Heterogeneous feature suitability-based hyperspectral adaptive ensemble classification method
By extracting multi-dimensional features and constructing an adaptive feature-classifier matching mechanism, the problem of insufficient adaptability between heterogeneous features and classifiers in hyperspectral image classification is solved, achieving high-precision and stable classification results and enhancing cross-scene applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HAINAN NORMAL UNIV
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-01
AI Technical Summary
Existing hyperspectral image classification methods have significant limitations in terms of heterogeneous features and classifier adaptability. Especially in real-world application scenarios where labeled samples are scarce, it is difficult to effectively utilize heterogeneous features, resulting in insufficient classification accuracy, poor model stability, and weak cross-scenario generalization ability.
An initial heterogeneous feature set is formed by extracting spectral reflectance features, spatial neighborhood features, and transform domain features. A dedicated feature saliency mapping space is constructed. The inter-source adaptation weights are dynamically calculated based on the signal-to-noise ratio and label sample density. A multi-classifier prototype library is constructed, and feature sensitivity fingerprints are assigned to each classifier. Adaptive matching between features and classifiers is performed, and finally, weighted integration is carried out through the inter-source adaptation weights.
It improves the accuracy and stability of hyperspectral image classification, enhances the generalization ability across scenes, solves the problems of redundant interference and information conflict in the process of heterogeneous feature fusion, and improves the adaptability of the classifier.
Smart Images
Figure CN121962764A_ABST
Abstract
Description
Hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation Technical Field
[0001] This invention relates to the field of image processing and pattern recognition technology, and more specifically, to a hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation. Background Technology
[0002] With the rapid development of remote sensing technology and spectral imaging equipment, hyperspectral images, with their rich spectral information and sophisticated ground feature identification capabilities, have become important data support for fields such as precision agriculture monitoring, ecological environment assessment, resource exploration, and urban planning. In complex ground feature classification tasks, accurately and efficiently extracting and utilizing effective information from hyperspectral data is of great significance for improving identification accuracy and reducing the risk of misjudgment. In order to fully explore the potential value of hyperspectral data, in recent years, feature selection and classification methods based on machine learning have been widely used. These methods extract various types of features such as spectral, texture, and morphology from the original spectral data and combine them with classifiers for pattern recognition, which has improved the accuracy and automation level of ground feature classification to a certain extent.
[0003] Existing hyperspectral image classification methods suffer from significant limitations in adapting to heterogeneous features and classifiers, a problem particularly pronounced in real-world applications where labeled samples are scarce. Specifically, hyperspectral data inherently possesses high spectral dimensionality, severe band redundancy, and difficulty in obtaining effectively labeled samples. When classification requires the simultaneous use of heterogeneous features from different data sources, the optimal feature subsets from each data source often exhibit significant differences, and the distribution characteristics and discriminative abilities of different feature types show significant inconsistencies. These inherent differences in heterogeneous features not only make it difficult to comprehensively measure feature effectiveness from a single data source but also introduce redundant interference and information conflicts during feature fusion, leading to… The current mainstream approach to address this issue is to use a unified feature selection criterion combined with a fixed classifier architecture. Features are selected using a single evaluation metric and fed into a pre-defined classification model. However, this approach ignores the essential differences between heterogeneous features and the dynamic changes in data source quality. It fails to establish an adaptive matching mechanism between feature subsets and classifiers. Furthermore, under limited label samples, traditional feature selection models struggle to effectively expand the training sample size to support reliable feature evaluation. This results in problems such as insufficient classification accuracy, poor model stability, and weak cross-scene generalization ability, ultimately limiting the widespread application of hyperspectral image classification technology in accurate recognition tasks.
[0004] In view of this, the present invention proposes a hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptability to solve the above problems. Summary of the Invention
[0005] To overcome the aforementioned deficiencies of existing technologies and achieve the above objectives, this invention provides the following technical solution: a hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptability, comprising: Step S1: Simultaneously extracting spectral reflectance features, spatial neighborhood features, and transform domain features from a hyperspectral image to form an initial heterogeneous feature set; Step S2: Constructing a dedicated feature saliency mapping space for each type of feature, and generating candidate feature subsets through band correlation suppression and inter-class dispersion enhancement; Step S3: Dynamically calculating inter-source adaptation weights based on the signal-to-noise ratio statistical characteristics and label sample density of each data source, and performing weighted cross-projection on the candidate feature subsets to obtain a global fusion feature candidate pool; Step S4: Constructing a multi-classifier prototype library and assigning feature sensitivity fingerprints to each classifier, and generating classifier-specific feature subsets by matching the adaptability of the feature sensitivity fingerprints with the features in the global fusion feature candidate pool; Step S5: Inputting each classifier-specific feature subset into the corresponding classifier for independent training and outputting class probabilities, and adaptively weighting and integrating the class probabilities of all classifiers through inter-source adaptation weights to obtain the final land cover category determination result.
