Intelligent vegetation classification method and device for power transmission channel based on unet and hyperspectrum, and medium

Abstract

The application discloses a power transmission channel vegetation intelligent classification method and device based on UNET and hyperspectrum, and a medium, belongs to the field of computer vision and image processing, including collecting hyperspectrum images of the power transmission channel area by a hyperspectrum remote sensing device and preprocessing; designing and constructing a UNET neural network, training the model using labeled tree category data, classifying using newly collected test data, outputting the tree category to which each pixel point belongs, performing post-processing operations, removing small area noise through morphological operations, using a conditional random field optimization model output, and improving the classification results of similar categories between different tree species. The application improves the accuracy of vegetation classification, improves the processing efficiency and automation degree, effectively solves the classification problem of similar tree species, enhances the spatial consistency of the classification results, saves the labor cost and reduces the error rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer vision and image processing, and particularly relates to a power transmission channel vegetation intelligent classification method and device based on UNET and hyperspectrum and a medium. BACKGROUND

[0002] With the increasing requirements of the power industry on the safety and stability of power transmission lines, vegetation monitoring and management of power transmission channels have gradually become an important issue. Especially with the changes in vegetation growth, different trees in the power transmission channel area may pose a threat to the power transmission line, and if not monitored and handled in time, it may lead to power equipment failure, shutdown or even safety accidents. Therefore, accurate and efficient classification and monitoring of vegetation in the power transmission channel area has become the key to ensuring the safe operation of power transmission lines. Traditional vegetation monitoring methods rely on manual inspection or ground observation, which is time-consuming and labor-intensive and difficult to cover a large area.

[0003] With the development of unmanned aerial vehicles, satellites and other remote sensing technologies, vegetation classification methods based on remote sensing images have become the mainstream of current research and application. Hyperspectral remote sensing technology, as a remote sensing method that can provide fine-grained spectral information, has been widely used in vegetation classification, environmental monitoring and other fields in recent years.

[0004] Through hyperspectral remote sensing data, rich spectral information of each pixel can be obtained, thereby realizing accurate identification and classification of ground objects. Especially in vegetation identification, different vegetation types have different spectral characteristics, which enables hyperspectral data to provide more useful information. Although hyperspectral data has high spectral resolution and can effectively distinguish different vegetation types, its high-bit characteristics and noise problems are still important factors affecting classification accuracy. Traditional classification methods, such as maximum likelihood classification (MLC) and support vector machine (SVM), although can perform vegetation classification to some extent, but since these methods usually cannot fully exploit the spatial features in hyperspectral data and have limited processing capacity for high-dimensional data, often result in low classification accuracy, especially in the classification of vegetation and background junctions and complex terrain areas.

[0005] In recent years, the successful application of deep learning techniques, especially convolutional neural networks (CNNs), in remote sensing image processing has provided a new direction for addressing the shortcomings of traditional methods in hyperspectral remote sensing data processing. In particular, the UNET neural network, due to its superior image segmentation capabilities, has been widely applied to the semantic segmentation task of remote sensing images. UNET (a convolutional neural network based on deep learning) can effectively extract contextual information from images through its encoder-decoder structure and perform pixel-level classification, making it a powerful tool for spatial information processing. Using the UNET neural network for vegetation classification of hyperspectral remote sensing data not only fully exploits the spectral information of the image but also effectively processes the spatial features of the image, improving classification accuracy.

[0006] Despite the significant achievements of deep learning-based remote sensing image classification methods in various fields, they still face a series of challenges in hyperspectral data processing. The high dimensionality of the data can lead to excessive computational load, and traditional deep learning models may not be able to efficiently run on limited computing resources, especially when dealing with large-scale remote sensing data. Moreover, deep learning models typically require a large amount of labeled data for training, but obtaining high-quality labeled data is difficult, and there may be insufficient or biased data in specific environments (such as special weather conditions, different geographical regions of vegetation types, etc.), which makes the generalization ability of existing models poor and affects their reliability in practical applications.

[0007] For the vegetation classification task of power transmission corridors, existing technologies still face problems such as high dimensionality of data, insufficient classification accuracy, low computational efficiency, and lack of labeled data when processing hyperspectral remote sensing data. To address these issues, an intelligent vegetation classification method for power transmission corridors based on UNET and hyperspectral data is proposed. This method combines hyperspectral remote sensing data with the image segmentation capabilities of the UNET neural network, effectively extracting spectral and spatial features from hyperspectral data, and providing more accurate vegetation classification results. While ensuring classification accuracy, this method can also reduce dependence on labeled data through model optimization and data augmentation, improving the model's generalization ability, and through reasonable computational architecture design, improving the model's efficiency in large-scale data processing. By leveraging the advantages of UNET, it can better handle complex terrain, high similarity between vegetation and background, and improve the reliability and real-time performance of power transmission corridor monitoring. SUMMARY

[0008] In view of the above existing problems, the present invention addresses the shortcomings of existing technologies in vegetation classification in complex environments, especially for the specific application scenario of power transmission corridors, providing a more efficient and accurate vegetation monitoring method. Through this method, real-time monitoring capabilities can be improved while ensuring high classification accuracy, thereby providing strong technical support for the safe management of power lines.

[0009] To solve the above technical problems, an intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral is proposed, which includes,

[0010] Collect the hyperspectral image of the power transmission channel area through the hyperspectral remote sensing device, which covers the spectral information of different tree species, and preprocess the collected hyperspectral image; design and build a UNET neural network, which includes an encoder, a decoder and a skip connection, and perform pixel-level classification task on hyperspectral data; use the labeled tree species data to train the model, adjust the model parameters through back propagation and optimization algorithm; use the newly collected test data to classify based on the trained UNET model, output the tree species to which each pixel belongs, evaluate the classification performance of the model according to the classification result, and calculate the evaluation index; perform post-processing operation, remove small area noise through morphological operation, smooth the classification boundary, and use conditional random field optimization model output to further eliminate the confusion between categories; improve the classification result of similar categories between different tree species through misclassification correction strategy, and improve the overall classification accuracy.

