color indexes CIs according to the bands in the Sentinel-2 image, and extract the image values ​​at the location of the sample point;

[0024] The spectral reflectance values ​​of wheat ears were used to extract 13 vegetation indices, 9 texture features and 6 color indices at the grain filling and greening stages of wheat sample points, and single-phase and multi-phase features were constructed.

[0025] The extracted VIs, TFs and CIs of wheat in the grain filling stage were taken as single-temporal features, and the VIs, TFs and CIs data of wheat in the grain filling stage and the greening stage were normalized and used as multi-temporal features.

[0026] The screening of the optimal characteristics of wheat scab comprises the following steps:

[0027] The weight relationship between vegetation index, texture characteristics and color index and wheat fusarium rust under single-phase and multi-phase conditions was calculated using the relief algorithm, and 7 vegetation indices, 6 texture characteristics and 3 color indices with large weight values ​​were screened out.

[0028] Use the relief algorithm to calculate the feature weights and initialize the weight W of each feature i =0,

[0029] Where i = 1, 2, ..., m, m is the total number of sample features, randomly select a sample x from the data set, and find the two nearest neighbor samples x of x. H and x M ;

[0030] Update the weight of each feature and update the weight W according to the following formula i :

[0031]

[0032] Among them, W i is the weight of the i-th feature, x H is the nearest neighbor sample of the same type as x, x M is the heterogeneous nearest neighbor sample of x, n is the total number of samples, diff(x i ,x H,i ) represents samples x and x H The difference in the i-th feature, diff(x i ,x M,i ) represents samples x and x M The difference in the i-th feature;

[0033] Repeat the steps of randomly selecting samples and updating weights many times until all samples are traversed and feature weights W = {W1, W2, ..., W m}, W represents the importance of each wheat characteristic;

[0034] The synthetic minority oversampling technique was used to balance the training samples, increasing the 44 disease samples to 123, so that the healthy plots and the Fusarium fusarium infected plots accounted for 50% each;

[0035] The minority class sample set is Y = {y1,y2,…,y b}, b is the number of minority class samples, for minority class samples y i , randomly select d nearest neighbor samples, and randomly select a sample y from these d nearest neighbor samples j , generate a new synthetic sample ynew ; d is the specified number of nearest neighbors, d = 5;

[0036] Generate new synthetic samples y using the SMOTE algorithm new The formula is as follows:

[0037] y new =y i +λ(y j -y i ),

[0038] Among them, λ is a random number in the interval [0,1], which is used to control the position of the synthetic sample; y i is a minority class sample; y new is a synthetic sample generated by interpolation; y j From y i A sample randomly selected from the nearest neighbor samples of ;

[0039] For each sample y i Generate a specified number of synthetic samples until the balance target is reached;

[0040] The optimal features were screened using the sequential forward selection method, and six VIs, four TFs, and two CIs were screened in the single-phase, and three VIs, four TFs, and two CIs were screened in the multi-phase;

[0041] The formula for screening the optimal features using the sequential forward selection method is as follows:

[0042]

[0043] Among them, arg max represents the parameter value of the maximization operation; ∈ is the belonging symbol; ∪ is the union symbol; the feature set Z after the SMOTE algorithm is synthesized is {z1,z2,…,z h}, h is the total number of samples after synthesis; S represents the currently selected feature subset; Z\S represents the feature set that has not been selected; z i Represents a single feature in the feature set Z; z i * represents the optimal feature selected in the current round; J(S) represents the objective function, which evaluates the performance of feature subset S;

[0044] Repeatedly select the feature z that maximizes the gain of the objective function i , the selected feature z i * Add to the feature subset S, and finally filter out the best features;

[0045] The optimal features screened were combined respectively to obtain the optimal characteristics of wheat fusarium rust in the form of VIs, VIs+TFs, VIs+CIs, and VIs+TFs+CIs feature combinations.

[0046] The construction and training of the wheat scab prediction model includes the following steps:

[0047] The optimal features of wheat fusarium head blight were used as the input features of the prediction model. The data set was divided into a training set and a test set with a ratio of 7:3. The winter wheat sample data was divided into two categories: healthy and diseased. 0 represented diseased winter wheat and 1 represented healthy wheat.

