Mikania micrantha and fragrant eupatorium herb classification and identification method based on hyperspectrum of unmanned aerial vehicle
By using UAV hyperspectral technology and the deep learning model FD-U-Net, the problem of distinguishing between Mikania micrantha and Agrostis pilosa has been solved, achieving high-precision and automated classification and identification, improving identification efficiency and accuracy, and supporting precise prevention and control.
Patent Information
- Application Number
- CN202511625002.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-02-13
AI Technical Summary
Existing technologies are insufficient to efficiently and accurately distinguish and monitor invasive plants such as Mikania micrantha and Clerodendrum trichotomum, which are highly similar in morphology, flowering period, and habitat. Traditional monitoring methods are inefficient and prone to misjudgment, while satellite remote sensing is not accurate enough to meet the needs of precise prevention and control.
Using UAV hyperspectral technology, through preprocessing, spectral transformation and dimensionality reduction, multi-type classification models are constructed, especially the FD-U-Net model, to achieve high-precision classification and recognition of Mikania micrantha and Agrostis pilosa, and output the classification results.
It achieves high-precision automated differentiation between Mikania micrantha and Clerodendrum trichotomum, with an overall classification accuracy of over 90% and a Kappa coefficient as high as 0.78. This significantly improves identification efficiency and accuracy, reduces subjective errors from manual surveys, and provides accurate distribution maps to guide management.
Smart Images
Figure CN121527652A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of ecological environment monitoring, and particularly relates to a classification and identification method for Mikania micrantha and Chromolaena odorata based on unmanned aerial vehicle hyperspectral. BACKGROUND
[0002] Invasive alien species has become a global ecological and economic problem, which seriously threatens biodiversity, ecosystem stability and the safety of agricultural and forestry production. Mikania micrantha and Chromolaena odorata, as the malignant invasive plants listed in the "List of Key Management Alien Invasive Species", have extremely strong reproductive capacity and environmental adaptability, and are rapidly spreading and diffusing in many regions of Yunnan and other places in China. Through occupying the survival space of native vegetation, competing for nutrients and light, the two invasive plants lead to the degradation of native vegetation and the reduction of crop yield, causing great ecological damage and economic loss.
[0003] However, the niches of Mikania micrantha and Chromolaena odorata are highly coincident, the flowering period is concentrated from November to February of the next year, the flower morphology and plant structure have high similarity, and they often coexist, making it difficult for traditional monitoring methods to effectively distinguish between the two. The existing monitoring technology mainly relies on manual visual investigation, which requires a large amount of manpower and material resources to carry out survey operations block by block, not only has the problems of low efficiency and time-consuming and laborious, but also is affected by subjective factors such as the professional experience of the investigators, and is prone to misjudgment and omission. The ordinary satellite remote sensing technology is limited by spatial resolution, and the monitoring accuracy of small-scale or early-stage invasive patches is insufficient, which cannot meet the actual needs of precise prevention and control.
[0004] In recent years, unmanned aerial vehicle (UAV) remote sensing technology has been widely applied in vegetation monitoring due to its significant advantages of mobility, controllable cost and high spatial resolution. Especially the UAV system equipped with a hyperspectral imager can obtain continuous spectral reflectance information of ground objects in narrow bands, forming high-dimensional data of "graph and spectrum integration", which can capture the subtle differences in plant spectra that cannot be distinguished by the human eye, providing core technical support for the precise differentiation of similar vegetation types.
[0005] At present, the existing technology mainly uses UAV remote sensing to identify single invasive species or distinguish target species from background objects, such as the monitoring of Spartina alterniflora and Acanthospermum hispidum, which relies on the significant spectral differences between the target species and the surrounding vegetation to construct an identification model. However, for Mikania micrantha and Chromolaena odorata, which have similar morphology, flowering period and habitat, there is no public and special hyperspectral precise classification and identification technology scheme, which makes it difficult to meet the actual application needs of precise differentiation and large-scale dynamic monitoring of the two species. Therefore, how to realize the rapid, precise and efficient classification and identification of Mikania micrantha and Chromolaena odorata has become a technical problem that needs to be solved in the current invasive species prevention and control field. SUMMARY
[0006] The application provides a method for classifying and identifying gueldenstaedtia pekinensis and eriochloa villosa based on unmanned aerial vehicle hyperspectral, aiming to solve at least one of the above technical problems.
[0007] To achieve the above-mentioned purpose, the application provides a method for classifying and identifying gueldenstaedtia pekinensis and eriochloa villosa based on unmanned aerial vehicle hyperspectral, which comprises the following steps:
[0008] S1: obtaining unmanned aerial vehicle hyperspectral original image data of a target area;
[0009] S2: performing a preprocessing operation on the unmanned aerial vehicle hyperspectral original image data to obtain a standardized hyperspectral data set; wherein the preprocessing operation comprises radiation correction processing, geometric correction processing, image registration processing and remote sensing image inlay processing;
[0010] S3: performing spectral transformation and dimensionality reduction processing on the standardized hyperspectral data set to obtain a plurality of groups of optimized feature data sets formed by different combinations of spectral transformation and dimensionality reduction processing;
[0011] S4: constructing a multi-type classification model based on the plurality of groups of optimized feature data sets, screening out an optimal classification model through model training and performance evaluation, and realizing classification and identification of gueldenstaedtia pekinensis and eriochloa villosa in the target area by using the optimal classification model, and outputting a classification result.
