Hyperspectral long-distance target positioning and denoising method based on double models
By adopting a dual-model denoising method in hyperspectral long-distance target positioning, using SAM to screen pixels and combine dynamic weight adjustment BP neural network and random forest model, the problem of serious noise interference is solved, and high-precision and robust long-distance target positioning is achieved.
Patent Information
- Application Number
- CN202510023531.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-07
- Publication Date
- 2025-05-30
AI Technical Summary
In hyperspectral long-distance target positioning tasks, noise interference is severe, and existing denoising methods are difficult to maintain accuracy and robustness while improving positioning accuracy.
Using a dual-model denoising method, pixels similar to the target spectrum are first filtered out by spectral angle matching (SAM) method, reducing data dimensions and reducing noise interference. Then, the screened data was deeply extracted and classified using dynamic weight-adjusted BP neural network and random forest model to remove residual noise and enhance the classification effect.
It significantly improves the accuracy and robustness of long-distance target positioning, and can effectively handle noise in hyperspectral data under medium computing resource configuration, and is suitable for applications in multiple scenarios.
Smart Images
Figure CN120071126A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of hyperspectral image processing, and particularly to a denoising method for long-distance target localization in hyperspectral images based on a dual model. Background Art
[0002] Hyperspectral imaging technology is a technology that integrates imaging and spectral analysis, featuring high spectral resolution and high spatial resolution. It has wide applications in fields such as long-distance target localization, environmental monitoring, agriculture, and medicine. However, due to the high dimensionality of hyperspectral data and noise interference under long-distance acquisition conditions, the processing of hyperspectral images faces numerous challenges, especially in long-distance target localization tasks over three kilometers, where the impact of noise is particularly significant.
[0003] To address such problems, the denoising methods in long-distance hyperspectral target localization can be mainly divided into the following three categories:
[0004] ① Traditional filtering denoising methods: Such methods preprocess and filter hyperspectral data, such as mean filtering, median filtering, etc., to remove noise. The methods are simple and direct, but are prone to losing details while retaining effective information, affecting the localization accuracy.
[0005] ② Denoising methods based on statistical models: Such methods establish noise models and use statistical methods to model and remove noise, such as principal component analysis (PCA) and other methods. Although such methods can effectively reduce some noise, the noise processing effect for complex backgrounds or high-dimensional data is limited.
[0006] ③ Denoising methods based on machine learning: In recent years, deep learning methods have gradually been applied to the denoising and localization tasks of hyperspectral images. Although neural network models perform excellently in denoising and feature extraction, deep learning methods usually require a large amount of training data and high-performance computing resources, and there are certain robustness and interpretability problems in practical applications.
[0007] In summary, in this field, there is an urgent need for a localization denoising method for long-distance hyperspectral targets that can improve the localization accuracy while taking into account the accuracy and robustness of long-distance target localization. This method has high practical value and promotion potential. Summary of the Invention
[0008] Objective of the Invention: To solve the problems mentioned in the background art, the present invention discloses a denoising method for hyperspectral long-distance target localization based on a dual model. The method uses the SAM method to screen out pixels similar to the target spectrum, reduces the data dimension and noise interference, and uses a dual model with dynamic weight adjustment to deeply extract features and classify the hyperspectral data screened by SAM, removing residual noise and enhancing the classification effect. This method can not only effectively process the noise in hyperspectral data under medium computing resource configurations, but also significantly improve the accuracy and robustness of long-distance target localization.
[0009] Technical Solution:
[0010] The present invention discloses a denoising method for hyperspectral long-distance target localization based on a dual model, and the method comprises the following steps:
[0011] S1: Use a hyperspectral spectrometer to collect hyperspectral data of a long-distance portrait as a sample;
[0012] S2: Use a data processing library to load the collected hyperspectral data, obtain complete hyperspectral image information, load HDR and SPE files, and obtain the band information of the hyperspectral image; Mark the portrait data as 1 and other data as 0 to construct a dataset for a binary classification problem;
[0013] S3: Perform block processing on the loaded hyperspectral image row by row, and normalize the spectral data of each pixel; Use the Spectral Angle Matcher (SAM) method to calculate the spectral angle between each pixel and the target average spectrum, and screen out pixels similar to the target spectrum;
[0014] S4: For the pixel data screened by SAM, use a BP neural network model for feature extraction and prediction to initially complete classification;
[0015] For the pixel data screened by SAM, use a random forest model for classification to obtain the prediction results of the random forest;
[0016] S5: Based on dynamic weight adjustment, perform weighted fusion on the prediction results of the Bp neural network and the random forest to obtain the denoised hyperspectral data of the long-distance portrait.