[0048] The XGBoost algorithm was used to build an initial wheat scab prediction model and train a binary classification winter wheat recognition model;

[0049] The simulated annealing algorithm is used to tune the hyperparameters of XGBoost. Simulated annealing searches within the set hyperparameter space, and the simulated annealing temperature T is updated according to the annealing strategy.

[0050] T q+1 =σ·T q ,

[0051] Where σ is the cooling coefficient, 0<σ<1; q is the iteration round, the initial temperature starts from T0, and through multiple iterations (q=1,2,…), the temperature T is gradually reduced. q , until it converges to the global optimal solution and stops iterating;

[0052] The state acceptance probability formula used to find the optimal hyperparameters using simulated annealing is as follows:

[0053]

[0054] Where, ΔE=f(a new )-f(a curr ent ), f(a) is the XGBoost hyperparameter function; a current Represents the currently selected hyperparameter combination; a new is a new candidate solution, a new hyperparameter combination randomly generated in the current hyperparameter space; ΔE is the change in the objective function value between the current solution and the new candidate solution; P(ΔE,T) is the state acceptance probability, which determines whether to accept the new candidate solution;

[0055] The hyperparameter ranges set using the XGBoost model are as follows:

[0056] n_estimators(50,500), max_depth(3,10), learning_rate(0.01,0.3), subsample(0.5,1), colsample_bytree(0.5,1),gamma(0,5);

[0057] The final XGBoost model was trained by using simulated annealing to find the optimal parameters in the hyperparameters set by the XGBoost model, and finally a binary classification wheat fusarium scaly training model with multiple phases and multiple features was obtained after training.

[0058] The prediction function of the XGBoost model is as follows:

[0059]

[0060] Among them, K represents the number of decision trees in the XGBoost model, e represents the characteristic value of the winter wheat fusarium rust sample; g k (e) represents the prediction of the kth decision tree for the input sample e, and f(e) is the final output value of the prediction function, which is used to determine the severity of the winter wheat fusarium rust disease;

[0061] During the XGBoost model training process, the structure and leaf node weight of each decision tree are determined by minimizing the objective function. The objective function consists of the following two parts:

[0062]

[0063] Among them, ∑ is the summation operation, is the loss function used to measure the predicted value and the true value r i The error between them, N is the total number of samples;

[0064] is the regularization term,

[0065]

[0066] V is the number of leaf nodes in the tree; ω is the weight of the leaf node, γ and is a regularization hyperparameter used to control model complexity.

[0067] Beneficial Effects

[0068] The invention discloses a remote sensing prediction method for wheat fusarium scabiei based on the combination of multi-phase and multi-feature with explainable machine learning algorithm. Compared with the prior art, it combines the multi-phase and multi-feature of wheat, searches for the optimal XGBoost hyperparameter training model based on the simulated annealing algorithm, and obtains a rational classification model, thereby realizing the accurate prediction of remote sensing images of wheat fusarium scabiei.

[0069] The invention uses wheat data of the greening period and the filling period to construct multi-temporal features; extracts multi-temporal and single-temporal vegetation indexes, texture features and color indexes for predicting wheat fusarium wilt; uses a relief algorithm to preliminarily screen features; uses a synthetic minority oversampling technique (SMOTE) algorithm to resample an unbalanced training data set, and uses a sequential forward selection (SFS) method to screen out optimal features; uses the optimal VIs, TFs and CIs as input features, uses a simulated annealing algorithm to optimize the hyperparameters of XGBoost, establishes an interpretable machine learning method, and uses the Shapley additive explained (SHAP) value to achieve the interpretability of the model; and evaluates the performance of a wheat fusarium wilt prediction model based on the XGBoost method.