[0012] Optionally, in step S1, the unmanned aerial vehicle hyperspectral original image data is collected by selecting a multi-rotor unmanned aerial vehicle equipped with a hyperspectral imager, and performing hyperspectral original image data collection of the target area under the conditions of preset flight parameters, flight time period and flight path.
[0013] Optionally, in step S3, the spectral transformation and dimensionality reduction processing is performed on the standardized hyperspectral data set to obtain a plurality of groups of optimized feature data sets formed by different combinations of spectral transformation and dimensionality reduction processing, specifically comprising:
[0014] S31: reading the standardized hyperspectral data set and extracting the spectral vector of each pixel , wherein m is the number of hyperspectral bands, and the missing pixels are filled by using the neighborhood mean filling method;
[0015] S32: using three independent spectral transformation methods to enhance the spectral differentiation characteristics of gueldenstaedtia pekinensis and eriochloa villosa and the dimensionality reduction processing of principal component analysis, combining the three spectral transformation methods with the dimensionality reduction of principal component analysis to form a plurality of combined optimized feature data sets;
[0016] S33: Save the optimized feature datasets of several combinations as ENVI format, and annotate the ENVI format files with metadata including the combination processing method, the number of principal components, and the cumulative contribution rate, and wait for them to be called when building the subsequent classification model.
[0017] Optionally, in step S32, three independent spectral transformation methods enhance the spectral distinguishing features between Mikania micrantha and Clerodendrum trichotomum, specifically including: first derivative processing, Gaussian smoothing, and standard normal variable transformation, with the following expressions:
[0018] ;
[0019] ;
[0020] ;
[0021] ;
[0022] ;
[0023] In the formula, For the first The center wavelength of each band The band spacing is ≤5nm. wavelength The corresponding reflectance value, For the first The first derivative of the center wavelength of each band The standard deviation of the Gaussian kernel ranges from 0.5 to 2.0 and is adaptively adjusted according to the spectral noise level. To smooth the center wavelength, The size of the convolution window (5-7 bands). This is the smoothed reflectance value. For the first Reflectance values for each band, The mean of the spectral vector. denoted as the standard deviation of the spectral vector.
[0024] Optionally, in step S32, the dimensionality reduction processing of principal component analysis specifically includes:
[0025] S321: Calculate the covariance matrix of the spectral data The expression is as follows:
[0026] ;
[0027] in, Number of pixel samples for 3D spectral data matrix For a spectrum mean vector;
[0028] S322: solve the eigenvalue and the corresponding eigenvector of the covariance matrix ;
[0029] S323: according to the cumulative contribution rate, the first principal components are selected, and the principal component projection matrix is constructed ;
[0030] S324: project the transformed spectral data into the principal component space to obtain the reduced dimension data set , wherein is a dimensional spectral data matrix, is a dimensional reduced dimension data set, and .
[0031] Optionally, step S4: based on the multiple sets of optimized feature data sets, a multi-type classification model is constructed, the optimal classification model is selected through model training and performance evaluation, the optimal classification model is used to realize the classification and identification of target area A. ganoderma and aircraft grass, and the classification result is output. Specifically, it includes:
[0032] S41: based on the standardized hyperspectral data set, a plurality of sample data of the optimized feature data set for each combination is selected, and based on the ROI region label and the optimized feature data set corresponding to each sample data, the training sample data set of each combination of the optimized feature data set is constructed;
[0033] S42: the support vector machine model, the random forest model and the U-Net model are combined with the training sample data set of each combination of the optimized feature data set to obtain a plurality of classification model combinations. The training sample data set of the optimized feature data set for each combination in each classification model combination is used to train and evaluate the performance of the classification model, and the optimal classification model combination is selected according to the performance evaluation result;
[0034] S43: based on the optimal classification model combination and the standardized hyperspectral data set, the classification and identification of target area A. ganoderma and aircraft grass are performed, and the classification result is output.
[0035] Optionally, in step S41, based on the ROI region label and the optimized feature data set corresponding to each sample data, the training sample data set of each combination of the optimized feature data set is constructed, specifically including:
[0036] S411: The ROI regions of gueldenstaedtia pekinensis, erigeron bonariensis and other ground objects are manually drawn on each sample data as ROI region labels in advance; wherein, the gueldenstaedtia pekinensis label is 0, the erigeron bonariensis label is 1, and the other ground object label is 2;
[0037] S412: The pixel feature vectors of each combination of the optimized feature data set corresponding to each ROI region are extracted, and a training sample data set of each combination of the optimized feature data set composed of the pixel feature vectors and the ROI region labels is constructed.
