Wheat stripe rust monitoring method and device based on unmanned aerial vehicle multi-spectral multi-temporal feature weighting

Abstract

The present application relates to a kind of wheat stripe rust monitoring methods based on unmanned aerial vehicle multispectral multi-temporal feature weighting, comprising: obtaining the unmanned aerial vehicle multispectral image of three time phases in study area and pre-processing;Three single-time features are constituted;Multi-temporal fusion feature set is constructed;Wheat stripe rust multi-temporal fusion data set is constituted;Form the combined features for model training and test;Output the stripe disease health or infection determination result of wheat sample.The present application can be according to the response difference of various features in different disease development stages, adaptively determine the contribution size of each observation period, to improve the pertinence and rationality of weight distribution;Give full play to the complementary action between different types of features, enhance the characterization ability to wheat stripe rust;Can effectively reduce feature redundancy, reduce noise interference, improve the representativeness and compactness of feature subset, to improve the stability and generalization ability of model training.

Description

Technical Field

[0001] This invention relates to the fields of agricultural crop disease monitoring and UAV multispectral information processing technology, and in particular to a method and equipment for monitoring wheat stripe rust based on weighted multispectral and multi-temporal features of UAVs. Background Technology

[0002] Wheat stripe rust is one of the important fungal diseases affecting the stable and high yield of winter wheat, characterized by rapid spread, strong outbreaks, and a wide range of damage. Timely and accurate understanding of the occurrence and spatial distribution of stripe rust is an important prerequisite for disease early warning, control decision-making, and precision pesticide application. Traditional field surveys mainly rely on manual visual inspection and sampling statistics, which is not only labor-intensive and inefficient, but also significantly affected by the experience of the surveyors, making it difficult to quickly obtain continuous spatial information.

[0003] With the development of UAV remote sensing technology, using UAVs equipped with multispectral sensors for farmland disease monitoring has become an important technical approach. Existing methods typically extract vegetation indices, texture features, or color features from images from a single observation period, and then combine them with machine learning models to complete health and infection identification. While this type of method has the advantage of simplicity, the monitoring effect is easily affected by the choice of a single observation time. When the disease is at different infection stages, single-phase features are often insufficient to fully represent the differences in disease evolution.

[0004] To improve monitoring accuracy, some existing studies have introduced bi-temporal normalized difference features to characterize the changes between two observation periods. However, the bi-temporal difference method only utilizes the relative changes between two observation periods, making it difficult to fully integrate the complementary information existing in three or more observation periods; if multi-temporal features are simply spliced ​​together, it is easy to cause problems such as excessive dimensionality, redundancy enhancement, and noise accumulation.

[0005] Therefore, a wheat stripe rust monitoring method is needed that can adaptively allocate temporal weights for different features and integrate multiple types of information such as vegetation index, texture features and color index to improve monitoring accuracy and stability. Summary of the Invention

[0006] To address the issues of existing single-phase methods being sensitive to observation time and dual-phase difference methods lacking sufficient information utilization, the primary objective of this invention is to provide a wheat stripe rust monitoring method based on UAV multispectral and multi-phase feature weighting. This method can adaptively determine the contribution of each observation period based on the response differences of various features at different disease development stages, thereby improving the pertinence and rationality of weight allocation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for monitoring wheat stripe rust based on multispectral and multi-temporal characteristics weighted by unmanned aerial vehicles (UAVs), the method comprising the following sequential steps:

[0008] (1) Obtain UAV multispectral images of the study area in three time phases, plan sampling points and collect ground disease survey data corresponding to the sampling points, classify the ground disease survey data into healthy or infected according to the threshold, and save the classification results as labels;

[0009] (2) Preprocess the UAV multispectral image to obtain preprocessed data. The preprocessing includes radiometric correction, reflectivity conversion, geometric correction and multi-temporal georegistration. Mark the location of the sampling point in the preprocessed data. The wheat image at each sampling point location forms a sample.

[0010] (3) Extract the vegetation index feature, texture feature and color index feature of each sample from the UAV images of the three time phases respectively. The three features of each time phase constitute a single time phase feature set. The three time phase images constitute three single time phase feature sets respectively.