[0070] Through verification, the XGBoost model trained based on the simulated annealing algorithm proposed in the present invention was tested on four feature data sets of multi-temporal and single-temporal VIs, VIs+TFs, VIs+CIs, and VIs+TFs+CIs. The experimental results show that the simulated annealing training XGBoost model proposed in the present invention has the highest model accuracy on multi-temporal and multi-feature data sets, and can obtain relatively ideal wheat remote sensing image classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a method sequence diagram of the present invention;

[0072] Figure 2 The weighted pseudo-color image distribution diagram of VIs, TFs and CIs in single phase and multi-phase of the present invention;

[0073] Figure 3 A pseudo-color image line graph showing how the accuracy of the model changes with the increase in the number of considered features under the SFS algorithm of the present invention;

[0074] Figure 4 This is a pseudo-color image prediction map of wheat fusarium in the study area under the XGBoost model of the present invention. DETAILED DESCRIPTION

[0075] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0076] like Figure 1 As shown, the remote sensing prediction method for wheat scab based on multi-phase multi-features combined with an interpretable machine learning algorithm described in the present invention comprises the following steps:

[0077] The first step is to obtain and preprocess single-phase and multi-phase data: use Sentinel-2 images to extract the filling and greening period data of wheat sample points, construct single-phase and multi-phase data and preprocess them. The specific steps are as follows:

[0078] (1) Data from 167 winter wheat fields during the grain filling period were collected. The infection status of wheat was recorded by visual inspection. The disease severity of the fields was divided into two categories: healthy and diseased. Sample point data for 44 fusarium head blight-infected fields and 123 healthy fields were obtained.

[0079] (2) Upload the sampled wheat sample point vector data and the study area vector file on the GEE platform, obtain the Sentinel-2 image data of wheat in the study area during the grain filling and greening periods, select images with cloud cover less than 10, perform cropping and fusion processing, and obtain image data of the grain filling area and greening period of the study area;

[0080] (3) The data of the grain filling area and the greening period of the study area were extracted on the GEE cloud platform. The wheat grain filling period data were taken as single-phase data. The wheat grain filling period and greening period data were normalized and differenced to construct multi-phase data. The formula for constructing multi-phase data T is:

[0081]

[0082] The second step is to construct multi-temporal features: extract vegetation index VIs, texture index TFs and color index CIs of wheat filling period as single-temporal features, select the features of wheat greening period, and perform normalized difference processing on the feature data of wheat filling period and greening period to construct multi-temporal features. The specific steps are as follows:

[0083] (1) Calculate vegetation indices VIs, texture indices TFs, and color indices CIs based on the bands in the Sentinel-2 image, and extract the image values ​​at the sample point;

[0084] (2) Using the spectral reflectance values ​​of wheat ears, 13 vegetation indices, 9 texture features, and 6 color indices were extracted at the wheat sample points during the filling and greening stages, and single-phase and multi-phase features were constructed;

[0085] (3) The extracted VIs, TFs and CIs of wheat during the filling period were used as single-temporal features, and the VIs, TFs and CIs data of wheat during the filling period and the greening period were normalized and used as multi-temporal features.

[0086] The third step is to screen the optimal features for wheat fusarium wilt. Using the Relief algorithm for preliminary screening and screening of features that are strongly correlated with the disease can reduce the interference of irrelevant features on the model. Using the SMOTE algorithm to resample the unbalanced data set, increase the minority class samples, and improve the classification model's recognition ability for the minority class, the prediction error caused by data bias can be reduced. Using the SFS algorithm to gradually add important features, screen out the optimal feature combination, prevent the dimensionality disaster caused by too many irrelevant features, and improve the generalization ability of the model. At the same time, the SFS algorithm needs to gradually test the contribution of each feature, and the calculation complexity is high, especially when the number of features is large, which may bring a large time cost. By screening features and reducing the number of features, the complexity of the model can be effectively controlled, thereby reducing the risk of overfitting and improving the accuracy of the model.

[0087] The specific steps are as follows:

[0088] (1) The relief algorithm was used to calculate the weight relationship between vegetation index, texture characteristics and color index and wheat fusarium rust under single-phase and multi-phase conditions, and 7 vegetation indices, 6 texture characteristics and 3 color indices with large weight values ​​were screened out;

[0089] Use the relief algorithm to calculate the feature weights and initialize the weight W of each feature i =0,

[0090] Where i = 1, 2, ..., m, m is the total number of sample features, randomly select a sample x from the data set, and find the two nearest neighbor samples x of x. H and x M ;

[0091] Update the weight of each feature and update the weight W according to the following formula i :

[0092]

[0093] Among them, W i is the weight of the i-th feature, x H is the nearest neighbor sample of the same type as x, x M is the heterogeneous nearest neighbor sample of x, n is the total number of samples, diff(x i ,x H,i ) represents samples x and x H The difference in the i-th feature, diff(x i ,x M,i ) represents samples x and x M The difference in the i-th feature;