[0011] As a preferred scheme of the intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral, wherein: the hyperspectral image includes, selecting a typical power transmission channel area, using a hyperspectral remote sensing sensor to collect the vegetation in the current area, the spectral data collected by the hyperspectral sensor covers the electromagnetic spectrum from 400nm to 2500nm, a total of 210 bands, and the spectral resolution of each band is 10nm;

[0012] For each tree, collect spectral data several times from different angles and heights, and the collected hyperspectral image data contains spectral information of 9 tree species, and each tree species has collected not less than 30 sample points;

[0013] The preprocessing includes spectral data calibration and denoising, spectral feature extraction and enhancement, and data standardization;

[0014] The spectral data calibration and denoising includes calibration and denoising in the order of radiation calibration and noise removal; wherein, the radiation calibration, by comparing the collected spectral data with the reflectivity of the known standard target, the measurement results of the hyperspectral sensor are calibrated by radiation, and the influence of light change and sensor characteristics on the data is eliminated;

[0015] Noise removal includes using discrete wavelet transform to denoise the collected hyperspectral data, decomposing the signal into components of different frequencies, and removing high-frequency noise; the spectral data is expressed in vector form Wherein, represents the spectral value of the i-th band, Wavelet decomposition is performed on each spectrum data for the total number of wave bands to obtain low-frequency components and high-frequency components, the high-frequency components are subjected to threshold processing to remove noise, and the processed signals are reconstructed to obtain denoised spectrum data;

[0016] The noise removal also includes performing atmospheric correction, using an atmospheric correction algorithm to perform atmospheric correction on the data to eliminate the influence of atmospheric components on spectral reflection;

[0017] The spectrum feature extraction and enhancement includes wave band selection, using principal component analysis to select wave bands, performing mean value reduction centering processing on the spectrum data of each sample, calculating the covariance matrix of the sample data, calculating the eigenvalues and eigenvectors of the covariance matrix C, selecting the first k principal components, and projecting the original data onto the k principal components to obtain the reduced dimension data;

[0018] Feature enhancement, using spectral ratio method to enhance the spectrum of different wave bands, emphasizing the characteristics of vegetation by calculating the ratio of any wave band to the remaining wave band, and using the ratio of red wave band and near infrared wave band to enhance the reflection characteristics of vegetation, which has significant difference in vegetation and non-vegetation areas;

[0019] The data standardization is to normalize the spectrum value of each wave band.

[0020] As a preferred scheme of the power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum, the UNET neural network is designed and constructed, including designing a UNET neural network model and training the UNET neural network model.

[0021] The UNET neural network model includes an encoder, a decoder and a skip connection, inputting hyperspectral image data, the encoder uses a convolution kernel to extract local features of the input data in the convolution layer, a 3x3 convolution kernel is selected, the number of convolution kernels increases layer by layer after each convolution operation, and the spatial size of the hyperspectral image is reduced through a pooling operation.

[0022] The decoder performs up-sampling on the image through a deconvolution layer, maps the features extracted by the encoder to the same spatial size as the input image, and obtains the class label of each pixel, i.e., the tree species, through pixel classification, at the same time, the low-level features of the encoder and the high-level features of the decoder are combined through a skip connection, a softmax activation function is used to convert the prediction value of each pixel into a class probability, and a classification map is output, each pixel point is labeled as one of the nine tree species.

[0023] The training UNET neural network model comprises: using the preprocessed hyperspectral data as input, calculating the output labeled vegetation type image through forward propagation, and dividing the data set into 70% as a training set and 30% as a test set;

[0024] The cross-entropy loss function is used to measure the difference between the model prediction value and the real label, the classification error of each pixel is calculated, and the sum of all pixel errors is the loss of the entire image :

[0025] ;

[0026] wherein, is the real label, is the prediction probability, is the total number of pixels;

[0027] The Adam optimizer is used to adjust the update learning rate, and the update rule is as follows:

[0028] ;

[0029] wherein, is the parameter, is the learning rate, and are the first moment and the second moment estimates respectively, is a constant to prevent division by zero;

[0030] The training rule is set, the number of samples of each training batch is set to 16, the total number of training rounds is set to 100 rounds, the learning rate decay strategy is set, and the learning rate is reduced to 0.1 times of the original every 10 rounds;

[0031] The gradient of the loss function with respect to each weight is calculated by back propagation, and the gradient descent algorithm is used to update the weight and bias, at the end of each training round, the test set data is used to evaluate the model, and the model with the highest accuracy and generalization ability is selected.

[0032] As a preferred scheme of the power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum, wherein: the classification result comprises, after the training of the UNET neural network is completed, performing a vegetation classification task on the model;

[0033] In the classification task phase, the preprocessed test data is used as input, and the hyperspectral data of each test sample is obtained through the forward propagation of the network to obtain a pixel-level classification result, which is input into the trained UNET model for prediction, and the output is the probability distribution of each pixel point under the set 9 kinds of tree categories, and each pixel point is assigned to the category with the highest probability through the softmax activation function, as the prediction label;

[0034] Each test image generates an output image with the same size as the input image, wherein each pixel point is classified as one of the set tree species, and the predicted label and the real label image of the output are compared through a direct pixel-level classification matching method to determine whether it is a correct classification, the probability of each pixel point belonging to each category is calculated through a Softmax activation function, and the maximum probability category is selected as the final classification label;

[0035] The classification accuracy of the calculation model and the evaluation index include,

[0036] Accuracy, the proportion of all predicted correct pixels to the total pixels;

[0037] Precision, the proportion of the actual current category when predicted as any category as positive;

[0038] Recall, the proportion of the predicted current category when the actual is any category as positive;

[0039] F1-score, the harmonic mean of precision and recall, which comprehensively considers the performance of the classifier.

[0040] As a preferred scheme of the intelligent classification method for vegetation in power transmission channels based on UNET and hyperspectrum, the post-processing operation includes removing noise by morphology, smoothing the classification boundary, and optimizing the conditional random field.

[0041] The morphological noise removal includes using a 3x3 or 5x5 structure element for morphological operation, applying erosion operation to reduce noise in the region, restoring the shape of the target region through inflation operation, and removing small noise in the classification result by combining inflation and erosion operation, and retaining the identified vegetation region.