[0011] (4) For the sample vectors of the same feature in each phase of the three single-phase feature sets, the weight of each feature in different phases is calculated by the entropy weight method or the median absolute deviation method, and a multi-phase fusion feature set is constructed by weighted summation;

[0012] (5) Combine the multi-temporal fusion feature set with the labels to form a multi-temporal fusion dataset for wheat stripe rust;

[0013] (6) Divide the multi-temporal fusion dataset into a training set and a test set in a 7:3 ratio. In the training set, perform feature filtering on vegetation index features, texture features and color index features respectively to obtain the optimal feature subsets of each type of feature. Combine the optimal feature subsets of each type of feature to form a combined feature for model training and testing.

[0014] (7) Input the combined features into the machine learning classification model for training to obtain the trained model. Then input the test set into the trained model and output the stripe rust health or infection result of the wheat sample.

[0015] In step (1), the UAV multispectral imagery is acquired by a UAV equipped with a multispectral sensor, which includes blue, green, red, red-edge, and near-infrared bands; the three time phases are the wheat booting stage, heading stage, and flowering stage; the disease index DI is obtained based on ground disease survey data, and the wheat samples at the sampling points are divided into healthy and infected grades according to a preset threshold; the disease index DI is calculated based on the number of samples for each disease grade and the corresponding severity representative value, and the calculation formula is:

[0016]

[0017] In the formula, The disease severity is classified into 1 to 8 levels based on the proportion of stripe rust infection area to the total leaf area. The disease level of the sampled leaves is numbered as follows: Number of blades per stage; for The representative value for severity level, from arrive The percentages are 1%, 5%, 10%, 20%, 40%, 60%, 80%, and 100%, respectively. is the highest severity level of the sampled leaf in the current sample; m is the highest disease level of the actual sampled leaf in the current sample; when DI≤5%, it is judged as a healthy sample, and when DI>5%, it is judged as an infected sample, thus forming the supervision label required for subsequent classification modeling.

[0018] Step (3) specifically includes the following:

[0019] (3a) Construct a 7×7 pixel neighborhood window centered on the sampling point and extract the average reflectance of each band; the neighborhood window is preferably a 7×7 pixel window;

[0020] (3b) Calculate vegetation index characteristics based on reflectance of each band;

[0021] (3c) Based on the gray-level co-occurrence matrix, extract the texture features of the window in four directions and calculate the mean of the four directions as the final texture features; the four directions include 45°, 90°, 135° and 180°;

[0022] (3d) Calculate color index features based on the reflectance of red, green and blue bands; the vegetation index features include NDVI, MSR, TVI, SIPI, EVI, RVI, PSRI, GNDVI, RG and RDVI; the texture features include mean, variance, uniformity, contrast, variability, information entropy, second moment and correlation; the color index features include ExG, ExR, ExGR, GLI, NGRDI and VARI.

[0023] In step (4), the calculation of the weights of each feature at different time phases using the entropy weight method or the median absolute deviation method specifically includes the following steps:

[0024] (4a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the impact of outliers on features... The interference of weight calculation results at different time phases is first addressed by considering the features. The sample vectors at each time phase are subjected to truncation robustness processing, compressing extreme values ​​outside the quantile interval to the corresponding quantile range, thus obtaining the truncated feature values. :

[0025]

[0026] In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature;

[0027] (4b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on subsequent entropy calculations, the truncated characteristic values ​​are... Further normalization is performed to make the same feature comparable across different time periods, resulting in a sample. In the The first phase of time The eigenvalues ​​after normalization of each feature :

[0028]

[0029] In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0;

[0030] (4c) represents the normalized eigenvalues. This is transformed into a relative proportion that can be used for information entropy calculation. For each time phase, it is calculated according to the... Calculate the probability value corresponding to each sample by using the normalized feature values ​​of each feature across all samples. :

[0031]

[0032] In the formula, Indicates the total number of samples;

[0033] (4d) To measure the dispersion and information richness of the same feature across different time phases, based on the samples in the first time phase... Calculate the information entropy value corresponding to each time point based on the probability value at each time point. :

[0034]

[0035] In the formula, The entropy value is the standardized coefficient. This is used to constrain entropy values ​​to a comparable range;

[0036] (4e) In order to further transform the differences reflected by information entropy into an evaluation quantity that can be used for weighting, the information entropy values ​​of each time phase are... Calculate the information utility value separately :

[0037]

[0038] In the formula, The larger the value, the higher the value. The first phase of time The stronger the ability of a feature to distinguish differences between samples;

[0039] (4f) To map the information utility values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the first... The feature in the first The first initial weight at each time phase :

[0040] In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution. The smaller the value, the easier it is for the weight to concentrate on the phases with higher scores, and vice versa.