[0038] Optionally, in step S42, the support vector machine model, the random forest model and the U-Net model are constructed, and specifically comprising:
[0039] S421: A radial basis function is used as a kernel function, a penalty coefficient C=1.0 is set, a kernel function coefficient gamma='scale' is set, a maximum iteration number max_iter=-1 is set, a decision function shape decision_function_shape='ovr' is set, break_ties=False, random_state=None, and a support vector machine is constructed;
[0040] S422: The number of decision trees n_estimators=100 is set, the maximum depth of a single decision tree max_depth=256 is set, the minimum number of samples required for internal node splitting min_samples_split=2 is set, the minimum number of samples of a leaf node min_samples_leaf=1 is set, the maximum number of features max_features='auto' is set, the maximum number of leaf nodes max_leaf_nodes=None is set, and a random forest model is constructed;
[0041] S423: A U-Net network structure containing an encoder and a decoder is constructed, the network input is set as an image block containing the first N main principal component bands, the encoder uses multiple convolutional layers and maximum pooling layers to gradually extract multi-level features, the decoder fuses shallow details and deep semantic information through upsampling and skip connection, the network output end uses a convolutional layer with a Softmax activation function for pixel classification, the total number of model training rounds is set to 200, and the optimizer uses stochastic gradient descent to obtain the final U-Net model.
[0042] Optionally, in step S42, the model training and performance evaluation are performed, and the optimal classification model combination is selected according to the performance evaluation result, specifically comprising:
[0043] S424: Divide the training sample data set into a training set and a validation set according to a preset proportion, train the classification model in the classification model combination using the training set, and verify the classification model using the validation set;
[0044] S425: According to the overall classification accuracy OA, the Kappa coefficient, the producer accuracy PA, the user accuracy UA and the F1-Score obtained by verification calculation, select the optimal classification model combination from the several classification model combinations.
[0045] Optionally, S43: based on the optimal classification model combination and the standardized hyperspectral data set, performing classification and identification of target area Ageratina adenophora and Bidens pilosa, and outputting a classification result, specifically comprising:
[0046] S431: From the plurality of sets of optimized feature data sets of the standardized hyperspectral data set, extract the optimized feature data set belonging to the target set in the optimal classification model combination;
[0047] S432: Transfer the optimized feature data set to the classification model belonging to the optimal classification model combination for classification prediction, and generate an Ageratina adenophora and Bidens pilosa spatial distribution classification map of the standardized hyperspectral data set.
[0048] The beneficial effects of the present application are:
[0049] (1) High recognition accuracy, especially significant enhancement of the ability to distinguish similar species: the present application uses the subtle differences of hyperspectral data and combines the optimized FD-U-Net model to fundamentally solve the problem that traditional methods cannot distinguish Ageratina adenophora and Bidens pilosa. The test data shows that the overall classification accuracy (OA) of the preferred scheme of the present application is more than 90%, and the Kappa coefficient is as high as 0.78 or more. In particular, the recognition F1 score of the easily confused species Bidens pilosa reaches an extremely high level of 0.90, which is significantly better than the comparative scheme (such as the F1 score of about 0.55-0.65 of traditional machine learning methods such as SVM and RF), and realizes high-precision automatic differentiation.
[0050] (2) Efficiency and accuracy are improved: compared with traditional manual field investigation, the present application adopts unmanned aerial vehicle automatic operation to quickly obtain large-area data. At the same time, compared with traditional machine learning methods (such as SVM and RF) which need a lot of tuning and have a lower upper limit of performance, the deep learning model (U-Net) used in the present application can process the entire image at one time after training, has high inference efficiency, and the final accuracy has achieved a qualitative leap, realizing the dual improvement of efficiency and accuracy.
[0051] (3) Strong objectivity, stable and reliable results: the entire process is based on data-driven, through standardized data processing and deep learning algorithm for classification, excluding subjective error of artificial investigation. Multiple experimental results show that the preferred FD-U-Net scheme of the present application shows stable and excellent performance on different data subsets, ensuring the objectivity, repeatability and reliability of the results.
[0052] (4) Strong pertinence, direct empowerment and precise prevention and control: the present application is the first complete technical solution specially designed for the specific and difficult problem of distinguishing between fleabane and aircraft grass. The high-precision distribution map produced can clearly define the specific distribution range of the two invasive plants, directly guide the ground personnel to carry out "point-to-point" precise removal work, effectively reduce the management cost, improve the prevention and control effect, and has very high application value. BRIEF DESCRIPTION OF DRAWINGS
[0053] Figure 1 It is a flowchart of the classification and identification method of fleabane and aircraft grass based on unmanned aerial vehicle hyperspectral.
[0054] Figure 2 It is a schematic diagram of hyperspectral image after data preprocessing.
[0055] Figure 3 It is a schematic diagram of spectral data of the original image.
[0056] Figure 4 It is a schematic diagram of spectral data after Gaussian smoothing processing.
[0057] Figure 5 It is a schematic diagram of spectral data after first derivative processing.
[0058] Figure 6 It is a schematic diagram of spectral data after standard normal variable change processing.
[0059] Figure 7 It is a schematic diagram of support vector machine classification results based on different preprocessing methods.
[0060] Figure 8 It is a schematic diagram of random forest classification results based on different preprocessing methods.
[0061] Figure 9 It is a schematic diagram of U-Net classification results based on different preprocessing methods.
[0062] Figure 10 It is a confusion matrix diagram of different processing methods. DETAILED DESCRIPTION
[0063] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.