[0094] Repeat the steps of randomly selecting samples and updating weights many times until all samples are traversed and feature weights W = {W1, W2, ..., Wm}, W represents the importance of each wheat characteristic;

[0095] (2) Using synthetic minority oversampling techniques to balance the training samples, the 44 disease samples were increased to 123, so that the healthy plots and the Fusarium fusarium infected plots accounted for 50% each;

[0096] The minority class sample set is Y = {y1,y2,…,y b}, b is the number of minority class samples, for minority class samples y i , randomly select d nearest neighbor samples, and randomly select a sample y from these d nearest neighbor samples j , generate a new synthetic sample y new ; d is the specified number of nearest neighbors, d = 5;

[0097] Generate new synthetic samples y using the SMOTE algorithm new The formula is as follows:

[0098] y new =y i +λ(y j -y j ),

[0099] Among them, λ is a random number in the interval [0,1], which is used to control the position of the synthetic sample; y j is a minority class sample; y new is a synthetic sample generated by interpolation; y j From y j A sample randomly selected from the nearest neighbor samples of ;

[0100] For each sample y j Generate a specified number of synthetic samples until the balance target is reached;

[0101] (3) The optimal features were screened using the sequential forward selection method, and 6 VIs, 4 TFs, and 2 CIs were screened in the single-phase, and 3 VIs, 4 TFs, and 2 Cis were screened in the multi-phase;

[0102] The formula for screening the optimal features using the sequential forward selection method is as follows:

[0103]

[0104] Among them, arg max represents the parameter value of the maximization operation; ∈ is the belonging symbol; ∪ is the union symbol; the feature set Z after the SMOTE algorithm is synthesized is {z1,z2,…,z h}, h is the total number of samples after synthesis; S represents the currently selected feature subset; Z\S represents the feature set that has not been selected; z iRepresents a single feature in the feature set Z; z i * represents the optimal feature selected in the current round; J(S) represents the objective function, which evaluates the performance of feature subset S;

[0105] Repeatedly select the feature z that maximizes the gain of the objective function i , the selected feature z i * Add to the feature subset S, and finally filter out the best features;

[0106] (4) The optimal features screened were combined respectively to obtain the optimal features of wheat fusarium rust in the form of VIs, VIs+TFs, VIs+CIs, and VIs+TFs+CIs feature combinations.

[0107] Step 4: Construction and training of wheat fusarium scabiei prediction model: A wheat fusarium scabiei prediction model was established based on the XGBoost algorithm. The optimal features were input into the wheat fusarium scabiei prediction model and trained in combination with the simulated annealing algorithm. The XGBoost model has many hyperparameters and has nonlinear or complex interactions. The simulated annealing algorithm can efficiently explore complex high-dimensional parameter spaces and find hyperparameter combinations with better performance through random perturbations and probability acceptance mechanisms. By optimizing hyperparameters, the complexity of the XGBoost model can be controlled, thereby effectively reducing the risk of overfitting and improving the generalization ability of the model on test data. The simulated annealing algorithm needs to evaluate the performance of the XGBoost model multiple times, so a large amount of model training and verification is required in the high-dimensional hyperparameter space, resulting in high computational overhead and a long optimization process. The optimization results may be affected by the random initial state or search path, resulting in unstable results. Using the simulated annealing algorithm to optimize the XGBoost hyperparameters can effectively improve the model performance, have global optimization capabilities and flexibility, and are suitable for complex parameter optimization problems. However, issues such as the algorithm's computational cost, parameter setting, and convergence speed are the main difficulties in the implementation process.

[0108] The specific steps are as follows:

[0109] (1) The optimal features of wheat fusarium head blight were used as the input features of the prediction model. The data set was divided into a training set and a test set with a ratio of 7:3. The winter wheat sample data was divided into healthy and diseased categories, with 0 representing diseased winter wheat and 1 representing healthy wheat.

[0110] (2) Use the XGBoost algorithm to build an initial wheat fusarium head blight prediction model and train a binary classification winter wheat recognition model;

[0111] (3) Use the simulated annealing algorithm to tune the hyperparameters of XGBoost. Simulated annealing searches within the set hyperparameter space, and the simulated annealing temperature T is updated according to the annealing strategy.