[0041] (4g) To avoid the first In a certain time phase, a feature receives too little weight due to a low score, thus weakening the complementary effect of multi-time phase information. Therefore, a preset lower limit constraint is applied to the initial weights to obtain the [previous] feature. The feature in the first The first intermediate weight in each phase :

[0042]

[0043] In the formula, This represents the lower limit of the weight, set to 0.02, to prevent the weight from being lowered by the first weight. The weight of a feature is too small at a certain time;

[0044] (4h) To ensure the first The sum of the weights of each feature across the three time phases is 1. Further normalization is applied to the intermediate weights to obtain the result. The feature in the first The first final weight of each phase :

[0045]

[0046] (4i) According to the first final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Entropy weighting method for multi-temporal fusion results of individual features :

[0047]

[0048] In step (4), the calculation of the weights of each feature at different time phases using the median absolute deviation method specifically includes the following steps:

[0049] (5a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the interference of outliers on the calculation results of the same feature's dispersion at different time phases, firstly, for the... The sample vectors of each feature at each time phase are truncated and robustly processed to limit extreme values ​​outside the quantile interval to a preset range, resulting in truncated feature values. :

[0050] In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature;

[0051] (5b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on the dispersion evaluation, the truncated eigenvalues ​​are normalized to obtain the normalized eigenvalues. :

[0052]

[0053] In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0;

[0054] (5c) To obtain the robust central location of this feature in the sample population at each time phase, calculate the first... The first phase of time The median of each feature in all samples :

[0055] In the formula, Indicates the first The first phase of time The median of each feature across all samples;

[0056] (5d) To measure the degree of deviation of each sample feature value from the median, the absolute deviation is calculated. :

[0057]

[0058] (5e) To further obtain the robustness of this feature across different time phases, the absolute deviation is... Take the median to get the first... The corresponding absolute deviation of the median at each time point :

[0059]

[0060] In the formula, The larger it is, the more likely it is to be the first At this time, the first The more significant the differences between samples, the stronger the discriminative information that a feature may contain.

[0061] (5f) To map the median absolute deviation values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the th... The feature in the first The second initial weights obtained based on MAD score mapping at each time phase :

[0062]

[0063] In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution;

[0064] (5g) To avoid a certain time phase receiving too low a weight due to its small dispersion, thus affecting the comprehensive utilization of multi-time phase information, a preset lower limit constraint on the weight is applied based on the initial weights to obtain the first... The feature in the first The second intermediate weight in each phase :

[0065]

[0066] In the formula, This represents the lower limit of the weight, set to 0.02;

[0067] (5h) To ensure the first The sum of the weights of each feature in each time phase is 1, and the requirements for subsequent weighted fusion are satisfied. This applies to the second intermediate weight. After normalization, we obtain the first... The feature in the first The second final weight of each phase :

[0068]

[0069] (5i) According to the obtained second final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Multi-temporal fusion results of the median absolute deviation method for each feature :

[0070]

[0071] In step (6), the feature selection of vegetation index features, texture features, and color index features in the training set specifically includes the following steps:

[0072] (6a) For each type of feature, the Relief algorithm is first used to evaluate the importance of the feature and the top half of the features are retained according to their importance.

[0073] (6b) For the features retained by the Relief algorithm, the mRMR algorithm is used for further filtering, and the first half of the features after the Relief algorithm is retained;

[0074] (6c) For the features preserved by the mRMR algorithm, the SFS algorithm is used in the training set to perform stepwise forward filtering to obtain the optimal feature subset of this type of feature;

[0075] (6d) Combine the optimal feature subsets obtained from vegetation index features, texture features and color index features respectively to form combined features.

[0076] In step (7), the machine learning classification model is the XGBoost model.

[0077] Another object of the present invention is to provide an electronic device comprising:

[0078] Processor; and

[0079] The memory stores computer program instructions that, when executed by the processor, cause the processor to perform the wheat stripe rust monitoring method based on UAV multispectral and multitemporal feature weighting as described above.

[0080] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the wheat stripe rust monitoring method based on multispectral and multitemporal features of unmanned aerial vehicles as described above.