[0064] The embodiment of the present application provides a classification and identification method for gueldenstaedtia pekinensis and eriochloa villosa based on unmanned aerial vehicle hyperspectrum. Figure 1 , Figure 1 The embodiment of the present application provides a classification and identification method for gueldenstaedtia pekinensis and eriochloa villosa based on unmanned aerial vehicle hyperspectrum.
[0065] In the embodiment, a classification and identification method for gueldenstaedtia pekinensis and eriochloa villosa based on unmanned aerial vehicle hyperspectrum comprises the following steps:
[0066] S1: acquiring unmanned aerial vehicle hyperspectrum original image data of a target area;
[0067] S2: performing a preprocessing operation on the unmanned aerial vehicle hyperspectrum original image data to obtain a standardized hyperspectrum data set; wherein the preprocessing operation comprises radiation correction processing, geometric correction processing, image registration processing and remote sensing image inlay processing;
[0068] S3: performing spectral transformation and dimension reduction processing on the standardized hyperspectrum data set to obtain a plurality of groups of optimized feature data sets formed by different combinations of spectral transformation and dimension reduction processing;
[0069] S4: constructing a plurality of types of classification models based on the plurality of groups of optimized feature data sets, screening out an optimal classification model through model training and performance evaluation, realizing classification and identification of gueldenstaedtia pekinensis and eriochloa villosa in the target area by using the optimal classification model, and outputting a classification result.
[0070] In the preferred embodiment, in step S1, the unmanned aerial vehicle hyperspectrum original image data is collected by selecting a multi-rotor unmanned aerial vehicle carrying a hyperspectrum imager to perform hyperspectrum original image data collection of the target area under the conditions of preset flight parameters, flight time period and flight path.
[0071] In the preferred embodiment, step S3: performing spectral transformation and dimension reduction processing on the standardized hyperspectrum data set to obtain a plurality of groups of optimized feature data sets formed by different combinations of spectral transformation and dimension reduction processing, specifically comprises:
[0072] S31: reading the standardized hyperspectrum data set and extracting a spectral vector of each pixel , wherein m is the number of hyperspectrum bands, and the missing pixels are filled by using a neighborhood mean filling method;
[0073] S32: Three independent spectral transformation methods are used to enhance the spectral distinguishing features of Mikania micrantha and Agrostis spp. as well as the dimensionality reduction processing of principal component analysis. The three spectral transformation methods are combined with principal component analysis dimensionality reduction to form several optimized feature datasets.
[0074] S33: Save the optimized feature datasets of several combinations as ENVI format, and annotate the ENVI format files with metadata including the combination processing method, the number of principal components, and the cumulative contribution rate, and wait for them to be called when building the subsequent classification model.
[0075] In a preferred embodiment, in step S32, three independent spectral transformation methods enhance the spectral distinguishing features between Mikania micrantha and Clerodendrum trichotomum, specifically including: first derivative processing, Gaussian smoothing, and standard normal variable transformation, the expressions of which are as follows:
[0076] ;
[0077] ;
[0078] ;
[0079] ;
[0080] ;
[0081] In the formula, For the first The center wavelength of each band The band spacing is ≤5nm. wavelength The corresponding reflectance value, For the first The first derivative of the center wavelength of each band The standard deviation of the Gaussian kernel ranges from 0.5 to 2.0 and is adaptively adjusted according to the spectral noise level. To smooth the center wavelength, The size of the convolution window (5-7 bands). This is the smoothed reflectance value. For the first Reflectance values for each band, The mean of the spectral vector. denoted as the standard deviation of the spectral vector.
[0082] In a preferred embodiment, step S32, the dimensionality reduction processing of principal component analysis, specifically includes:
[0083] S321: Calculate the covariance matrix of the spectral data The expression is as follows:
[0084] ;
[0085] wherein, is the number of pixel samples, is a matrix of spectral data of dimension is a vector of spectral mean values of dimension
[0086] S322: solve the eigenvalues and the corresponding eigenvectors of the covariance matrix ;
[0087] S323: according to the cumulative contribution rate, screen the principal components, select the first principal components, and construct a principal component projection matrix ;
[0088] S324: project the transformed spectral data into the principal component space to obtain a reduced dimension data set , wherein is a matrix of spectral data of dimension is a reduced dimension data set of dimension .
[0089] In the preferred embodiment, step S4: based on the multiple sets of optimized feature data sets, construct multiple types of classification models, select the optimal classification model through model training and performance evaluation, use the optimal classification model to realize the classification and identification of the target area of A. japonicus and E. japonica, and output the classification result, specifically including:
[0090] S41: based on the standardized hyperspectral data set, select a plurality of sample data of the optimized feature data set for each combination, and based on the ROI region label and the optimized feature data set corresponding to each sample data, construct a training sample data set of the optimized feature data set of each combination;
[0091] S42: construct a support vector machine model, a random forest model and a U-Net model, and combine each combination of the training sample data set of the optimized feature data set to obtain a plurality of classification model combinations, use the training sample data set of the optimized feature data set for each combination in each classification model combination to train and evaluate the performance of the classification model, and select the optimal classification model combination according to the performance evaluation result;
[0092] S43: based on the optimal classification model combination and the standardized hyperspectral data set, perform classification and identification of the target area of A. japonicus and E. japonica, and output the classification result.