[0112] T q+1 =σ·T q

[0113] Where σ is the cooling coefficient, 0<σ<1; q is the iteration round, the initial temperature starts from t0, and through multiple iterations (q=1,2,…), the temperature T is gradually reduced. q , until it converges to the global optimal solution, stop the iteration;

[0114] The state acceptance probability formula used to find the optimal hyperparameters using simulated annealing is as follows:

[0115]

[0116] Where, ΔE=f(a new )-f(a curr ent ), f(a) is the XGBoost hyperparameter function; a current Represents the currently selected hyperparameter combination; a new is a new candidate solution, a new hyperparameter combination randomly generated in the current hyperparameter space; ΔE is the change in the objective function value between the current solution and the new candidate solution; P(ΔE,T) is the state acceptance probability, which determines whether to accept the new candidate solution;

[0117] (4) The hyperparameter ranges set by the XGBoost model are as follows:

[0118] n_estimators(50,500), max_depth(3,10), learning_rate(0.01,0.3), subsample(0.5,1), colsample_bytree(0.5,1),gamma(0,5);

[0119] (5) The optimal parameters found in the hyperparameters set by the XGBoost model were used to train the final XGBoost model, and finally a binary classification wheat fusarium scabiei training model with multiple phases and multiple features was obtained after training;

[0120] The prediction function of the XGBoost model is as follows:

[0121]

[0122] Among them, K represents the number of decision trees in the XGBoost model, e represents the characteristic value of the winter wheat fusarium rust sample; g k(e) represents the prediction of the kth decision tree for the input sample e, and f(e) is the final output value of the prediction function, which is used to determine the severity of the winter wheat fusarium rust disease;

[0123] During the XGBoost model training process, the structure and leaf node weight of each decision tree are determined by minimizing the objective function. The objective function consists of the following two parts:

[0124]

[0125] Among them, ∑ is the summation operation, is the loss function used to measure the predicted value and the true value r i The error between them, N is the total number of samples;

[0126] is the regularization term.

[0127]

[0128] V is the number of leaf nodes in the tree; ω is the weight of the leaf node, γ and is a regularization hyperparameter used to control model complexity.

[0129] The fifth step is to obtain the remote sensing image to be predicted: obtain the remote sensing image to be predicted and perform preprocessing.

[0130] Step 6: Obtaining remote sensing prediction results of wheat fusarium scabiei: Input the preprocessed remote sensing image to be predicted into the trained wheat fusarium scabiei prediction model to obtain the remote sensing prediction results of wheat fusarium scabiei.

[0131] The effect of the present invention is further described below in conjunction with simulation experiments:

[0132] 1. Simulation experiment conditions:

[0133] The computer hardware environment of the experiment of the present invention is Intel Core i5-11400, CPU 2.60 GHz, GPU uses GTX1050Ti, RAM 8 GB, the software environment is Windows 10 (64) bit operating system, and the compilation environment is Pycharm. All subsequent training and testing experiments are based on this platform.

[0134] The experiment of the present invention randomly divides healthy and diseased wheat samples into training sets and test sets (ratio is 7:3). The training set is used to train the model, and the samples in the training set are used to learn features and make predictions, while the test set is used to evaluate the performance of the trained model. The overall accuracy (OA), user's accuracy (UA), producer's accuracy (PA) and Kappa coefficient of the prediction model are used to evaluate the performance of the prediction model.

[0135] 2. Simulation experiment content and result analysis:

[0136] The present invention uses a relief algorithm to calculate the weight relationship between different characteristics and wheat fusarium wilt under single-phase and multi-phase conditions to reduce redundant characteristics. Figure 2 The importance distribution of VIs, TFs and CIs in different time phases is described. The larger the weight value, the more important the corresponding feature. As can be seen from the figure, in the case of multi-phase, the difference in weight values ​​of different features is more significant than that in the single phase. Using the relief algorithm, 7 VIs, 6 TFs and 3 CIs were preliminarily screened out, and the selected features are shown in Table 1.