[0081] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, the present invention independently assigns weights to data from different observation periods based on features, which can adaptively determine the contribution of each observation period according to the response differences of various features at different disease development stages, thereby improving the pertinence and rationality of weight allocation and avoiding the problem that traditional uniform weighting methods cannot reflect the stage-specific differences of features; Second, the present invention synergistically integrates vegetation index, texture features, and color index, which can comprehensively characterize diseases from multiple levels such as plant physiological state, canopy spatial structure, and visible symptom manifestations, giving full play to the advantages of different types of features. The complementary effects of different features enhance the characterization ability of wheat stripe rust. Third, this invention combines a grouped feature screening method to screen and optimize different types of features separately, effectively reducing feature redundancy, noise interference, and improving the representativeness and compactness of feature subsets, thereby enhancing the stability and generalization ability of model training. Fourth, this invention does not rely on complex manual empirical rules, achieving efficient disease identification while possessing good objectivity, repeatability, and application value. It can not only be used for rapid monitoring of wheat stripe rust at the field scale but also provide technical reference for multi-temporal remote sensing identification and precise control of other crop diseases. Attached Figure Description

[0082] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0083] like Figure 1As shown, a method for monitoring wheat stripe rust based on multispectral and multi-temporal features of unmanned aerial vehicles (UAVs) is described. This method includes the following steps in sequence:

[0084] (1) Obtain UAV multispectral images of the study area in three time phases, plan sampling points and collect ground disease survey data corresponding to the sampling points, classify the ground disease survey data into healthy or infected according to the threshold, and save the classification results as labels;

[0085] (2) Preprocess the UAV multispectral image to obtain preprocessed data. The preprocessing includes radiometric correction, reflectance conversion, geometric correction and multi-temporal georegistration. Mark the location of the sampling point in the preprocessed data. The wheat image at each sampling point location forms a sample. In order to ensure the comparability between multi-temporal data, radiometric correction tools can be used to complete reflectance conversion, and geometric correction and orthorectification can be performed in combination with ground control points or aerial triangulation results. Then, the images of each observation period are unified under the same coordinate frame.

[0086] (3) Extract the vegetation index feature, texture feature and color index feature of each sample from the UAV images of the three time phases respectively. The three features of each time phase constitute a single time phase feature set. The three time phase images constitute three single time phase feature sets respectively.

[0087] (4) For the sample vectors of the same feature in each phase of the three single-phase feature sets, the weight of each feature in different phases is calculated by the entropy weight method or the median absolute deviation method, and a multi-phase fusion feature set is constructed by weighted summation; thus, it can reflect the response differences of different features in different disease development stages and improve the characterization ability of multi-phase fusion features for wheat stripe rust.

[0088] (5) Combine the multi-temporal fusion feature set with the labels to form a multi-temporal fusion dataset for wheat stripe rust;

[0089] (6) Divide the multi-temporal fusion dataset into a training set and a test set in a 7:3 ratio. In the training set, perform feature filtering on vegetation index features, texture features and color index features respectively to obtain the optimal feature subsets of each type of feature. Combine the optimal feature subsets of each type of feature to form a combined feature for model training and testing.

[0090] (8) Input the combined features into the machine learning classification model for training to obtain the trained model. Then input the test set into the trained model and output the stripe rust health or infection result of the wheat sample.

[0091] In step (1), the UAV multispectral imagery is acquired by a UAV equipped with a multispectral sensor, which includes blue, green, red, red-edge, and near-infrared bands; the three time phases are the wheat booting stage, heading stage, and flowering stage; the disease index DI is obtained based on ground disease survey data, and the wheat samples at the sampling points are divided into healthy and infected grades according to a preset threshold; the disease index DI is calculated based on the number of samples for each disease grade and the corresponding severity representative value, and the calculation formula is:

[0092]

[0093] In the formula, The disease severity is classified into 1 to 8 levels based on the proportion of stripe rust infection area to the total leaf area. The disease level of the sampled leaves is numbered as follows: Number of blades per stage; for The representative value for severity level, from arrive The percentages are 1%, 5%, 10%, 20%, 40%, 60%, 80%, and 100%, respectively. is the highest severity level of the sampled leaf in the current sample; m is the highest disease level of the actual sampled leaf in the current sample; when DI≤5%, it is judged as a healthy sample, and when DI>5%, it is judged as an infected sample, thus forming the supervision label required for subsequent classification modeling.