[0042] The smoothing of the classification boundary includes denoising by using a boundary smoothing algorithm to smooth the boundary by reducing the mutation in the image:

[0043] ;

[0044] Wherein, is the smoothed image, is the original classification image, represents the image gradient, is the balance term, is the total number of images, and a is a point defined on the image.

[0045] As a preferred scheme of the intelligent classification method for vegetation in power transmission channels based on UNET and hyperspectrum, the conditional random field optimization includes adjusting the classification boundary by considering the spatial and spectral consistency between pixels :

[0046] ;

[0047] wherein, is the output label, is the input image, is the similarity measure between adjacent pixels and , and , is the weight parameter, is the normalization constant.

[0048] As a preferred scheme of the power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum, wherein: the misclassification correction strategy includes that for the area with misclassification, including edge area division low contrast area, further correction is carried out, and the classification result of the area similar to the surrounding area is corrected to the correct category by setting the correction rule;

[0049] The correction rule includes,

[0050] Rule one, based on the category consistency of the pixel neighborhood, when the classification of a pixel is consistent with the category of the surrounding pixels and the consistency is higher than 85%, the current classification result is maintained, and when the classification of a pixel is not consistent with the category of the surrounding pixels and the consistency is not more than 85%, correction is carried out.

[0051] Rule two, based on the similarity of spectral features, when the spectral features of a pixel are similar to the typical features of any tree category and the similarity is higher than 70%, the category correction is carried out in the post-processing stage.

[0052] A computer device comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum when executing the computer program.

[0053] A computer readable storage medium, which stores a computer program, and the computer program implements the steps of the power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum when executed by a processor.

[0054] The beneficial effects of the present application are: (1) by combining hyperspectral remote sensing data, more rich spectral information than traditional RGB images can be obtained, especially in distinguishing similar tree species, hyperspectral data provide more accurate spectral features. In this way, the classification error caused by the similarity of spectral features between tree species can be effectively reduced.

[0055] The UNET network is particularly suitable for image tasks with complex spatial structures due to its unique encoder-decoder structure and skip connections. In vegetation classification, UNET can effectively capture the spatial distribution characteristics of different trees, further improving classification accuracy.

[0056] Post-processing techniques such as morphological operations, boundary smoothing, and conditional random field (CRF) optimization significantly improve the quality of classification results. In particular, correcting misclassified areas and smoothing classification boundaries reduces inaccurate classification caused by small regional noise and edge effects, making the final output more stable and consistent.

[0057] (2) By using deep learning models to automatically classify hyperspectral data, the classification speed is greatly improved compared to traditional manual classification methods (such as classification based on spectral indices), and human error is reduced.

[0058] By training UNET neural networks, the system can automatically identify and classify different tree types, avoiding human intervention, adapting to large-scale data sets, and efficiently processing large-area vegetation monitoring tasks for power transmission corridors.

[0059] (3) In vegetation monitoring of power transmission corridors, different tree species often have similar spectral characteristics, which can easily lead to classification confusion. This method combines hyperspectral data with UNET neural networks, not only improving the ability to identify fine-grained tree species, but also optimizing classification results through post-processing techniques, significantly improving the ability to distinguish between similar tree species.

[0060] (4) Traditional image classification methods may produce inconsistent or jagged classification boundaries in the transition area between vegetation boundaries and different tree species, while UNET neural networks can better maintain spatial structure. On this basis, using morphological operations, total variation denoising, and smoothing boundary post-processing methods makes the final classification results more consistent in space, especially in tree boundary and dense vegetation areas. CRF optimization further enhances the spatial consistency of classification results, reducing misclassification caused by differences between adjacent pixels during classification.

[0061] (5) By using hyperspectral data and UNET neural networks for joint modeling, it can handle vegetation monitoring tasks for power transmission corridors under different environmental conditions. Compared with traditional methods, this deep learning-based model has stronger generalization ability and can adapt to various types of trees and different regional spectral characteristics. After sufficient training, the model can identify and classify unseen tree species and regions, with strong adaptability.

[0062] (6) Greatly reduces the dependence on manual inspection and manual classification, and reduces the interference of human factors on classification accuracy through an automated classification process. For monitoring of large-scale power transmission channels, the use of the patent technology can not only significantly improve work efficiency, but also reduce potential risks caused by human errors. The results of post-processing optimization can greatly reduce the phenomenon of incorrect classification caused by unclear classification boundaries and noise, further improving the reliability of the results.

[0063] (7) Combines hyperspectral remote sensing data and deep learning methods (UNET neural network), making up for the shortcomings of traditional methods in dealing with complex vegetation classification tasks. This method realizes efficient and accurate classification in various tree species through innovative model design and post-processing procedures. The invention is not only suitable for vegetation monitoring of power transmission channels, but also can be extended to other types of remote sensing data analysis tasks, and has a wide application prospect. In the safety monitoring of power facilities, through accurate vegetation classification, potential risks (such as tree barriers) on the power transmission line can be found in time, providing effective data support and decision basis for power companies, thereby improving the safety and reliability of the power transmission line. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0065] Fig. 1 The overall flowchart of the vegetation intelligent classification method for power transmission channel based on UNET and hyperspectrum provided by an embodiment of the present application.

[0066] Fig. 2 The UNET neural network model structure diagram of the vegetation intelligent classification method for power transmission channel based on UNET and hyperspectrum provided by an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0068] In the following description, a lot of specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can be practiced in other manners different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the present application, therefore, the present application is not limited to the specific embodiments disclosed below.

[0069] Embodiment 1, reference Figs. 1-2 As a first embodiment of the present application, the embodiment provides a power transmission channel vegetation intelligent classification method based on UNET and hyperspectrum, comprising:

[0070] S1: Collect the hyperspectral image of the power transmission channel area through the hyperspectral remote sensing device, which covers the spectral information of different tree species, and pre-process the collected hyperspectral image.

[0071] It should be noted that a typical power transmission channel area is selected, and a hyperspectral remote sensing sensor (such as ASDFieldSpec 4 hyperspectral imager) is used to collect the vegetation in the current area, and the spectral data collected by the hyperspectral sensor covers the electromagnetic spectrum from 400nm to 2500nm, a total of 210 bands, and the spectral resolution of each band is 10nm;

[0072] For each tree, spectral data is collected several times from different angles and heights, and the collected hyperspectral image data contains the spectral information of the set 9 tree species (for example: eucalyptus, camphor tree, construction tree, blue oak, tung tree, cypress, pine, cedar, and bamboo); Each tree species has collected not less than 30 sample points; During the collection process, different parts of the tree (such as leaves, trunks, branches, etc.) are included to ensure the comprehensiveness and diversity of the data.