[0006] Furthermore, step S1 includes: directly using the spectral curve of each pixel in the hyperspectral image as the spectral reflectance feature; constructing local windows of multiple sizes centered on the pixel to extract the gray-level co-occurrence matrix parameters as spatial neighborhood features; and taking the first few principal components as transform domain features after performing minimum noise separation transformation on the pixel spectral vector.
[0007] Furthermore, step S2 includes: calculating the inter-band Pearson correlation coefficient matrix within each feature class and setting a dynamic correlation threshold; clustering and merging band pairs that exceed the dynamic correlation threshold to obtain band clusters; calculating the improved Fisher inter-class dispersion to intra-class clustering ratio as a significance score within the merged band clusters; sorting the significance scores from high to low and retaining them in a fixed proportion to form a unique subset of candidate features.
[0008] Furthermore, step S3 includes: calculating the mean pixel signal-to-noise ratio and the proportion of available label samples corresponding to each type of feature to form a quality evaluation vector; normalizing the quality evaluation vector and using it as the inter-source adaptation weight; performing orthogonal projection stitching on each candidate feature subset under the guidance of the inter-source adaptation weight, and performing redundant feature diagonalization removal on the stitched features to obtain a single global fusion feature candidate pool.
[0009] Furthermore, step S4 includes: pre-constructing a multi-classifier prototype library containing support vector machines, random forests, extreme learning machines, and one-dimensional convolutional networks; conducting feature perturbation experiments on multiple public datasets for each classifier, recording the sensitivity coefficients of various feature changes on classification accuracy, and forming feature sensitivity fingerprints; calculating the cosine similarity between each feature in the global fusion feature candidate pool and the feature sensitivity fingerprints of each classifier; and dynamically selecting a fixed proportion of features for each classifier according to the similarity from high to low, forming a classifier's exclusive feature subset.
[0010] Furthermore, step S5 includes: inputting the specific feature subsets of each classifier into the corresponding classifier for forward inference. Forward inference includes a training phase and an inference phase. In the training phase, the four classifiers are trained independently using labeled samples. The training data is divided into training data and validation data according to the proportion of labeled samples. In the inference phase, the specific feature subsets of the pixels to be classified are input into the corresponding classifier, and the probability distribution of the pixel belonging to each category is output as the independent probability of each land cover category, and the probability distribution satisfies the category probability condition. The source-to-source adaptation weights of each classifier are used as the ensemble confidence of the classifier. The probability outputs of all classifiers are weighted, summed, and normalized according to the ensemble confidence, and the category with the highest category probability is taken as the final land cover category determination result.
[0011] Furthermore, the method of conducting feature perturbation experiments includes: adding Gaussian noise to each feature in the training set and recording the decrease in classification accuracy, and using the normalized reciprocal of the decrease as the sensitivity coefficient of the classifier to that feature.
[0012] Furthermore, the dynamic correlation threshold is determined by the product of the logarithmic function of the current feature dimension and the inverse of the number of labeled samples.
[0013] Furthermore, multiple local windows of different sizes are constructed centered on pixels to extract gray-level co-occurrence matrix parameters as spatial neighborhood features. This includes: simultaneously constructing three nested square local windows with a side length of 2k+1 centered on pixels, where k takes values of 1, 2, and 3; performing a minimum noise separation transformation on the pixel spectral vector within each window, and taking the three components with the highest signal-to-noise ratio after the transformation to form three single-band gray-level images; calculating gray-level co-occurrence matrices with step sizes of 1 and 2 along the horizontal and vertical orthogonal directions on each single-band gray-level image; and extracting gray-level co-occurrence matrix parameters from each co-occurrence matrix. Three texture statistics—contrast, entropy, and inverse difference moment—are extracted from the array. Redundancy is suppressed by differential extraction of the same texture statistics with the same component, the same direction, and different timings. Max pooling is used to retain the most significant spatial patterns of the same texture statistics with the same component but different directions. The processed texture statistics are weighted and fused using variable weight coefficients based on local label density adaptation. The weights decrease nonlinearly as the number of known label samples around the current pixel increases. Finally, the fused compact texture feature vector is appended with spatial scale and component source labels and used as spatial neighborhood features.