[0094] Step (3) specifically includes the following:

[0095] (3a) Construct a 7×7 pixel neighborhood window centered on the sampling point and extract the average reflectance of each band; the neighborhood window is preferably a 7×7 pixel window;

[0096] (3b) Calculate vegetation index characteristics based on reflectance of each band;

[0097] (3c) Based on the gray-level co-occurrence matrix, extract the texture features of the window in four directions and calculate the mean of the four directions as the final texture features; the four directions include 45°, 90°, 135° and 180°;

[0098] (3d) Color index features are calculated based on the reflectance of the red, green, and blue bands; the vegetation index features include NDVI, MSR, TVI, SIPI, EVI, RVI, PSRI, GNDVI, RG, and RDVI; the texture features include mean, variance, uniformity, contrast, variability, information entropy, second moment, and correlation; the color index features include ExG, ExR, ExGR, GLI, NGRDI, and VARI. This invention uses 56 features in total, including 10 vegetation index features, 40 texture features, and 6 color index features. Each sample possesses these 56 features at different observation periods.

[0099] In step (4), the calculation of the weights of each feature at different time phases using the entropy weight method or the median absolute deviation method specifically includes the following steps:

[0100] (4a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the impact of outliers on features... The interference of weight calculation results at different time phases is first addressed by considering the features. The sample vectors at each time phase are subjected to truncation robustness processing, compressing extreme values ​​outside the quantile interval to the corresponding quantile range, thus obtaining the truncated feature values. :

[0101]

[0102] In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature;

[0103] (4b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on subsequent entropy calculations, the truncated characteristic values ​​are... Further normalization is performed to make the same feature comparable across different time periods, resulting in a sample. In the The first phase of time The eigenvalues ​​after normalization of each feature :

[0104]

[0105] In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0;

[0106] (4c) represents the normalized eigenvalues. This is transformed into a relative proportion that can be used for information entropy calculation. For each time phase, it is calculated according to the... Calculate the probability value corresponding to each sample by using the normalized feature values ​​of each feature across all samples. :

[0107]

[0108] In the formula, Indicates the total number of samples;

[0109] (4d) To measure the dispersion and information richness of the same feature across different time phases, based on the samples in the first time phase... Calculate the information entropy value corresponding to each time point based on the probability value at each time point. :

[0110]

[0111] In the formula, The entropy value is the standardized coefficient. This is used to constrain the entropy value to a comparable range. The larger the entropy value, the more uniform the distribution of this feature among the samples in that time phase, and the weaker the difference. Conversely, it indicates that the time phase may contain stronger discriminative information.

[0112] (4e) In order to further transform the differences reflected by information entropy into an evaluation quantity that can be used for weighting, the information entropy values ​​of each time phase are... Calculate the information utility value separately :

[0113]

[0114] In the formula, The larger the value, the higher the value. The first phase of time The stronger the ability of a feature to distinguish differences between samples;

[0115] (4f) To map the information utility values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the first... The feature in the first The first initial weight at each time phase :

[0116] In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution. The smaller the value, the easier it is for the weight to concentrate on the phases with higher scores, and vice versa.

[0117] (4g) To avoid the first In a certain time phase, a feature receives too little weight due to a low score, thus weakening the complementary effect of multi-time phase information. Therefore, a preset lower limit constraint is applied to the initial weights to obtain the [previous] feature. The feature in the first The first intermediate weight in each phase :

[0118]

[0119] In the formula, This represents the lower limit of the weight, set to 0.02, to prevent the weight from being lowered by the first weight. The weight of a feature is too small at a certain time;

[0120] (4h) To ensure the first The sum of the weights of each feature across the three time phases is 1. Further normalization is applied to the intermediate weights to obtain the result. The feature in the first The first final weight of each phase :

[0121]

[0122] (4i) According to the first final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Entropy weighting method for multi-temporal fusion results of individual features :

[0123]

[0124] The multi-temporal fusion results corresponding to each feature are summarized to construct a multi-temporal fusion feature set.

[0125] In step (4), the calculation of the weights of each feature at different time phases using the median absolute deviation method specifically includes the following steps:

[0126] (5a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the interference of outliers on the calculation results of the same feature's dispersion at different time phases, firstly, for the... The sample vectors of each feature at each time phase are truncated and robustly processed to limit extreme values ​​outside the quantile interval to a preset range, resulting in truncated feature values. :

[0127] In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature;

[0128] (5b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on the dispersion evaluation, the truncated eigenvalues ​​are normalized to obtain the normalized eigenvalues. :

[0129]

[0130] In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0;

[0131] (5c) To obtain the robust central location of this feature in the sample population at each time phase, calculate the first... The first phase of time The median of each feature in all samples :

[0132] In the formula, Indicates the first The first phase of time The median of each feature across all samples;

[0133] (5d) To measure the degree of deviation of each sample feature value from the median, the absolute deviation is calculated. :

[0134]

[0135] (5e) To further obtain the robustness of this feature across different time phases, the absolute deviation is... Take the median to get the first... The corresponding absolute deviation of the median at each time point :

[0136]

[0137] In the formula, The larger it is, the more likely it is to be the first At this time, the first The more significant the differences between samples, the stronger the discriminative information that a feature may contain.