[0137] Table 1 Features after preliminary screening of relief algorithm

[0138]

[0139] SFS was used to extract the best feature combination of multiple features at different phases. For 44 diseased plots and 123 healthy plots, the SMOTE algorithm was used to balance the training samples. After data balancing, the number of diseased plots increased to 123, with healthy plots and Fusarium head blight infected plots accounting for 50% each, as shown in Table 2.

[0140] Table 2 SMOTE balanced training data

[0141]

[0142] The balanced training samples are applied to SFS based on XGBoost. Figure 3The results show that the accuracy of the model changes with the increase in the number of features considered. It can be seen from the figure that the accuracy of the multi-temporal model is always higher than that of the single-temporal model. As the number of vegetation indices increases, the accuracy increases overall. When the number of features of the multi-temporal vegetation index is 3, the model accuracy reaches the highest, among which the combination of NDVIre2, RVI and TVI performs best. The combination of single-temporal vegetation indices GNDVI, EVI, RDVI, TVI, RVI and NDRE has the highest accuracy; in multi-temporal, the combination of texture features contrast, idm, savg and svar has the highest accuracy, while in single-temporal, the combination of contrast, idm, sent and var has the highest accuracy; the combination of GLI and EG obtains the highest accuracy of multi-temporal color index, and the combination of ER and VARI obtains the highest accuracy of single-temporal color index. The optimal feature combination is shown in Table 3.

[0143] Table 3 Optimal features obtained by SFS

[0144]

[0145] After proposing the optimal features in different phases, VIs, TFs and CIs were combined as model input features, and the wheat scab prediction model was established using the XGBoost method. The hyperparameters of XGBoost were optimized by the simulated annealing algorithm, and the model parameters are shown in Table 4.

[0146] Table 4 Optimizing XGBoost hyperparameters using simulated annealing algorithm

[0147]

[0148] The model accuracy varies significantly with different time phases and feature combinations. Under single-phase conditions, when only a single feature VIs is used as input, the overall accuracy of the XGBoost model is 83.8%, and the Kappa coefficient is 0.676. With the introduction of TFs and CIs features, the model accuracy gradually improves. When VIs+TFs and VIs+CIs are used as input, the overall accuracy reaches 85.1% and 87.8%, respectively, and the Kappa coefficients are 0.703 and 0.756, respectively. When multiple features (VIs+TFs+CIs) are further combined as input, the overall accuracy of the model reaches 89.2%, and the Kappa coefficient is 0.784, which is 5.4% and 0.108 higher than that of single feature VIs.

[0149] Under multi-temporal conditions, when VIs, VIs+TFs, VIs+CIs and VIs+TFs+CIs were used as input, the overall accuracy was 86.5%, 89.2%, 90.5% and 94.6%, and the Kappa coefficient was 0.730, 0.783, 0.811 and 0.892, which were significantly better than single-temporal feature input. When multi-temporal combined with multiple features (VIs+TFs+CIs) was used as input, the overall accuracy of the model was improved by 10.8% compared with single-temporal single feature (VIs). The overall prediction accuracy of the XGBoost model is shown in Table 5.

[0150] Table 5. Validation of prediction model based on XGBoost

[0151]

[0152]

[0153] The prediction results of spatial distribution of wheat fusarium scaly obtained by training different models under multi-temporal and multi-feature conditions are as follows: Figure 4 As shown. Figure 4 It can be seen that the infection area of ​​fusarium head blight is widely distributed and dispersed, especially in areas close to water bodies, where the incidence of the disease is significantly higher. This may be closely related to the environmental characteristics and high humidity around the water bodies. The humidity conditions near the water bodies are more suitable for the growth and spread of wheat fusarium head blight pathogens, making the wheat in this area more susceptible to infection.