[0073] The sample collection positions are widely distributed and cover different climate conditions, tree health conditions and vegetation growth stages to simulate the influence of real environmental changes on spectral data. When each sample point is collected, the GPS position of the point is recorded while measuring the spectral data to ensure the spatial accuracy of the subsequent data.

[0074] Further, the purpose of data preprocessing is to remove noise, reduce data redundancy, and ensure high-quality data to improve the performance of the subsequent classification model;

[0075] Specifically, it includes spectral data calibration and denoising, spectral feature extraction and enhancement, and data standardization.

[0076] The spectral data calibration and denoising includes calibration and denoising in the order of radiation calibration and noise removal; wherein,

[0077] First, radiation calibration is performed. The measurement results of the hyperspectral sensor are calibrated by comparing the collected spectral data with the reflectivity of known standard targets (such as white and black plates), so as to eliminate the influence of light changes and sensor characteristics on the data;

[0078] Second, noise removal including wavelet transform. During the collection of hyperspectral data, noise interference is inevitable. Common noise sources include sensor noise, atmospheric interference, object shadow, etc.

[0079] The collected hyperspectral data are denoised using discrete wavelet transform, the signal is decomposed into components of different frequencies, and high-frequency noise is removed:

[0080] The spectral data are expressed in vector form , wherein represents the spectral value of the i-th waveband, is the total number of wavebands;

[0081] Each spectral data is decomposed by wavelet to obtain low-frequency components and high-frequency components;

[0082] The high-frequency components are thresholded to remove noise;

[0083] The processed signal is reconstructed to obtain denoised spectral data.

[0084] Noise removal also includes atmospheric correction. An atmospheric correction algorithm (such as FLAASH algorithm) is used to correct the data, so as to eliminate the influence of atmospheric components (such as water vapor, aerosols, etc.) on spectral reflection;

[0085] The spectral feature extraction and enhancement includes, first, band selection, and second, feature enhancement;

[0086] First, band selection. Hyperspectral data contain a large number of wavebands, not all of which are useful for vegetation classification. Some spectral information of some wavebands may be redundant or even noise sources. In order to improve the calculation efficiency and avoid overfitting, principal component analysis (PCA) is used to select the wavebands. PCA can project the original data onto a new set of orthogonal bases, retaining the main variation information in the original data.

[0087] The principal component analysis is used to select the wavebands. The spectral data of each sample are centered by mean subtraction;

[0088] The covariance matrix C of the sample data is calculated:

[0089] ;

[0090] , wherein is the i-th sample, is the mean, N is the number of samples, i is the variable index, and T is the transpose matrix;

[0091] Eigenvalues and eigenvectors of the covariance matrix C are calculated, and the first k principal components are selected;

[0092] The original data is projected onto the k principal components to obtain the reduced dimension data;

[0093] Secondly, feature enhancement, in order to further improve the accuracy of the model, the spectral ratio method is used to enhance the spectrum of different wave bands, and the characteristics of the vegetation are emphasized by calculating the ratio of any wave band to the remaining wave band, and the reflection characteristics of the vegetation are enhanced by using the ratio of the red wave band and the near infrared wave band. The ratio has significant difference between vegetation and non-vegetation area; for different tree species, the ratio difference between different wave bands can significantly improve the distinguishability in classification;

[0094] The data standardization is, because the reflectivity range of different wave bands in hyperspectral data is different, in order to avoid that the spectral information of some wave bands has too great influence on model training, the data needs to be standardized.

[0095] Specifically, the spectral value of each wave band is normalized according to the following formula to obtain the normalized result :

[0096] ;

[0097] Wherein, x is the original spectral value, μ is the mean of the wave band, and σ is the standard deviation of the wave band.

[0098] It should be noted that in order to verify the effect of the model, the collected spectral data needs to be divided into training set and test set. The spectral data of each tree is randomly allocated 70% of the samples as the training set, and 30% of the samples as the test set. Through this division method, the diversity and generalization ability of the data are guaranteed.

[0099] S2: design and build UNET neural network, including encoder, decoder and jump connection, for pixel-level classification task of hyperspectral data, use the labeled tree category data to train the model, adjust the model parameters through back propagation and optimization algorithm.

[0100] When classifying the vegetation in the power transmission channel, the UNET neural network model based on hyperspectral remote sensing data can fully excavate the spectral characteristics between different tree species and perform pixel-level classification.

[0101] Further, the UNET neural network model is designed and the UNET neural network model is trained;

[0102] The design of the UNET neural network model includes an encoder, a decoder and a jump link.

[0103] Encoder: responsible for extracting features from input images (hyperspectral data).

[0104] Decoder: responsible for restoring the extracted features to the original image size and performing pixel-level classification.

[0105] In this invention, the design of UNET takes into account the particularity of hyperspectral data (i.e. each pixel contains spectral information of multiple bands), and the convolutional and pooling layers in the UNET network need to be designed to adapt to the multi-dimensional features of hyperspectral data.

[0106] (1) Encoder part

[0107] The encoder extracts spatial and spectral features of different scales from the hyperspectral image through a series of convolution and pooling operations. Specifically,

[0108] Convolutional layer: local feature extraction is performed on the input data using a convolution kernel, and a 3x3 size convolution kernel is selected. After each convolution operation, the number of convolution kernels increases layer by layer, from 16 to 256.

[0109] Pooling layer: maximum pooling operation is used to reduce the spatial resolution of the data, so as to retain the main features and reduce the computational complexity. The pooling operation gradually reduces the spatial size of the hyperspectral image, but retains the main feature information.

[0110] (2) Decoder part

[0111] The decoder part is responsible for mapping the features extracted by the encoder to the same spatial size as the input image, and obtaining the class label of each pixel (i.e. tree species) through pixel classification.