[0138] (5f) To map the median absolute deviation values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the th... The feature in the first The second initial weights obtained based on MAD score mapping at each time phase :

[0139]

[0140] In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution;

[0141] (5g) To avoid a certain time phase receiving too low a weight due to its small dispersion, thus affecting the comprehensive utilization of multi-time phase information, a preset lower limit constraint on the weight is applied based on the initial weights to obtain the first... The feature in the first The second intermediate weight in each phase :

[0142]

[0143] In the formula, This represents the lower limit of the weight, set to 0.02;

[0144] (5h) To ensure the first The sum of the weights of each feature in each time phase is 1, and the requirements for subsequent weighted fusion are satisfied. This applies to the second intermediate weight. After normalization, we obtain the first... The feature in the first The second final weight of each phase :

[0145]

[0146] (5i) According to the obtained second final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Multi-temporal fusion results of the median absolute deviation method for each feature :

[0147]

[0148] The multi-temporal fusion results corresponding to each feature are summarized to construct a multi-temporal fusion feature set.

[0149] In step (6), the feature selection of vegetation index features, texture features, and color index features in the training set specifically includes the following steps:

[0150] (6a) For each type of feature, the Relief algorithm is first used to evaluate the importance of the feature and the top half of the features are retained according to their importance.

[0151] (6b) For the features retained by the Relief algorithm, the mRMR algorithm is used for further filtering, and the first half of the features after the Relief algorithm is retained;

[0152] (6c) For the features preserved by the mRMR algorithm, the SFS algorithm is used in the training set to perform stepwise forward filtering to obtain the optimal feature subset of this type of feature;

[0153] (6d) Combine the optimal feature subsets obtained from vegetation index features, texture features and color index features respectively to form combined features.

[0154] In step (7), the machine learning classification model is the XGBoost model.

[0155] In summary, this invention independently weights data from different observation periods based on features, adaptively determining the contribution of each observation period according to the response differences of various features at different disease development stages. This improves the pertinence and rationality of weight allocation, avoiding the problem that traditional uniform weighting methods cannot reflect the stage-specific differences of features. This invention synergistically integrates vegetation indices, texture features, and color indices, enabling a comprehensive characterization of diseases from multiple levels, including plant physiological state, canopy spatial structure, and visible symptom representation. It fully leverages the complementary effects between different types of features, enhancing the characterization ability of wheat stripe rust. This invention combines a grouped feature screening method to separately screen and optimize different types of features, effectively reducing feature redundancy, minimizing noise interference, and improving the representativeness and compactness of feature subsets, thereby enhancing the stability and generalization ability of model training. This invention does not rely on complex manual empirical rules, achieving efficient disease identification while possessing good objectivity, repeatability, and application value. It can not only be used for rapid monitoring of wheat stripe rust at the field scale but also provide technical reference for multi-temporal remote sensing identification and precise control of other crop diseases.

[0156] The foregoing has shown and described 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 embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for monitoring wheat stripe rust based on multispectral and multi-temporal characteristics of unmanned aerial vehicles (UAVs), characterized in that: The method includes the following steps in sequence: (1) Obtain UAV multispectral images of the study area in three time phases, plan sampling points and collect ground disease survey data corresponding to the sampling points, classify the ground disease survey data into healthy or infected according to the threshold, and save the classification results as labels; (2) Preprocess the UAV multispectral image to obtain preprocessed data. The preprocessing includes radiometric correction, reflectivity conversion, geometric correction and multi-temporal georegistration. Mark the location of the sampling point in the preprocessed data. The wheat image at each sampling point location forms a sample. (3) Extract the vegetation index feature, texture feature and color index feature of each sample from the UAV images of the three time phases respectively. The three features of each time phase constitute a single time phase feature set. The three time phase images constitute three single time phase feature sets respectively. (4) For the sample vectors of the same feature in each phase of the three single-phase feature sets, the weight of each feature in different phases is calculated by the entropy weight method or the median absolute deviation method, and a multi-phase fusion feature set is constructed by weighted summation; (5) Combine the multi-temporal fusion feature set with the labels to form a multi-temporal fusion dataset for wheat stripe rust; (6) Divide the multi-temporal fusion dataset into a training set and a test set in a 7:3 ratio. In the training set, perform feature filtering on vegetation index features, texture features and color index features respectively to obtain the optimal feature subsets of each type of feature. Combine the optimal feature subsets of each type of feature to form a combined feature for model training and testing. (7) Input the combined features into the machine learning classification model for training to obtain the trained model. Then input the test set into the trained model and output the stripe rust health or infection result of the wheat sample.

2. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: In step (1), the UAV multispectral imagery is acquired by a UAV equipped with a multispectral sensor, which includes blue, green, red, red-edge, and near-infrared bands; the three time phases are the wheat booting stage, heading stage, and flowering stage; the disease index DI is obtained based on ground disease survey data, and the wheat samples at the sampling points are divided into healthy and infected grades according to a preset threshold; the disease index DI is calculated based on the number of samples for each disease grade and the corresponding severity representative value, and the calculation formula is: ； In the formula, The disease severity is classified into 1 to 8 levels based on the proportion of stripe rust infection area to the total leaf area. The disease level of the sampled leaves is numbered as follows: Number of blades per stage; for The representative value for severity level, from arrive They are 1%, 5%, 10%, 20%, 40%, 60%, 80%, and 100%, respectively. is the highest severity level of the sampled leaf in the current sample; m is the highest disease level of the actual sampled leaf in the current sample; when DI≤5%, it is judged as a healthy sample, and when DI>5%, it is judged as an infected sample, thus forming the supervision label required for subsequent classification modeling.

3. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: Step (3) specifically includes the following: (3a) Construct a 7×7 pixel neighborhood window centered on the sampling point and extract the average reflectance of each band; the neighborhood window is preferably a 7×7 pixel window; (3b) Calculate vegetation index characteristics based on reflectance of each band; (3c) Based on the gray-level co-occurrence matrix, extract the texture features of the window in four directions and calculate the mean of the four directions as the final texture features; the four directions include 45°, 90°, 135° and 180°; (3d) Calculate color index features based on the reflectance of red, green and blue bands; the vegetation index features include NDVI, MSR, TVI, SIPI, EVI, RVI, PSRI, GNDVI, RG and RDVI; the texture features include mean, variance, uniformity, contrast, variability, information entropy, second moment and correlation; the color index features include ExG, ExR, ExGR, GLI, NGRDI and VARI.

4. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: In step (4), the calculation of the weights of each feature at different time phases using the entropy weight method or the median absolute deviation method specifically includes the following steps: (4a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the impact of outliers on features... The interference of weight calculation results at different time phases is first addressed by considering the features. The sample vectors at each time phase are subjected to truncation robustness processing, compressing extreme values ​​outside the quantile interval to the corresponding quantile range, thus obtaining the truncated feature values. : ； In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature; (4b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on subsequent entropy calculations, the truncated characteristic values ​​are... Further normalization is performed to make the same feature comparable across different time periods, resulting in a sample. In the The first phase of time The eigenvalues ​​after normalization of each feature : ； In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0; (4c) represents the normalized eigenvalues. This is transformed into a relative proportion that can be used for information entropy calculation. For each time phase, it is calculated according to the... Calculate the probability value corresponding to each sample by using the normalized feature values ​​of each feature across all samples. : ； In the formula, Indicates the total number of samples; (4d) To measure the dispersion and information richness of the same feature across different time phases, based on the samples in the first time phase... Calculate the information entropy value corresponding to each time point based on the probability value at each time point. : ； In the formula, The entropy value is the standardized coefficient. This is used to constrain entropy values ​​to a comparable range; (4e) In order to further transform the differences reflected by information entropy into an evaluation quantity that can be used for weighting, the information entropy values ​​of each time phase are... Calculate the information utility value separately : ； In the formula, The larger the value, the higher the value. The first phase of time The stronger the ability of a feature to distinguish differences between samples; (4f) To map the information utility values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the first... The feature in the first The first initial weight at each time phase : ; In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution. The smaller the value, the easier it is for the weight to concentrate on the phases with higher scores, and vice versa. (4g) To avoid the first In a certain time phase, a feature receives too little weight due to a low score, thus weakening the complementary effect of multi-time phase information. Therefore, a preset lower limit constraint is applied to the initial weights to obtain the [previous] feature. The feature in the first The first intermediate weight in each phase : ； In the formula, This represents the lower limit of the weight, set to 0.02, to prevent the weight from being lowered by the first... The weight of a feature is too small at a certain time; (4h) To ensure the first The sum of the weights of each feature across the three time phases is 1. Further normalization is applied to the intermediate weights to obtain the result. The feature in the first The first final weight of each phase : ； (4i) According to the first final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Entropy weighting method for multi-temporal fusion results of individual features : 。 5. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: In step (4), the calculation of the weights of each feature at different time phases using the median absolute deviation method specifically includes the following steps: (5a) Let Indicates the feature number, Indicates the phase number, Indicates the sample number; for any feature Based on features The sample vectors at each time phase are calculated as follows: To reduce the interference of outliers on the calculation results of the same feature's dispersion at different time phases, firstly, for the... The sample vectors of each feature at each time phase are truncated and robustly processed to limit extreme values ​​outside the quantile interval to a preset range, resulting in truncated feature values. : ; In the formula, Indicates sample In the The first phase of time The original feature values ​​of each feature; and They represent the first At the given time phase, the first of all samples The 1st percentile and 99th percentile of each feature; (5b) To eliminate the influence of differences in characteristic dimensions and value ranges between different time phases on the dispersion evaluation, the truncated eigenvalues ​​are normalized to obtain the normalized eigenvalues. : ； In the formula, and They represent the first The first phase of time The maximum and minimum values ​​of each feature in all samples; It is a very small positive number, used to avoid the denominator being 0; (5c) To obtain the robust central location of this feature in the sample population at each time phase, calculate the first... The first phase of time The median of each feature in all samples : ; In the formula, Indicates the first The first phase of time The median of each feature across all samples; (5d) To measure the degree of deviation of each sample feature value from the median, the absolute deviation is calculated. : ； (5e) To further obtain the robustness of this feature across different time phases, the absolute deviation is... Take the median to get the first... The corresponding absolute deviation of the median at each time point : ； In the formula, The larger it is, the more likely it is to be the first At this time, the first The more significant the differences between samples, the stronger the discriminative information that a feature may contain. (5f) To map the median absolute deviation values ​​of each time phase to non-negative and comparable initial weights, a softmax function is used for transformation to obtain the th... The feature in the first The second initial weights obtained based on MAD score mapping at each time phase : ； In the formula, This represents the temperature parameter, set to 0.35, used to adjust the smoothness of the weight distribution; (5g) To avoid a certain time phase receiving too low a weight due to its small dispersion, thus affecting the comprehensive utilization of multi-time phase information, a preset weight lower limit constraint is applied to the initial weights to obtain the first... The feature in the first The second intermediate weight in each phase : ； In the formula, This represents the lower limit of the weight, set to 0.02; (5h) To ensure the first The sum of the weights of each feature in each time phase is 1, and the requirements for subsequent weighted fusion are satisfied. This applies to the second intermediate weight. After normalization, we obtain the first... The feature in the first The second final weight of each phase : ； (5i) According to the obtained second final weight A weighted summation is performed on the values ​​of the same feature at different time phases to obtain the sample. The Multi-temporal fusion results of the median absolute bias method for each feature : 。 6. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: In step (6), the feature selection of vegetation index features, texture features, and color index features in the training set specifically includes the following steps: (6a) For each type of feature, the Relief algorithm is first used to evaluate the importance of the feature and the top half of the features are retained according to their importance. (6b) For the features retained by the Relief algorithm, the mRMR algorithm is used for further filtering, and the first half of the features after the Relief algorithm is retained; (6c) For the features preserved by the mRMR algorithm, the SFS algorithm is used in the training set to perform stepwise forward filtering to obtain the optimal feature subset of this type of feature; (6d) Combine the optimal feature subsets obtained from vegetation index features, texture features and color index features respectively to form combined features.

7. The wheat stripe rust monitoring method based on UAV multispectral and multi-temporal feature weighting according to claim 1, characterized in that: In step (7), the machine learning classification model is the XGBoost model.

8. An electronic device, comprising: processor; as well as A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the wheat stripe rust monitoring method based on UAV multispectral and multitemporal feature weighting as described in any one of claims 1-7.

9. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the wheat stripe rust monitoring method based on UAV multispectral and multitemporal feature weighting as described in any one of claims 1-7.

Images