[0017] Further, the specific process of S1 for collecting hyperspectral data is as follows:
[0018] Use a hyperspectral spectrometer covering the spectral band of 400 - 1000 nm for data collection. The dispersion width of the instrument on the detector target surface is 6.9 mm, the detector width is 7.03 mm, to obtain the hyperspectral data of the long-distance portrait. The image covers 300 spectral bands, capturing detailed information of the target under different spectra. The size of the collected hyperspectral image is 1920*1920*300, and it is saved as an ENVI format file.
[0019] Furthermore, the specific steps of S3 are as follows:
[0020] S301 Calculate the average spectrum of the target category samples and normalize it to obtain the normalized vector of the target average spectrum;
[0021] S302 Normalize the spectral vector of each pixel in the hyperspectral image;
[0022] S303 Use the Spectral Angle Matcher (SAM) method to calculate the cosine similarity between the spectral vector of each pixel and the target average spectrum, and obtain the spectral angle through the inverse cosine function;
[0023] S304 Set the upper and lower threshold values of the spectral angle, and filter out the pixels similar to the target spectrum according to the threshold to generate a SAM mask, marking the pixel positions that meet the conditions;
[0024] Furthermore, the training and execution processes of the S4 BP neural network model are as follows:
[0025] S401 Use the multi-layer perceptron structure of the BP neural network to perform feature extraction and prediction on the pixel data screened by SAM. The BP neural network contains a hidden layer with 400 neurons;
[0026] S402 Adopt the ReLU activation function and train the network through the Adam optimizer, with a maximum number of iterations of 500 times;
[0027] S403 Use the trained BP neural network model to calculate the prediction probability that the pixels screened by SAM belong to the target category;
[0028] Furthermore, the training and execution processes of the S4 random forest model are as follows:
[0029] S404 Use the random forest model to process the pixel data screened by SAM. The random forest contains 100 decision trees, which are used to enhance the classification ability of features and suppress noise;
[0030] S405 Use the trained random forest model to calculate the prediction probability that the pixels screened by SAM belong to the target category;
[0031] S406 Compare the output of the random forest with the output of the BP neural network as the basis for subsequent fusion judgment.
[0032] Furthermore, the specific steps of S5 are as follows:
[0033] S501 Calculate the initial weights according to the performance of the BP neural network and the random forest on the validation set and normalize them;
[0034] For each pixel, according to the prediction probabilities of the BP neural network and the random forest, calculate their respective prediction uncertainties. The closer the prediction probability is to 0.5, the higher the uncertainty;
[0035] Combine the global initial weights and the local prediction uncertainties to dynamically adjust the weights of the BP neural network and the random forest;
[0036] Normalize the adjusted weights to ensure that the sum of the weights is 1;
[0037] Use the dynamically adjusted weights to perform weighted averaging on the prediction probabilities of the BP neural network and the random forest to obtain the fused prediction probability. Convert the probability to the final class label; for the pixels predicted as the target class, retain their corresponding spectral information; for the pixels of non-target classes, set their spectral information to zero to achieve effective denoising and accurate positioning of the hyperspectral image.
[0038] Beneficial effects:
[0039] 1. The present invention uses the SAM method to screen out the pixels similar to the target spectrum. Compared with the conventional machine learning method for processing data sets, on the one hand, it can further reduce the data dimension and reduce noise interference, improving the positioning accuracy. On the other hand, in the field of hyperspectral denoising and recognition, it is applicable to more types of image selection, has good generalization ability, and further improves the adaptability of the method in multi-scene applications.
[0040] 2. The present invention classifies and extracts features from the hyperspectral data through the parallel mode of dual models. Compared with the conventional machine learning mode of a single model, when distinguishing different ground object categories, it shows higher robustness and can effectively process hyperspectral data under medium computing resource conditions, further saving computing power and reducing costs.