[0093] In the preferred embodiment, in step S41, based on the ROI region label corresponding to each sample data and the optimization feature data set, a training sample data set of each combined optimization feature data set is constructed, specifically comprising:
[0094] S411: The ROI regions of eucalyptus, aircraft grass and other ground objects are manually drawn on each sample data as ROI region labels by means of artificial ROI drawing; wherein the eucalyptus label is 0, the aircraft grass label is 1, and the other ground object label is 2 in the ROI region label;
[0095] S412: The pixel feature vector of each combined optimization feature data set corresponding to each ROI region is extracted, and a training sample data set of each combined optimization feature data set composed of pixel feature vectors and ROI region labels is constructed.
[0096] In the preferred embodiment, in step S42, a support vector machine model, a random forest model and a U-Net model are constructed, specifically comprising:
[0097] S421: A radial basis function is used as a kernel function, a penalty coefficient C=1.0 is set, a kernel function coefficient gamma='scale' is set, a maximum iteration number max_iter=-1 is set, a decision function shape decision_function_shape='ovr' is set, break_ties=False, random_state=None, and a support vector machine is constructed;
[0098] S422: The number of decision trees n_estimators=100 is set, the maximum depth of a single decision tree max_depth=256 is set, the minimum number of samples required for internal node splitting min_samples_split=2 is set, the minimum number of leaf nodes min_samples_leaf=1 is set, the maximum number of features max_features='auto' is set, the maximum number of leaf nodes max_leaf_nodes=None is set, and a random forest model is constructed;
[0099] S423: A U-Net network structure containing an encoder and a decoder is constructed, the network input is set to an image block containing the first N main principal component bands, the encoder uses multiple convolutional layers and max-pooling layers to gradually extract multi-level features, the decoder fuses shallow details and deep semantic information through upsampling and skip connection, the network output end uses a convolutional layer with a Softmax activation function for pixel classification, the total number of model training rounds is set to 200, the optimizer uses stochastic gradient descent, and the final U-Net model is obtained.
[0100] In a preferred embodiment, in step S42, model training and performance evaluation are performed, and the optimal classification model combination is selected according to the performance evaluation result, specifically including:
[0101] S424: dividing the training sample data set into a training set and a validation set according to a preset ratio, training the classification model in the classification model combination using the training set, and verifying the classification model using the validation set;
[0102] S425: selecting the optimal classification model combination from the several classification model combinations according to the overall classification accuracy OA, the Kappa coefficient, the producer accuracy PA, the user accuracy UA and the F1-Score obtained by verification calculation.
[0103] In a preferred embodiment, S43: based on the optimal classification model combination and the standardized hyperspectral data set, performing classification and identification of target area Eupatorium odoratum and Bidens pilosa, and outputting a classification result, specifically including:
[0104] S431: extracting an optimized feature data set belonging to a target combination in the optimal classification model combination from the multiple groups of optimized feature data sets of the standardized hyperspectral data set;
[0105] S432: transmitting the optimized feature data set to the classification model belonging to the optimal classification model combination for classification prediction, and generating a Eupatorium odoratum and Bidens pilosa spatial distribution classification map of the standardized hyperspectral data set.
[0106] It should be noted that the prior art mainly uses unmanned aerial vehicle remote sensing to identify a single invasive species or distinguish a target species from background objects, for example, monitoring of species such as Spartina alterniflora and Acanthospermum hispidum, which relies on the construction of an identification model based on the significant spectral difference between the target species and the surrounding vegetation. However, there is no public and specialized hyperspectral precision classification and identification technology scheme for Eupatorium odoratum and Bidens pilosa, which are highly similar in morphology, flowering period and habitat, making it difficult to meet the practical application needs of precise differentiation and large-scale dynamic monitoring of the two species.
[0107] To solve the above problems, the present embodiment establishes a complete, efficient and standardized classification and identification method based on unmanned aerial vehicle hyperspectral technology, which is specifically used for precise differentiation of Eupatorium odoratum and Bidens pilosa. By analyzing hyperspectral data, key spectral features and optimal identification bands that can stably differentiate Eupatorium odoratum and Bidens pilosa are mined and determined, and a “data preprocessing + data dimensionality reduction + classifier” model combination most suitable for classification of Eupatorium odoratum and Bidens pilosa is constructed and optimized, to realize high-precision automatic identification. At the same time, the effectiveness and stability of the method in different terrains and actual environments can be verified, providing reliable technical support for large-scale monitoring and precise management of the two types of invasive plants.
[0108] In order to more clearly explain the present application, a specific application example of a classification and identification method of gueldenstaedtia pekinensis and eriochloa speciosa based on unmanned aerial vehicle hyperspectral is provided below.
[0109] 1. Unmanned aerial vehicle hyperspectral data acquisition and preprocessing, including the following steps:
[0110] Step 1.1, using an unmanned aerial vehicle to carry a hyperspectral imager, flying according to a preset route to obtain hyperspectral image data. The unmanned aerial vehicle should fly at a moderate height and at a steady speed to ensure the integrity and accuracy of the data. Step 1.2, image data rectification and radiation correction. Step 1.3, geometric correction to solve the problems of projection distortion and geometric accuracy error. Step 1.4, selecting control points for image registration. Step 1.5, remote sensing image mosaicking is to splice two or more remote sensing images together to form a whole image. Step 1.6, outputting the processing results as data sets or images. As shown in Figure 2 .