[0112] The decoder restores the feature map extracted by the encoder to the same spatial size as the input image through the deconvolution layer, and

[0113] obtains the class label of each pixel, i.e. tree species, through pixel classification, and at the same time, uses skip connection to combine the low-level features of the encoder with the high-level features of the decoder, which can better preserve the detail information of the image.

[0114] The output layer uses the softmax activation function to convert the prediction value of each pixel to a class probability, and outputs a classification map, each pixel is labeled as one of the 9 tree species;

[0115] Further, the training UNET neural network model includes using the preprocessed hyperspectral data as input, calculating the output labeled vegetation type image through forward propagation, each pixel point is labeled as one of the nine trees, and the data set is divided into 70% as a training set and 30% as a test set.

[0116] The cross-entropy loss function is used to measure the difference between the model prediction value and the true label, calculate the classification error of each pixel, and the sum of all pixel errors is the loss of the entire image :

[0117] ;

[0118] Wherein, is the true label, is the predicted probability, is the total number of pixels;

[0119] The Adam optimizer is used to optimize the weight of the UNET model, and the update learning rate is adjusted, and the update rule is as follows:

[0120] ;

[0121] Wherein, is the parameter, is the learning rate, and are the first moment and the second moment estimates respectively, is a constant to prevent division by zero;

[0122] The present application sets the optimized training rule as follows according to the characteristics of high-dimensional hyperspectral data, large network parameter quantity and power transmission channel vegetation classification task:

[0123] ①Set the sample number of each training batch to 16 to avoid memory overflow; the hyperspectral image contains a large number of bands, and the single sample feature map occupies memory significantly higher than ordinary images; if the batch is too large, it is easy to cause memory overflow, and if the batch is too small, it will cause unstable gradient estimation, and the sample number of each training batch is 16, which can obtain stable and reliable gradient update under the condition of controllable memory, and it is the optimal value of the network structure.

[0124] ②Set the total number of training rounds to 100 rounds, and in the training process, all training samples will be traversed every round. This number of rounds can make UNET fully learn the spectral-spatial joint features and realize loss convergence. Insufficient training (such as 50 rounds) will cause insufficient feature extraction, and excessive training (such as more than 150 rounds) will cause the model to overfit to the spectral noise of limited samples, and 100 rounds can achieve the best balance between convergence efficiency and generalization performance. After each round of training, verification is performed on the test set to avoid overfitting.

[0125] ③ A learning rate decay strategy was implemented, reducing the learning rate to 0.1 times its original value every 10 epochs to improve the model's convergence and stability. Hyperspectral classification requires a high learning rate for rapid searching in the early stages of training, while a lower learning rate is needed in the later stages for finer optimization. Too high a decay frequency can lead to premature entrapment in local optima, while too slow decay can cause gradient oscillations. Verification showed that decaying the learning rate once every 10 epochs effectively improves convergence stability and final classification accuracy.

[0126] The gradient of the loss function with respect to each weight is calculated through backpropagation, and the weights and biases are updated using the gradient descent algorithm. At the end of each training round, the model is evaluated using test set data, and the model with the highest accuracy and generalization ability is selected.

[0127] Assuming the input image size is 480×480×20 (480 pixels high, 480 pixels wide, and 20 spectral bands), during training, each input image undergoes several convolution and pooling operations, ultimately outputting a 480×480×9 classification map (9 class probabilities for each pixel). Calculate the cross-entropy loss and update the model parameters based on the gradient.

[0128] It should be noted that at the end of each training round, the model is evaluated using test set data, and metrics such as classification accuracy, recall, and F1-score are calculated. The model with the highest accuracy and good generalization ability is ultimately selected.

[0129] S3: Based on the trained UNET model, classify the data using newly collected test data, output the tree category to which each pixel belongs, evaluate the classification performance of the model based on the classification results, and calculate the evaluation index.

[0130] Furthermore, after training the UNET neural network, the model is used to perform a vegetation classification task.

[0131] In the classification task stage, the preprocessed test data is used as input. The hyperspectral data of each test sample is passed through the forward propagation of the network to obtain a pixel-level classification result, which is then fed into the trained UNET model for prediction. The output is the probability distribution of each pixel under the set 9 tree categories. Through the softmax activation function, each pixel is assigned to the category with the highest probability as the predicted label.

[0132] Each test image generates an output image of the same size as the input image, where each pixel is classified as one of the specified tree species. The predicted label and the true label are compared using a direct pixel-level classification matching method. If the predicted category of a pixel matches the true label, it is considered a correct classification; otherwise, it is considered an incorrect classification.

[0133] The final output of the UNET model is the probability distribution of each pixel point in 9 tree categories. The probability of each pixel point belonging to each category is calculated by the Softmax activation function, and the maximum probability category is selected as the final classification label.

[0134] Compare the output image with the real label image to calculate the classification accuracy of the model and the evaluation indicators include,

[0135] Accuracy , the proportion of all correctly predicted pixels in the total pixels :

[0136] ;

[0137] Precision , the proportion of the actual current category when predicted as any positive class:

[0138] ;

[0139] Recall , the proportion of the current category when predicted as any positive class:

[0140] ;

[0141] F1-score, the harmonic mean of precision and recall , which comprehensively considers the performance of the classifier:

[0142] ;

[0143] Where, is the number of correctly predicted pixels, is the number of incorrectly predicted pixels, is the number of actual positive pixels but predicted as the current category, is the number of correctly predicted pixels.

[0144] S4: Perform post-processing operations, remove small area noise through morphological operations, smooth the classification boundary, and use conditional random fields to optimize the model output, further eliminating confusion between categories.

[0145] Through the misclassification correction strategy, improve the classification results of similar categories between different tree species, and improve the overall classification accuracy.

[0146] Specifically, in the vegetation classification task, although the preliminary classification based on the UNET neural network can achieve high accuracy, post-processing operations are usually needed to further improve the classification effect and accuracy. Post-processing not only improves the noise and errors in the classification results, but also further optimizes the output of the model, making the final results more accurate; the images classified by the UNET neural network may contain the following problems:

[0147] Small area noise: Some small classification areas may be caused by noise or errors, which usually have no actual meaning in real scenes.

[0148] Inter-class confusion: Certain tree species may be misclassified due to similar spectral characteristics or the influence of the surrounding environment, especially similar species such as oak and pine, eucalyptus and construction trees, etc.