[0041] 3. The present invention effectively reduces the misrecognition rate and significantly improves the accuracy and robustness of long-distance human portrait positioning through a dynamic weight adjustment mechanism, comprehensively considering the prediction confidence of the two models. In application scenarios such as long-distance target positioning, environmental monitoring, and medical image analysis, the method of the present invention can achieve high denoising and classification effects. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 is the flowchart of the method of the present invention;
[0043] Figure 2 a is the original image of the embodiment of the present invention, Figure 2 b is the prediction result of the dual models of the embodiment of the present invention;
[0044] Figure 3a is the denoising result graph of a single model in the embodiment of the present invention, Figure 3 b is the denoising result graph of a dual model in the embodiment of the present invention. Detailed implementation manners
[0045] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0046] As Figure 1 shown, the present invention discloses a hyperspectral long-distance target positioning and denoising method based on a dual model, which includes the following steps:
[0047] S1 Collect hyperspectral data of a long-distance portrait using a hyperspectral spectrometer as a sample;
[0048] S2 Load the collected hyperspectral data using a data processing library to obtain complete hyperspectral image information, load HDR and SPE files to obtain the band information of the hyperspectral image; mark the portrait data as 1 and other data as 0 to construct a dataset for a binary classification problem;
[0049] S3 Perform block processing on the loaded hyperspectral image row by row, and perform normalization processing on the spectral data of each pixel; use the Spectral Angle Matcher (SAM) method to calculate the spectral angle between each pixel and the target average spectrum, and screen out the pixels similar to the target spectrum. The specific steps are as follows:
[0050] S301 Calculate the average spectrum of the target category samples and perform normalization processing on it to obtain the standardized vector of the target average spectrum.
[0051] S302 Perform normalization processing on the spectral vector of each pixel in the hyperspectral image.
[0052] S303 Use the Spectral Angle Matcher (SAM) method to calculate the cosine similarity between the spectral vector of each pixel and the target average spectrum, and obtain the spectral angle through the inverse cosine function.
[0053] S304 Set the upper and lower threshold values of the spectral angle, and screen out the pixels similar to the target spectrum according to the threshold values to generate a SAM mask and mark the pixel positions that meet the conditions.
[0054] For the pixel data screened by SAM, use the BP neural network model to extract features and make predictions, and initially complete classification. For the pixel data screened by SAM, use the random forest model to perform classification to obtain the prediction results of the random forest. The specific steps are as follows:
[0055] S401 Use the multi-layer perceptron structure of the BP neural network to extract features and make predictions on the pixel data screened by SAM. The BP neural network contains a hidden layer, and the number of neurons in the hidden layer is 400;
[0056] S402 Adopt the ReLU activation function and train the network through the Adam optimizer. The maximum number of iterations is 500 times;
[0057] S403 Use the trained BP neural network model to calculate the prediction probability that the pixels screened by SAM belong to the target category.
[0058] S5 For the prediction results of the Bp neural network and the random forest, perform weighted fusion based on dynamically adjusted weights to obtain the denoised long-distance portrait hyperspectral data. The specific steps are as follows:
[0059] S501 Calculate the initial weights according to the performance of the BP neural network and the random forest on the validation set and normalize them;
[0060] S502 For each pixel, calculate their respective prediction uncertainties according to the prediction probabilities of the BP neural network and the random forest. The closer the prediction probability is to 0.5, the higher the uncertainty;
[0061] S503 Combine the global initial weights and local prediction uncertainties to dynamically adjust the weights of the BP neural network and the random forest.
[0062] S504 Normalize the adjusted weights to ensure that the sum of the weights is 1.
[0063] S505 Use the dynamically adjusted weights to perform weighted averaging on the prediction probabilities of the BP neural network and the random forest to obtain the fused prediction probability. Convert the probability to the final class label; for the pixels predicted as the target category, retain their corresponding spectral information; for the pixels of non-target categories, set their spectral information to zero to achieve effective denoising and accurate positioning of the hyperspectral image.
[0064] The specific implementation process of the embodiment of the present invention is as follows:
[0065] 1) Data acquisition
[0066] First, hyperspectral data is collected using a hyperspectral spectrometer that covers the spectral band of 400 - 1000nm. The dispersion width of the instrument on the detector target surface is 6.9mm, and the detector width is 7.03mm. Hyperspectral data of long - distance human portraits is obtained in Huai'an. These images cover 300 spectral bands, capturing detailed information of the target under different spectra. The size of the collected hyperspectral images is 1920 * 1920 * 300, and they are saved as ENVI format files (including.hdr and.spe files) for subsequent data loading and processing.