[0111] 2. Hyperspectral data processing and feature optimization, including the following steps:
[0112] Step 2.1, inputting the data processing results generated in step 1. Step 2.2, performing spectral processing through differentiation, smoothing, and standard normal variable transformation methods. Step 2.3, performing spectral dimension reduction processing through principal component analysis. Step 2.4, processing data through different combinations of processing methods. Step 2.5, outputting the data sets or images obtained by different processing methods. As shown in Figures 3-6 .
[0113] 3. Gueldenstaedtia pekinensis and eriochloa speciosa classification model construction and optimization, including the following steps:
[0114] Step 3.1, inputting the combined results of different data processing methods generated in step 2.
[0115] Input the multiple hyperspectral data cubes obtained by step two (i.e. the results under different preprocessing method combinations, such as Gaussian smoothing-PCA dimension reduction data set, first derivative-PCA dimension reduction data set, standard normal variable transformation-PCA dimension reduction data set, etc.). For each data combination, the following operations are performed: a plant classification expert accurately outlines the region of interest (Region Of Interest, ROI) of gueldenstaedtia pekinensis, eriochloa speciosa and other ground objects (such as native vegetation, soil, water, buildings, etc.) on the preprocessed hyperspectral image according to the GPS points of the on-site investigation. Each ground object category should have a sufficient number of sample pixels (for example, not less than 5000 sample pixels for each of gueldenstaedtia pekinensis and eriochloa speciosa), and all sample points are randomly divided into a training set and a validation set in a ratio of about 7:3. The training set is used for model training, and the validation set is used for preliminary evaluation of the model and hyperparameter tuning.
[0116] Step 3.2, Constructing the combined classification model, by combining the support vector machine, random forest, and U-Net methods for classification processing.
[0117] For the multiple training datasets obtained in step 3.1, three different types of classification models are constructed and trained respectively:
[0118] a) Support vector machine (SVM) model: Radial basis function (RBF) is used as the kernel function. The key hyperparameter settings are as follows: penalty coefficient C=1.0, kernel function coefficient gamma='scale', maximum iteration number max_iter=-1 (i.e. no limit), decision function shape decision_function_shape='ovr' (one-versus-rest), break_ties=False, and set random_state=None. This study first uses this parameter for preliminary training, and then uses grid search (GridSearch) and cross-validation method to fine-tune around the key parameters C (try range such as [0.1,1,10,100]) and gamma (try range such as [0.001,0.01,0.1,1,'scale', 'auto']) to determine the optimal parameter combination for the current dataset.
[0119] b) Random forest (RF) model: The key hyperparameter settings are as follows: the number of decision trees n_estimators=100, the maximum depth of a single decision tree max_depth=256, the minimum number of samples required for internal node splitting min_samples_split=2, the minimum number of samples in a leaf node min_samples_leaf=1, the maximum number of features max_features='auto', the maximum number of leaf nodes max_leaf_nodes=None. To speed up the training process of large-scale data, set n_jobs=32 to fully utilize multi-core processors for parallel computing. To ensure reproducible results, fix random_state=42. Similarly, grid search and cross-validation are used to further optimize key parameters such as n_estimators (such as [100,200,300]) and max_depth (such as [128,256,None]). Record the feature importance ranking of each input feature (i.e. the principal component band after PCA dimensionality reduction) during training to assist in analyzing the key spectral features that distinguish the two types of invasive plants.
[0120] c) U-Net deep learning model: An adapted U-Net network structure was constructed for the spatial-spectral characteristics of hyperspectral images. The network input is an image block (Patch) containing the first N main principal component bands (e.g., the first 10 PCs). The network encoder part uses multiple convolutional layers (Conv2D) and maximum pooling layers (MaxPooling2D) to gradually extract multi-level features; the decoder part fuses shallow details and deep semantic information through upsampling (UpSampling2D) and skip connection (Skip Connection); the final output uses a convolutional layer with a Softmax activation function for pixel-by-pixel classification. The total number of training epochs (epochs) is set to 200. The optimizer uses stochastic gradient descent (SGD). To optimize the training process and improve model performance, the learning rate scheduling strategy combines exponential decay (Exponential Decay) and cosine annealing (CosineAnnealing) to dynamically adjust the learning rate. The model uses the classification cross-entropy loss function and introduces the early stopping (EarlyStopping) strategy to monitor the validation set loss to prevent overfitting.
[0121] Step 3.3: By comparing the overall classification accuracy, Kappa coefficient, and other parameters of each combined model, the model with the best overall performance is selected.
[0122] The trained classification model is used to predict the test set, and the classification results are shown in Figures 7-9 . The performance of each model combination is evaluated using quantitative indicators and qualitative visual interpretation. Taking the test results of "Plot One" as an example, the specific performance comparison is as follows:
[0123] Overall performance: Calculate the overall classification accuracy (OA) and Kappa coefficient of each model classification result. As shown in the table below, among all the method combinations, the "derivative-UNet" combination achieved the best overall performance (OA=0.902, Kappa=0.7834), significantly better than other combinations.