[0149] Spatial consistency problem: There may be pixel-level discontinuity or inconsistency in the classification results, affecting the final classification quality.

[0150] Further, morphological noise removal, smoothing of classification boundaries, and conditional random field optimization improve classification accuracy;

[0151] The morphological noise removal includes using a 3x3 or 5x5 structure element for morphological operation, applying erosion operation to reduce noise in the region, restoring the shape of the target region through dilation operation, and removing small noise in the classification result by combining dilation and erosion operations, and retaining the identified vegetation area;

[0152] To improve the spatial consistency of the classification results, smoothing the classification boundary includes using a boundary smoothing algorithm to reduce the mutation in the image and smooth the boundary to improve classification accuracy:

[0153] ;

[0154] wherein, is the smoothed image, is the original classification map, represents the image gradient, is the balance term, is the total number of images, and a is a point defined on the image.

[0155] It should be noted that the conditional random field CRF adjusts the classification boundary by considering the spatial and spectral consistency between pixels :

[0156] ;

[0157] wherein, is the output label, is an input image, is a neighboring pixel point and , and a similarity measure between is a weight parameter, is a normalization constant. The goal of CRF is to reduce the label variation between adjacent pixels by considering the mutual influence between pixels, so that similar pixels are assigned to the same class.

[0158] In summary, although the UNET neural network can achieve good results in preliminary classification, there may still be noise, misclassification or unclear boundary problems in the classification results. Traditional image processing methods often ignore the importance of post-processing for improving classification accuracy, while the present technology introduces morphological operations, conditional random field (CRF) optimization and class smoothing post-processing techniques to further improve the accuracy and spatial consistency of the classification results.

[0159] Morphological operations (dilation and erosion) remove small noise regions, fill small holes, and smooth the classification boundary;

[0160] CRF optimization enhances the spatial consistency between pixels, making the labels between adjacent pixels more consistent and avoiding blurred boundaries and label confusion; class smoothing and misclassification correction further improve the distinction between tree species, especially between similar species such as Qinggang and pine trees, reducing the occurrence of misclassification. These post-processing operations significantly improve the reliability and accuracy of classification, especially for complex vegetation areas, and have high practical value.

[0161] For areas with misclassification, including low-contrast areas in edge regions, further correction is performed by setting a correction rule to correct the classification results of areas with similar surrounding classes to the correct class;

[0162] The correction rule includes,

[0163] Rule one, based on the consistency of the pixel neighborhood class, the neighborhood class consistency threshold is set to 85%. The threshold is determined based on the local spatial structure characteristics of the hyperspectral vegetation image: the neighborhood pixels usually have strong spatial continuity, when the consistency is higher than 85%, it means that the current pixel is very likely to belong to the dominant class and does not need to be corrected; if the consistency is lower than 85%, it means that the pixel may fall on the classification edge or be disturbed by noise and needs to be corrected according to the majority vote. If the threshold is set too high (such as 95%), it will lead to a large number of real boundary pixels being misjudged as abnormal; if it is set too low (such as 60%), it will cause excessive correction of the class that should be kept, reducing the stability of the spatial structure. Therefore, 85% can achieve the best balance between "keeping the real boundary" and "removing noise errors".

[0164] Rule two, based on the similarity of spectral features, the spectral similarity threshold is set to 70%. Hyperspectral data has strong spectral resolution, but there is still local spectral overlap between different trees. If the threshold is too high (such as 85%), the tree species with similar spectra will not be corrected correctly by this rule; if the threshold is too low (such as 50%), a large number of different classes will be misjudged as high similarity, causing new misclassification; 70% can ensure the reliability of spectral matching while avoiding misjudgment, which is the optimal threshold verified by a large number of experiments. Then the Euclidean distance or cosine similarity is used to calculate the similarity between the spectral curve of the target pixel and the standard spectral curve of each tree class, and the corrected class is based on the similarity of spectral features.

[0165] First, the pixel neighborhood analysis (rule one) is performed to correct the local classification error; then the spectral features are verified again (rule two) to ensure that the corrected class matches the spectral features; finally, the global smoothing processing is performed to reduce the noise of the classification boundary, making the result more smooth and natural.

[0166] In order to improve the correction effect, a joint strategy is adopted:

[0167] By designing the class smoothing and misclassification correction mechanism, the problem of classification error in traditional vegetation classification method under the condition of similar tree species and spectral overlap is solved.

[0168] In vegetation monitoring of transmission corridors, many tree species have high spectral similarity, and traditional methods often rely on simple spectral features (such as NDVI, EVI, etc.) for classification, which has low recognition accuracy for similar trees. The technology of this patent application combines the multi-dimensional information of hyperspectral data, and through the learning ability of neural network and post-processing correction mechanism, it can classify these similar trees more accurately. The misclassification correction mechanism distinguishes between similar spectral tree species by analyzing the class consistency of the surrounding pixels, improving the overall classification accuracy.

[0169] In summary, the combination of hyperspectral remote sensing data and UNET neural network solves the shortcomings of traditional remote sensing data classification methods (such as classification methods based on spectral indices) in dealing with complex vegetation types and distinguishing similar tree species.

[0170] Hyperspectral remote sensing data contains rich spectral information, covering not only visible light bands but also infrared bands and other multiple bands of information, which helps to accurately distinguish different types of vegetation. UNET neural network, due to its strong spatial feature learning ability, is very suitable for processing pixel-level segmentation tasks. In the technology of this patent application, UNET can effectively extract the features of hyperspectral data through the encoder-decoder structure of the convolutional neural network, and perform fine vegetation classification. This innovative technology can fully utilize the advantages of hyperspectral data, significantly improving the classification accuracy, especially for areas with similar tree species, providing fine classification results.

[0171] Example 2, as a second embodiment of the present application, provides an intelligent classification method for vegetation in power transmission corridors based on UNET and hyperspectral data. To verify the beneficial effects of the present application, scientific demonstration is carried out through experiments.

[0172] (1) The classification results of 9 common tree species based on the UNET neural network model are shown in Table 1, including the accuracy, precision, recall, and F1-score of each tree species.