[0067] 2) Data Processing and Loading
[0068] The hyperspectral data of 40 long - distance human portraits obtained in Huai'an is processed. The pixels corresponding to the human portrait data are extracted and labeled as 1 (target class), and other background data is labeled as 0 (non - target class) to construct a binary classification dataset for training and testing. The Spectral library is used to load the hyperspectral image files, and image data containing multi - band information is obtained for later verification.
[0069] 3) Spectral Angle Matcher (SAM) Filtering
[0070] The Spectral Angle Matcher (SAM) method is used to perform preliminary filtering on the hyperspectral data. Specifically, the spectral angle between each pixel and the target average spectrum is calculated, and pixels similar to the target spectrum are selected to reduce noise interference and data dimension. This step effectively retains the spectral information related to the target and improves the accuracy of subsequent classification.
[0071] 4) BP Neural Network Feature Extraction
[0072] In this embodiment, a BP neural network is used to extract features from the data. The specific steps are as follows:
[0073] (1) Network Structure: The BP neural network consists of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the number of bands of the hyperspectral data. The hidden layer contains 400 neurons and uses the ReLU activation function; the output layer contains 1 node and uses the Sigmoid activation function to achieve the binary classification task.
[0074] (2) Forward Propagation:
[0075] Calculation from the input layer to the hidden layer:
[0076] h = ReLU(W 1 x + b 1 )
[0077] where x is the input feature vector, W 1 is the weight matrix from the input layer to the hidden layer, and b 1is the bias vector, and ReLU(z) = max(0, z) is the activation function.
[0078] Output layer calculation:
[0079]
[0080] where W 2 is the weight matrix from the input layer to the hidden layer, b 2 is the bias vector, is the Sigmoid activation function, is the final output.
[0081] (3) Loss function:
[0082] The binary cross-entropy loss function is used to measure the error between the predicted value and the true label:
[0083]
[0084] In the above formula, m is the number of samples, y (i) is the true label of the i-th sample, is the predicted output of the model.
[0085] (4) Backpropagation and weight update:
[0086] During the backpropagation process, the gradients of the loss function with respect to the weights of each layer need to be calculated using the chain rule:
[0087] Output layer error term:
[0088]
[0089] Hidden layer error term:
[0090] δ h idden = (δ output W 2 ) ⊙ ReLU'(W 1 x + b 1 )
[0091] where ⊙ represents element-wise multiplication (multiplication of each element), and ReLU′(z) = 1 when z > 0, otherwise 0.
[0092] Weight update:
[0093] The gradient descent method is used for update, and the update formula for the weights is as follows:
[0094]
[0095] where η is the learning rate, and are the partial derivatives of the loss function with respect to the weight W1 and W 2 gradient of
[0096] 5) Random Forest Refinement
[0097] To enhance the robustness and accuracy of feature extraction, a random forest classifier is introduced for refinement using different feature dimensions.
[0098] (1) Random Forest Structure:
[0099] The random forest consists of 100 decision trees. Each tree is trained using a random subset of samples during training to enhance the diversity and generalization ability of the model.
[0100] (2) Node Splitting Criterion:
[0101] Gini Impurity is used to select the best splitting feature, and its formula is as follows:
[0102]
[0103] where K is the number of classes, and p k is the proportion of samples of the k-th class in the current node.
[0104] (3) Information Gain:
[0105] When splitting each node, the feature that maximizes the information gain is selected as the splitting point. The formula for information gain is:
[0106]
[0107] where: H(T) is the Gini impurity of node T, T ν is the child node when the feature X takes the value ν, and |T| is the number of samples in node T.
[0108] 6) Model Fusion and Prediction
[0109] To improve the classification accuracy of the model, a decision fusion strategy of BP neural network and random forest is adopted, combined with a dynamic weight adjustment mechanism for final prediction.
[0110] (1) Independent Prediction:
[0111] Use the BP neural network to predict the test data and obtain the prediction result P BP .
[0112] Use the random forest to predict the test data and obtain the prediction result P RF .
[0113] (2) Result Fusion:
[0114] Through a dynamic weight adjustment mechanism, the output probabilities of each model are weighted and fused according to the prediction confidence of each model to generate a comprehensive prediction result. The specific formula is as follows:
[0115] Combined_Prob = ω BP ×P BP + ω RF ×P RF
[0116] Where P BP represents the prediction result of the BP neural network, P RF represents the prediction result of the random forest, ω BP and ω RF are the dynamically adjusted weights, and ω BP + ω RF = 1.