[0124] Subdivision performance: For the key distinction of guayule and horseweed classes, the producer accuracy (PA), user accuracy (UA) and F1-Score are calculated in detail. Analysis found that although the "standard normal variable transformation-UNet" combination has higher UA and F1 score for horseweed (UA=0.8974, F1=0.8861), its identification performance for guayule (F1=0.5638) is seriously unbalanced. The "first derivative processing-UNet" combination achieves the best overall balance in the identification of the two target species, with an F1 score of 0.9 for horseweed and 0.6614 for guayule, both at the leading level among all method combinations, and avoiding serious misjudgment imbalance.
[0125] Confusion matrix analysis: As shown in Figure 10 The confusion matrix diagram shows that the "first derivative processing-UNet" model correctly classifies 90% of the real horseweed pixels and 91.3% of the real guayule pixels, with a very low misclassification rate between the two, proving the strong distinguishing ability of the method for similar species.
[0126] Step 3.4, output the optimal classification model classification result data and image.
[0127] Apply the optimal classification model (FD-U-Net) selected in step 3.3 and its corresponding complete data processing process to the entire study area of the unmanned aerial high-spectral data to generate the final guayule and horseweed spatial distribution classification map.
[0128] Therefore, the present application proposes the first classification and identification method based on unmanned aerial high-spectral data specifically for distinguishing guayule and horseweed, two types of invasive plants with similar morphology and habitat. At the same time, a complete and optimized technical process is proposed and verified from data acquisition, various pre-processing (GS / FD / SNV), PCA dimensionality reduction to various classifiers (SVM / RF / U-Net) construction. Through experimental comparison, the best technical combination for accurate identification of guayule and horseweed is selected and determined (such as GS-PCA-SVM), which can effectively suppress noise and retain key distinguishing features, achieve high-precision classification. At the same time, the present application not only constructs a high-precision model, but also proves the stability and generalization ability of the method through verification on different terrains, making it have practical application value.
[0129] It is to be understood that the terms "one embodiment", "another embodiment", "other embodiments", "first embodiment", "second embodiment", etc. as may be found in the specification and / or in the claims, indicate that the alternative is included in at least one embodiment. These terms only specify the scope of claimable subject matter; and do not necessarily affect the scope of the application. Further, these terms only indicate particular embodiments of the applications. Other embodiments of the present application can be derived from the description, experimental results and / or the claims, without departing from the scope of the present application.
[0130] It is to be understood that the terms "including", "comprising", "consisting" and "consisting essentially of" or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.
[0131] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which is made according to the content of the specification and drawings of the present application, is also included in the patent protection scope of the present application.
Claims
1. A method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging, characterized in that, The method includes the following steps: S1: Acquire raw hyperspectral image data of the target area using a drone; S2: Perform preprocessing operations on the raw hyperspectral image data of the UAV to obtain a standardized hyperspectral dataset; wherein, the preprocessing operations include radiometric correction processing, geometric correction processing, image registration processing, and remote sensing image mosaicking processing; S3: Perform spectral transformation and dimensionality reduction on the standardized hyperspectral dataset to obtain multiple sets of optimized feature datasets formed by different combinations of spectral transformation and dimensionality reduction. S4: Construct multi-type classification models based on the multiple sets of optimized feature datasets. Select the optimal classification model through model training and performance evaluation. Use the optimal classification model to classify and identify Mikania micrantha and Clerodendrum trichotomum in the target area and output the classification results.
2. The method for classifying and identifying Mikania micrantha and Clematis armandii based on UAV hyperspectral imaging as described in claim 1, characterized in that, In step S1, the acquisition of the hyperspectral raw image data of the UAV is carried out by selecting a multi-rotor UAV equipped with a hyperspectral imager and performing the acquisition of the hyperspectral raw image data of the target area under preset flight parameters, flight time period and flight path conditions.
3. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 1, characterized in that, Step S3: Perform spectral transformation and dimensionality reduction on the standardized hyperspectral dataset to obtain multiple optimized feature datasets formed by different combinations of spectral transformation and dimensionality reduction, specifically including: S31: Read the standardized hyperspectral dataset and extract the spectral vector for each pixel. , where m is the number of hyperspectral bands, and the missing pixels are filled using the neighborhood mean fill method; S32: Three independent spectral transformation methods are used to enhance the spectral distinguishing features of Mikania micrantha and Agrostis spp. as well as the dimensionality reduction processing of principal component analysis. The three spectral transformation methods are combined with principal component analysis dimensionality reduction to form several optimized feature datasets. S33: Save the optimized feature datasets of several combinations as ENVI format, and annotate the ENVI format files with metadata including the combination processing method, the number of principal components, and the cumulative contribution rate, and wait for them to be called when building the subsequent classification model.
4. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 3, characterized in that, In step S32, three independent spectral transformation methods enhance the spectral distinguishability between Mikania micrantha and Clerodendrum trichotomum, specifically including: first derivative processing, Gaussian smoothing, and standard normal variable transformation, with the following expressions: ； ； ； ； ； In the formula, For the first The center wavelength of each band The band spacing is ≤5nm. wavelength The corresponding reflectance value, For the first The first derivative of the center wavelength of each band The standard deviation of the Gaussian kernel ranges from 0.5 to 2.0 and is adaptively adjusted according to the spectral noise level. To smooth the center wavelength, The size of the convolution window (5-7 bands). This is the smoothed reflectance value. For the first Reflectance values for each band, The mean of the spectral vector. denoted as the standard deviation of the spectral vector.