[0173] Table 1 Classification results based on UNET neural network model

[0174]

[0175] Result analysis:

[0176] ① Accuracy: From the table, the overall accuracy of the model reached 91.7%, proving that the UNET model has high performance in the vegetation classification task of hyperspectral data.

[0177] ② Precision: The precision of the model for different tree species is above 90%, indicating that in most cases, the model can accurately predict the correct tree species.

[0178] ③ Recall: The recall rate is also excellent, with all categories exceeding 88%, indicating that the model has strong ability in predicting most tree species and no obvious missed classification phenomenon.

[0179] ④ F1-score: F1-score combines the effects of precision and recall, and performs well overall, especially in eucalyptus, Chinese balsa, and tung tree categories, with F1-score exceeding 94%.

[0180] (2) Post-processing and precision improvement experiment

[0181] The results of the classification using the aforementioned UNET neural network model are used, and the post-processing operations described above are applied. For the 9 common tree species (Eucalyptus, Camphor tree, Papermulberry, Blue oak, Oil camphor tree, Chinese cypress, Pine, Fir, and Bamboo), the following operations are applied:

[0182] Morphological operations (dilation and erosion), boundary smoothing (total variation denoising), CRF optimization, class smoothing, and misclassification correction.

[0183] ① Experimental results, the classification results after model post-processing and precision improvement are shown in Table 2:

[0184] Table 2 Classification results after model post-processing and precision improvement

[0185]

[0186] ② Results analysis

[0187] The post-processing operations significantly improve the accuracy, precision, recall, and F1-score of the classification. In particular, in the Blue oak, Bamboo, and Pine categories, the post-processing operations help to correct misclassification problems and improve the recall and precision.

[0188] Morphological operations and CRF optimization effectively reduce noise areas and improve the spatial consistency of the classification boundaries.

[0189] Class smoothing further improves the confusion between similar tree species, especially the misclassification problem between Blue oak and Pine.

[0190] Through post-processing operations such as morphological operations, boundary smoothing, CRF optimization, and class smoothing, the precision of the UNET neural network in the classification of 9 common tree species can be significantly improved. These post-processing techniques help to remove noise, correct misclassification, and improve the accuracy of the final classification results.

[0191] In summary, by implementing the UNET and hyperspectral-based intelligent classification method for vegetation in power transmission corridors, accurate classification of 9 common tree species in power transmission corridors can be achieved, and high accuracy can still be achieved in complex environments. This method effectively solves the problems of high dimensionality, noise interference, and insufficient spatial information extraction in traditional vegetation classification, and has significant advantages in multi-source data fusion, model optimization, and computational efficiency. By using multi-temporal remote sensing data and dynamic monitoring technology, combined with real-time processing capabilities, this method can provide efficient, real-time, and accurate technical support for the safety management of power transmission lines.

[0192] Embodiment 3, third embodiment of the present application, which is different from the previous two embodiments is:

[0193] The functions described can be implemented in software, firmware, hardware, or any combination thereof. If implemented in software and as an independent application, it can be stored in a one or more of computer-readable storage media described above. The functions described can be stored as one or more instructions and / or data on non-transitory computer-readable storage medium and executed by one or more processors of the computer device. The general structure, logic, and / or the sequence of the steps of the functions described can be embodied in software, firmware, hardware, or any combination thereof. The computer-readable storage medium can be a machine-readable storage medium such as a floppy diskette, optical disk, hard disk, solid-state drive, magnetic tape, or other storage medium, as those skilled in the art will recognize. The computer-readable storage medium can be a random access memory (RAM), a read-only memory (ROM), or another type of memory. The computer-readable storage medium can also be a local memory connected to the computer device, a removable computer-readable storage medium, or other computer-readable storage medium.

[0194] The logic and / or steps represented in the flowcharts and / or otherwise described herein, for example, can be embodied in computer-readable storage media, which can be any available media that can be accessed by a general purpose or special purpose computer system including the functional design of a computer system such as a general purpose or special purpose computer system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or any combination thereof. For purposes of this specification, a "computer-readable storage medium" can be any medium that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0195] More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires (electrical connections), a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable storage medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer. In some embodiments, the computer-readable storage medium can be a non-transitory computer-readable storage medium with a single physical instance or multiple physical instances.

[0196] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies, alone or in any combination, can be used: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

Claims

1. A smart classification method for vegetation along power transmission corridors based on UNET and hyperspectral imaging, characterized by: include, Hyperspectral images of the power transmission corridor area were acquired using hyperspectral remote sensing equipment, covering spectral information of different tree species. The acquired hyperspectral images were then preprocessed. Design and build the UNET neural network, which includes an encoder, decoder, and skip connections, for pixel-level classification tasks on hyperspectral data. Train the model using labeled tree category data and tune the model parameters through backpropagation and optimization algorithms. The UNET neural network model includes an encoder, a decoder, and skip connections. The input is hyperspectral image data. The encoder uses convolution kernels in the convolutional layer to extract local features from the input data. A 3x3 convolution kernel is selected. After each convolution operation, the number of convolution kernels increases layer by layer. The spatial size of the hyperspectral image is reduced by pooling operation. The decoder upsamples the image through a deconvolution layer, restores the feature map extracted by the encoder to the same spatial size as the input image, and obtains the category label of each pixel, i.e. the tree species, through pixel classification. At the same time, skip connections are used to combine the low-level features of the encoder with the high-level features of the decoder. The softmax activation function is used to convert the predicted value of each pixel into the category probability, and outputs a classification map, in which each pixel is labeled as one of the nine tree species. Based on the trained UNET model, classification is performed using newly collected test data, outputting the tree category to which each pixel belongs, evaluating the classification performance of the model based on the classification results, and calculating the evaluation index. Post-processing operations are performed, including morphological operations to remove noise in small regions, smoothing classification boundaries, and using conditional random fields to optimize the model output and eliminate confusion between categories. By employing misclassification correction strategies, the classification results of similar categories among different tree species are improved, thereby enhancing the overall classification accuracy. The misclassification correction strategy includes correcting misclassified areas, including dividing edge areas into low-contrast areas, by setting correction rules to correct the classification results of surrounding areas with similar categories to the correct category. The correction rules include, Rule 1: Based on the class consistency of the pixel neighborhood, if the class consistency of a pixel with the class of its surrounding pixels is higher than 85%, the current classification result is maintained; if the class consistency of a pixel with the class of its surrounding pixels is less than 85%, it is corrected. Rule 2: Based on the similarity of spectral features, if the similarity between the spectral features of a pixel and the typical features of any type of tree is higher than 70%, then category correction will be performed in the post-processing stage.