[0117] Finally, if the combined probability Combined_Prob exceeds the set threshold (e.g., 0.5), then the data point is marked as the target category. In this way, by integrating the advantages of the two models, the misrecognition rate can be effectively reduced, and the overall recognition accuracy and robustness can be improved.
[0118] Model Evaluation
[0119] In this embodiment, the classification performances of two schemes based on the BP neural network and the dual model for hyperspectral long-distance portrait data are compared. By comparing their respective accuracies and confusion matrices, the performances of the two models in different aspects can be comprehensively evaluated, and the advantages of the dual model can be highlighted, as shown in Table 1.
[0120] Table 1
[0121]
[0122] The verification accuracy of the dual model (BP neural network + random forest) is 99.91%, higher than 99.83% of the single model (BP neural network). Although this improvement seems small, it is significant when dealing with a large test set (the total number of samples is 3,686,400 spectra). More importantly, the dual model reduces the number of false positives (FP) from 6,061 of the single model to 3,218, almost halving it. This significant noise reduction is of great significance for the system reliability and user experience in practical applications, especially in fields such as security monitoring that require high accuracy and low false alarm rates.
[0123] The verification precision of the dual model has increased from 0.00115 of the single model to 0.00279, and the recall rate has also increased from 0.4667 to 0.6000. This indicates that the dual model has higher accuracy when predicting the target class, significantly reducing the false positive rate. At the same time, it can identify more actual target class samples, reducing the situation of false negatives (FN). Specifically, the number of false negatives of the dual model has been reduced from 8 to 6, further improving the coverage and reliability of the model in the target recognition task.
[0124] The dual model is superior to the single model in terms of accuracy, precision, and recall rate, especially showing significant advantages in reducing false positives and improving target recognition ability. By combining the advantages of the BP neural network and random forest, the dual model achieves higher prediction accuracy and a wider target detection coverage, significantly reducing the number of false positives. This makes the dual model more reliable in practical applications.
[0125] In summary, the dual model (BP neural network + random forest) shows a significant improvement compared to the single model (BP neural network) in multiple key metrics, especially in terms of precision and recall rate. Therefore, in the hyperspectral data processing task of long-distance target localization, the combination of the dual model is an efficient and reliable solution.
[0126] Hyperspectral Image Processing and Visualization
[0127] In this embodiment, the hyperspectral image is processed row by row to make full use of the spectral information of each pixel point. The size of the processed hyperspectral image is 1920×1920×300, containing 3,686,400 pixel points, and each pixel point contains spectral features of 300 bands.
[0128] When processing the hyperspectral image row by row, each row contains 1920 pixel points. First, numerical screening of the SAM spectral angle is performed, and the spectral features of the screened pixel data are sequentially input into the BP neural network and random forest model for prediction. A dynamic weight fusion strategy is adopted, that is, according to the performance and prediction uncertainty of the two models on the validation set, their weights are dynamically adjusted, and the prediction probabilities are weighted and fused. The final prediction result is based on the weighted comprehensive probability, thereby effectively improving the detection ability of the model, reducing misjudgment, and ensuring more accurate marking of the target class.
[0129] As Figure 2As shown, in order to more intuitively verify the denoising and localization effects of the model, in this embodiment, a visual comparison is made of the original image, the BP neural network model, and the prediction results of the dual model; first, the comparison between the original image and the prediction results of the dual model is shown. The original image is displayed in grayscale, and the prediction results of the dual model are shown in pseudo-color, with the target pixel points marked in red and the non-target pixel points marked in blue.
[0130] As Figure 3 shown, the comparison between the prediction results of the BP neural network and the dual model is shown. The BP model has more false alarms, while the dual model effectively reduces false alarms, showing better accuracy and robustness. The prediction results of the dual model use the same color standard to facilitate the comparison of the differences in recognition accuracy and robustness between the two models.
[0131] The experimental results show that the present invention exhibits excellent denoising effects and classification performance in the hyperspectral long-distance target localization task. Through the feature extraction of the BP neural network and the multi-dimensional feature classification of the random forest, the method of the present invention demonstrates remarkable robustness and accuracy in the denoising and classification tasks.
[0132] The above description of the embodiments enables those skilled in the art to implement or use the present invention. Various modifications to the embodiments will be obvious to those skilled in the art. The general principles of the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should cover the widest scope that conforms to the principles and novel features disclosed in the present invention.