[0154] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A remote sensing prediction method for wheat scab based on multi-temporal and multi-feature combined with an interpretable machine learning algorithm, characterized in that: The following steps are involved: 11) Acquisition and preprocessing of single-phase and multi-phase data: Use Sentinel-2 images to extract the filling and greening period data of wheat sample points, construct single-phase and multi-phase data and perform preprocessing; 12) Constructing multi-temporal features: Extracting vegetation index VIs, texture index TFs and color index CIs of wheat in the grain filling period as single-temporal features, selecting the features of wheat in the greening period, and performing normalized difference processing on the feature data of wheat in the grain filling period and the greening period to construct multi-temporal features; 13) Screening of optimal characteristics of wheat fusarium head blight; 14) Construction and training of wheat fusarium scabiei prediction model: A wheat fusarium scabiei prediction model was established based on the XGBoost algorithm. The optimal features were input into the wheat fusarium scabiei prediction model and trained in combination with the simulated annealing algorithm. 15) Acquisition of remote sensing images to be predicted: Acquisition of remote sensing images to be predicted and preprocessing; 16) Obtaining remote sensing prediction results of wheat fusarium scabiei: Input the preprocessed remote sensing image to be predicted into the trained wheat fusarium scabiei prediction model to obtain the remote sensing prediction results of wheat fusarium scabiei.

2. The method for remote sensing prediction of wheat scab based on multi-phase multi-feature combined with an explainable machine learning algorithm according to claim 1, characterized in that: The acquisition and preprocessing of the single-phase and multi-phase data include the following steps: 21) Collect data from 167 winter wheat fields during the grain filling period, record the infection status of wheat by visual inspection, divide the disease severity of the fields into two categories: healthy and diseased, and obtain sample point data for 44 fusarium head blight infected fields and 123 healthy fields; 22) Upload the sampled wheat sample point vector data and the study area vector file on the GEE platform, obtain the Sentinel-2 image data of wheat in the study area during the grain filling and greening periods, filter out images with cloud cover less than 10, perform cropping and fusion processing, and obtain image data of the grain filling area and greening period of the study area; 23) The data of the grain filling area and the greening period of the study area were extracted on the GEE cloud platform. The wheat grain filling period data was taken as single-phase data. The wheat grain filling period and greening period data were normalized and differenced to construct multi-phase data. The formula for constructing multi-phase data T is:

3. The method for remote sensing prediction of wheat scab based on multi-phase multi-feature combined with an explainable machine learning algorithm according to claim 1, characterized in that: The multi-temporal feature construction comprises the following steps: 31) Calculate vegetation index VIs, texture index TFs and color index CIs according to the bands in Sentinel-2 images, and extract the image value at the location of the sample point; 32) The spectral reflectance values ​​of wheat ears were used to extract 13 vegetation indices, 9 texture features and 6 color indices at the filling and greening stages of wheat sample points, and single-phase and multi-phase features were constructed; 33) The extracted VIs, TFs and CIs of wheat during the filling stage were used as single-temporal features, and the VIs, TFs and CIs data of wheat during the filling stage and the greening stage were normalized and used as multi-temporal features.

4. The method for remote sensing prediction of wheat scab based on multi-phase multi-feature combined with an explainable machine learning algorithm according to claim 1, characterized in that: The screening of the optimal characteristics of wheat scab comprises the following steps: 41) The relief algorithm was used to calculate the weight relationship between vegetation index, texture characteristics and color index and wheat fusarium rust under single-phase and multi-phase conditions, and 7 vegetation indices, 6 texture characteristics and 3 color indices with large weight values ​​were screened out; Use the relief algorithm to calculate the feature weights and initialize the weight W of each feature i =0, Where i = 1, 2, ..., m, m is the total number of sample features, randomly select a sample x from the data set, and find the two nearest neighbor samples x of x. H and x M ; Update the weight of each feature and update the weight W according to the following formula i : Among them, W i is the weight of the i-th feature, x H is the nearest neighbor sample of the same type as x, x M is the heterogeneous nearest neighbor sample of x, n is the total number of samples, diff(x i ,x H,i ) represents samples x and x H The difference in the i-th feature, diff(x i ,x M,i ) represents samples x and x M The difference in the i-th feature; Repeat the steps of randomly selecting samples and updating weights many times until all samples are traversed and feature weights W = {W1, W2, ..., W m }, W represents the importance of each wheat characteristic; 42) The synthetic minority oversampling technique was used to balance the training samples, increasing the 44 disease samples to 123, so that the healthy plots and the Fusarium fusarium infected plots accounted for 50% each; The minority class sample set is Y = {y1, y2, ..., y b }, b is the number of minority class samples, for minority class samples y i , randomly select d nearest neighbor samples, randomly select a sample yj from these d nearest neighbor samples, and generate a new synthetic sample y new ; d is the specified number of nearest neighbors, d = 5; Generate new synthetic samples y using the SMOTE algorithm new The formula is as follows: and new =and i +λ(and j -and i ), Among them, λ is a random number in the interval [0, 1], which is used to control the position of the synthetic sample; y i is a minority class sample; y new is a synthetic sample generated by interpolation; y j From y i A sample randomly selected from the nearest neighbor samples of ; Generate a specified number of synthetic samples for each sample yi until the balance target is reached; 43) The optimal features were screened using the sequential forward selection method. Six VIs, four TFs, and two CIs were screened in the single-phase, and three VIs, four TFs, and two CIs were screened in the multi-phase; The formula for screening the optimal features using the sequential forward selection method is as follows: Among them, arg max represents the parameter value of the maximization operation; ∈ is the belonging symbol; ∪ is the union symbol; the feature set Z after the SMOTE algorithm is synthesized is {z1, z2, ..., z h }, h is the total number of samples after synthesis; S represents the currently selected feature subset; Z\S represents the feature set that has not been selected; z i Represents a single feature in the feature set Z; z i * represents the optimal feature selected in the current round; J(S) represents the objective function, which evaluates the performance of feature subset S; Repeatedly select the feature z that maximizes the gain of the objective function i , the selected feature z i * Add to the feature subset S, and finally filter out the best features; 44) The optimal features screened were combined respectively to obtain the optimal features of wheat fusarium wilt in the form of VIs, VIs+TFs, VIs+CIs, and VIs+TFs+CIs feature combinations.