2. The intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral imaging as described in claim 1, characterized in that: The hyperspectral image includes selecting a power transmission channel area and using a hyperspectral remote sensing sensor to collect data on the vegetation in the current area. The spectral data collected by the hyperspectral sensor covers the electromagnetic spectrum from 40nm to 2500nm, a total of 210 bands, and each band has a spectral resolution of 10nm. For each tree, spectral data were collected several times from different angles and heights. The collected hyperspectral image data contained the spectral information of the specified tree species, and no fewer than the specified number of sample points were collected for each tree species. The preprocessing includes spectral data calibration and denoising, spectral feature extraction and enhancement, and data standardization; The spectral data calibration and denoising includes calibration and denoising in the order of radiometric calibration and noise removal; wherein, radiometric calibration is performed by comparing the measured results of the hyperspectral sensor with the reflectance of a known standard target to eliminate the influence of illumination changes and sensor characteristics on the data. Noise removal involves using discrete wavelet transform to denoise the acquired hyperspectral data, decomposing the signal into components of different frequencies, and removing high-frequency noise; representing the spectral data as a vector form. ,in, This represents the spectral value of the i-th band. Given the total number of bands, wavelet decomposition is performed on each spectral data to obtain low-frequency and high-frequency components. The high-frequency components are then thresholded to remove noise. The processed signal is then reconstructed to obtain the denoised spectral data. Noise removal also includes atmospheric correction, which uses atmospheric correction algorithms to perform atmospheric correction on the data to eliminate the influence of atmospheric composition on spectral reflectance; The spectral feature extraction and enhancement includes: band selection, using principal component analysis to select bands, performing mean subtraction and centering processing on the spectral data of each sample, calculating the covariance matrix of the sample data, finding the eigenvalues ​​and eigenvectors of the covariance matrix C, selecting the first k principal components, and projecting the original data onto these k principal components to obtain the dimensionality-reduced data. Feature enhancement: The spectral ratio method is used to enhance the spectrum of different bands. The vegetation features are emphasized by calculating the ratio of any band to the remaining bands. The reflectance features of vegetation are enhanced by the ratio of the red band and the near-infrared band. This ratio has significant differences in vegetation and non-vegetation areas. The data standardization refers to normalizing the spectral values ​​of each band.

3. The intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral imaging as described in claim 2, characterized in that: The design and construction of the UNET neural network includes designing the UNET neural network model and training the UNET neural network model. The training of the UNET neural network model includes using preprocessed hyperspectral data as input, calculating and outputting labeled vegetation type images through forward propagation, and dividing the dataset into 70% as the training set and 30% as the test set. The cross-entropy loss function is used to measure the difference between the model's predicted values ​​and the true labels. The classification error for each pixel is calculated, and the sum of the errors of all pixels is the loss for the entire image. : in, For real labels, To predict probabilities, Total number of pixels; The Adam optimizer is used to adjust and update the learning rate, with the following update rules: in, For parameters, For learning rate, and These are estimates of the first and second moments, respectively. To prevent division by zero of constants; Set training rules, set the number of samples in each training batch to 16, set the total number of training rounds to 100, and set a learning rate decay strategy, reducing the learning rate to 0.1 times its original value every 10 rounds; The gradient of the loss function with respect to each weight is calculated through backpropagation, and the weights and biases are updated using the gradient descent algorithm. At the end of each training round, the model is evaluated using test set data, and the model with the highest accuracy and generalization ability is selected.

4. The intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral imaging as described in claim 3, characterized in that: The classification results include vegetation classification tasks performed on the model after training the UNET neural network. In the classification task stage, the preprocessed test data is used as input. The hyperspectral data of each test sample is passed through the forward propagation of the network to obtain a pixel-level classification result, which is then fed into the trained UNET model for prediction. The output is the probability distribution of each pixel under the set 9 tree categories. Through the softmax activation function, each pixel is assigned to the category with the highest probability as the predicted label. For each test image, an output image of the same size as the input image is generated. Each pixel is classified into one of the set tree species. The predicted label and the true label are compared by a direct pixel-level classification matching method to determine whether the classification is correct. The probability of each pixel belonging to each category is calculated by the Softmax activation function, and the category with the highest probability is selected as the final classification label. The classification accuracy and evaluation metrics of the computational model include, Accuracy, the percentage of all correctly predicted pixels out of the total number of pixels; Precision is the percentage of true positives when any predicted class is positive. Recall rate is actually the proportion of cases where the current category is predicted to be positive, given any given category as the positive category. F1-score, the harmonic mean of precision and recall, comprehensively considers the performance of the classifier.

5. The intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral imaging as described in claim 4, characterized in that: The post-processing operations include morphological noise removal, smoothing of classification boundaries, and conditional random field optimization. The morphological noise removal includes performing morphological operations using a 3x3 or 5x5 structuring element, applying erosion operations to reduce noise in the region, restoring the shape of the target region through dilation operations, and removing small noise in the classification results by combining dilation and erosion operations while retaining the identified vegetation regions. The smoothing of classification boundaries includes denoising using a boundary smoothing algorithm, which smooths the boundaries by reducing abrupt changes in the image: in, It is a smoothed image. This is the original classification diagram. Represents the image gradient. It is a balancing term. is the total number of images, and 'a' is a point defined on an image.

6. The intelligent vegetation classification method for power transmission channels based on UNET and hyperspectral imaging as described in claim 5, characterized in that: The conditional random field optimization includes adjusting the classification boundary by considering the spatial and spectral consistency between pixels. : in, It outputs tags. It is the input image. Adjacent pixels and , and Similarity measure between These are weight parameters. is a normalization constant, and i and j are variable indices.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent classification method for vegetation in power transmission channels based on UNET and hyperspectral imaging as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent classification method for vegetation in power transmission channels based on UNET and hyperspectral imaging as described in any one of claims 1 to 6.

Images