5. The method for classifying and identifying Mikania micrantha and Clematis armandii based on UAV hyperspectral imaging as described in claim 3, characterized in that, Step S32, the dimensionality reduction process of principal component analysis, specifically includes: S321: Calculate the covariance matrix of the spectral data The expression is as follows: ； in, Number of pixel samples for 3D spectral data matrix for 3D spectral mean vector; S322: Solving the covariance matrix eigenvalues and the corresponding feature vectors ; S323: Select principal components based on cumulative contribution rate, choosing the top... Construct the principal component projection matrix from the principal components. ; S324: Project the transformed spectral data onto the principal component space to obtain the dimensionality-reduced dataset. ,in for 3D spectral data matrix for Dimensionality reduction dataset, and .
6. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 3, characterized in that, Step S4: Construct multi-type classification models based on the multiple sets of optimized feature datasets. After model training and performance evaluation, select the optimal classification model. Use the optimal classification model to classify and identify Mikania micrantha and Clematis chinensis in the target area, and output the classification results, specifically including: S41: Based on the standardized hyperspectral dataset, select several sample data from the optimized feature dataset for each combination, and construct the training sample dataset for the optimized feature dataset for each combination based on the ROI region label and optimized feature dataset corresponding to each sample data. S42: Construct a support vector machine model, a random forest model, and a U-Net model, and combine them with the training sample dataset of the optimized feature dataset for each combination to obtain several classification model combinations. Use the training sample dataset of the optimized feature dataset for each combination to train and evaluate the classification model. Select the optimal classification model combination based on the performance evaluation results. S43: Based on the optimal classification model combination and the standardized hyperspectral dataset, perform classification and identification of Mikania micrantha and Clematis chinensis in the target region, and output the classification results.
7. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 6, characterized in that, In step S41, based on the ROI region label and optimized feature dataset corresponding to each sample data, a training sample dataset for each combination of optimized feature datasets is constructed, specifically including: S411: In advance, the ROI areas of Mikania micrantha, Herba Aspergillus niger and other land features are drawn on each sample data by manual ROI delineation as ROI area labels; wherein, in the ROI area labels, the label of Mikania micrantha is 0, the label of Herba Aspergillus niger is 1, and the label of other land features is 2; S412: Extract the pixel feature vector of the optimized feature dataset for each combination corresponding to each ROI region, and construct the training sample dataset of the optimized feature dataset for each combination consisting of the pixel feature vector and the ROI region label.
8. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 6, characterized in that, In step S42, the support vector machine model, random forest model, and U-Net model are constructed, specifically including: S421: Using radial basis functions as kernel functions, set the penalty coefficient C=1.0, kernel function coefficient gamma='scale', maximum number of iterations max_iter=-1, and set the decision function shape decision_function_shape='ovr', break_ties=False, random_state=None respectively to construct and obtain a support vector machine; S422: Set the number of decision trees n_estimators=100, the maximum depth of a single decision tree max_depth=256, the minimum number of samples required for internal node splitting min_samples_split=2, the minimum number of samples for leaf nodes min_samples_leaf=1, the maximum number of features max_features='auto', and the maximum number of leaf nodes max_leaf_nodes=None to build a random forest model; S423: Construct a U-Net network structure containing an encoder and a decoder. Set the network input to an image patch containing the first N principal component bands. The encoder uses multiple convolutional layers and max pooling layers to progressively extract multi-level features. The decoder fuses shallow details and deep semantic information through upsampling and skip connections. Set the network output to use a convolutional layer with the Softmax activation function for pixel-by-pixel classification. Set the total number of training epochs for the model to 200. Set the optimizer to use stochastic gradient descent to obtain the final U-Net model.
9. The method for classifying and identifying Mikania micrantha and Clerodendrum trichotomum based on UAV hyperspectral imaging as described in claim 6, characterized in that, In step S42, model training and performance evaluation are performed, and the optimal combination of classification models is selected based on the performance evaluation results. Specifically, this includes: S424: Divide the training sample dataset into a training set and a validation set according to a preset ratio. Use the training set to train the classification model in the classification model combination, and use the validation set to validate the classification model. S425: Based on the overall classification accuracy OA, Kappa coefficient, producer accuracy PA, user accuracy UA, and F1-Score obtained from the verification calculation, select the optimal classification model combination from several classification model combinations.
10. The method for classifying and identifying Mikania micrantha and Clematis armandii based on UAV hyperspectral imaging as described in claim 6, characterized in that, S43: Based on the optimal classification model combination and the standardized hyperspectral dataset, perform classification and identification of Mikania micrantha and Clematis armandii in the target region, and output the classification results, specifically including: S431: Extract the optimized feature dataset belonging to the target combination in the optimal classification model combination from the multiple optimized feature datasets of the standardized hyperspectral dataset; S432: The optimized feature dataset is transferred to the classification model in the optimal classification model combination for classification prediction, generating a standardized hyperspectral dataset of spatial distribution classification maps of Mikania micrantha and Clematis armandii.