Claims
1. A method for denoising a hyperspectral long-distance target positioning based on a dual model, characterized in that: The method comprises the following steps: S1 uses a hyperspectral instrument to collect hyperspectral data of distant portraits as samples; S2 uses the data processing library to load the collected hyperspectral data, obtain the complete hyperspectral image information, load HDR and SPE files, and obtain the band information of the hyperspectral image; mark the portrait data as 1 and other data as 0 to construct a data set for the binary classification problem; S3 processes the loaded hyperspectral image row by row and performs normalization on the spectral data of each pixel; it uses the spectral angle matching (SAM) method to calculate the spectral angle between each pixel and the target average spectrum, and screens out pixels similar to the target spectrum; S4 uses the BP neural network model to extract and predict features of the pixel data filtered by SAM and preliminarily completes the classification; For the pixel data filtered by SAM, the random forest model is used for classification to obtain the prediction results of the random forest; S5 performs weighted fusion on the prediction results of the Bp neural network and random forest based on dynamically adjusted weights to obtain denoised long-distance portrait hyperspectral data.
2. The method for denoising a hyperspectral long-distance target positioning based on a dual model according to claim 1, characterized in that: The specific process of S1 collecting hyperspectral data is as follows: A hyperspectral spectrometer covering the 400-1000nm spectral band is used for data collection. The dispersion width of the instrument on the detector target surface is 6.9mm, and the detector width is 7.03mm. Hyperspectral data of long-distance portraits are obtained. The image covers 300 spectral bands and captures detailed information of the target under different spectra. The size of the collected hyperspectral image is 1920*1920*300 and is saved as an ENVI format file.
3. The method for denoising a hyperspectral long-distance target positioning based on dual models according to claim 1 is characterized in that: The specific steps of S3 are as follows: S301 calculates the average spectrum of target category samples and performs normalization processing on it to obtain a standardized vector of the target average spectrum; S302 normalizes the spectral vector of each pixel in the hyperspectral image; S303 uses the spectral angle matching SAM method to calculate the cosine similarity between the spectral vector of each pixel and the target average spectrum, and obtains the spectral angle through the arccosine function; S304 sets the upper and lower thresholds of the spectral angle, selects pixels similar to the target spectrum according to the thresholds, generates a SAM mask, and marks the pixel positions that meet the conditions.
4. The method for denoising a hyperspectral long-distance target positioning based on dual models according to claim 1, characterized in that: The S4BP neural network model training and execution process is as follows: S401 uses the multi-layer perceptron structure of the BP neural network to extract and predict the features of the pixel data screened by SAM. The BP neural network contains one hidden layer with 400 neurons. S402 uses the ReLU activation function and trains the network using the Adam optimizer with a maximum number of iterations of 500; S403 uses the trained BP neural network model to calculate the predicted probability that the pixels screened by SAM belong to the target category.
5. The method for denoising a hyperspectral long-distance target positioning based on dual models according to claim 1 or 4, characterized in that: The S4 random forest model training and execution process is as follows: S404 uses a random forest model to process pixel data filtered by SAM; the random forest contains 100 decision trees to enhance the classification ability of features and suppress noise; S405 uses the trained random forest model to calculate the predicted probability that the pixels screened by SAM belong to the target category; S406 compares the output of the random forest with the output of the BP neural network as a basis for subsequent fusion judgment.
6. The method for denoising a hyperspectral long-distance target positioning based on dual models according to claim 1, characterized in that: The specific steps of S5 are as follows: S501 calculates the initial weights and normalizes them according to the performance of the BP neural network and the random forest on the validation set; S502 For each pixel, according to the prediction probabilities of the BP neural network and the random forest, the respective prediction uncertainties are calculated. The closer the prediction probability is to 0.5, the higher the uncertainty is. S503 dynamically adjusts the weights of the BP neural network and the random forest by combining the global initial weights and the local prediction uncertainty; S504 performs normalization processing on the adjusted weights to ensure that the sum of the weights is 1; S505 uses the dynamically adjusted weights to perform weighted averaging on the prediction probabilities of the BP neural network and the random forest, and obtains the fused prediction probability, which is converted into the final category label; for pixels predicted as the target category, their corresponding spectral information is retained; for pixels of the non-target category, their spectral information is set to zero, thereby achieving effective denoising and precise positioning of the hyperspectral image.