5. The method for remote sensing prediction of wheat scab based on multi-phase multi-feature combined with an explainable machine learning algorithm according to claim 1, characterized in that: The construction and training of the wheat scab prediction model includes the following steps: 51) The optimal features of wheat fusarium head blight were used as the input features of the prediction model. The data set was divided into a training set and a test set with a ratio of 7:

3. The winter wheat sample data was divided into healthy and diseased categories, with 0 representing diseased winter wheat and 1 representing healthy wheat. 52) Use the XGBoost algorithm to build an initial wheat scab prediction model and train a binary classification winter wheat recognition model; 53) Use simulated annealing algorithm to tune the hyperparameters of XGBoost. Simulated annealing searches in the set hyperparameter space. The simulated annealing temperature T is updated according to the annealing strategy. T q+1 =σ·T q , Where σ is the cooling coefficient, 0<σ<1; q is the iteration round, the initial temperature starts from T0, and through multiple iterations (q=1,2,…), the temperature T is gradually reduced. q , until it converges to the global optimal solution and stops iterating; The state acceptance probability formula used to find the optimal hyperparameters using simulated annealing is as follows: Where, ΔE=f(a new )-f(a current ), f(a) is the XGBoost hyperparameter function; a current Represents the currently selected hyperparameter combination; a new is a new candidate solution, a new hyperparameter combination randomly generated in the current hyperparameter space; ΔE is the change in the objective function value between the current solution and the new candidate solution; P(ΔE,T) is the state acceptance probability, which determines whether to accept the new candidate solution; 54) The hyperparameter ranges set using the XGBoost model are as follows: n_estimators(50,500), max_depth(3,10), learning_rate(0.01,0.3), subsample(0.5,1), colsample_bytree(0.5,1),gamma(0,5); 55) The final XGBoost model is trained by using simulated annealing to find the optimal parameters in the hyperparameters set by the XGBoost model, and finally a binary classification wheat fusarium scabiei training model with multiple phases and multiple features is obtained after training; The prediction function of the XGBoost model is as follows: Among them, K represents the number of decision trees in the XGBoost model, e represents the characteristic value of the winter wheat fusarium rust sample; g k (e) represents the prediction of the kth decision tree for the input sample e, and f(e) is the final output value of the prediction function, which is used to determine the severity of the winter wheat fusarium rust disease; During the XGBoost model training process, the structure and leaf node weight of each decision tree are determined by minimizing the objective function. The objective function consists of the following two parts: Among them, ∑ is the summation operation, is the loss function used to measure the predicted value and the true value r i The error between them, N is the total number of samples; is the regularization term, V is the number of leaf nodes in the tree; ω is the weight of the leaf node, γ and is a regularization hyperparameter used to control model complexity.

Images