[0014] Furthermore, each candidate feature subset is orthogonally projected and stitched together under the guidance of inter-source adaptation weights. Redundant features are then diagonally eliminated from the stitched features to obtain a single global fusion feature candidate pool. This includes: assigning corresponding inter-source adaptation weights to the candidate feature subsets of spectral reflectance features, spatial neighborhood features, and transform domain features; performing weighted principal component analysis with the inter-source adaptation weights as diagonal elements on each candidate feature subset to obtain three weighted orthogonal projection matrices; and projecting each candidate feature subset onto its respective weighted principal component space, retaining the cumulative... The principal components whose variance contribution rate reaches the preset contribution value form the dimension-reduced exclusive feature matrix; using the current inter-source adaptation weight as the scaling factor, the amplitude of each dimension-reduced exclusive feature matrix is modulated, and Gram-Schmidt orthogonalization process is performed sequentially between the modulated feature matrices to make the later-accessed feature matrix orthogonal to the concatenated part; the cubically orthogonalized feature matrices are horizontally concatenated in a fixed order of spectral-spatial-transform domain, and the highly correlated feature pairs that appear after concatenation are removed by quadratic orthogonalization, and finally a single global fusion feature candidate pool is output.
[0015] The technical effects and advantages of this invention's hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptability are as follows: This invention simultaneously extracts spectral reflectance features, spatial neighborhood features, and transform domain features to form an initial heterogeneous feature set, comprehensively capturing the information representation capabilities of different dimensions in hyperspectral images and overcoming the limitation of insufficient information coverage from a single feature source; a dedicated feature saliency mapping space is constructed for each type of feature, eliminating redundant interference within features from the same source through band correlation suppression, and enhancing the class discrimination ability of features through inter-class dispersion enhancement, ensuring that the candidate feature subsets of each source retain highly significant and effective features, avoiding information conflicts caused by inconsistent heterogeneous feature distribution characteristics; the inter-source adaptation weights are dynamically calculated based on the signal-to-noise ratio statistical characteristics and label sample density of each data source, adaptively adjusting the contribution ratio of each source in the feature fusion process according to the actual differences in data source quality, and performing weighted cross-projection on the candidate feature subsets and executing... Redundant features are diagonally removed to obtain a global fusion feature candidate pool, solving the problem that a single data source cannot comprehensively measure the effectiveness of features, while reducing the risk of introducing redundant interference during feature fusion. A multi-classifier prototype library is constructed and a feature sensitivity fingerprint is assigned to each classifier. By matching the feature sensitivity fingerprint with the fit of each feature in the global fusion feature candidate pool, a classifier-specific feature subset is generated. An adaptive matching mechanism between the feature subset and the classifier is established, enabling different types of classifiers to obtain the feature combination that best matches their discriminative characteristics, overcoming the defect of fixed classifier architecture ignoring the essential differences of heterogeneous features. The class probabilities of all classifiers are adaptively weighted and integrated through inter-source adaptation weights to obtain the final land cover category determination result. This comprehensively leverages the complementary advantages of multiple classifiers, effectively improving classification accuracy and model stability under conditions of scarce label samples, and enhancing the cross-scene generalization ability of hyperspectral image classification methods. Attached Figure Description
[0016] Figure 1 is a schematic diagram of the hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation of the present invention. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] Example 1: As shown in Figure 1, this example is a hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptability, which includes: Step S1: Simultaneously extracting spectral reflectance features, spatial neighborhood features and transform domain features from the hyperspectral image to form an initial heterogeneous feature set.
[0019] Hyperspectral images are typically acquired by imaging spectrometers mounted on remote sensing satellites or airborne platforms. Each pixel contains spectral reflectance values across tens to hundreds of consecutive bands, forming rich spectral curve information. In applications such as precision agriculture monitoring and ecological environment assessment, relying solely on a single type of feature is insufficient to fully characterize the complex attributes of ground features. Therefore, it is necessary to extract heterogeneous features from multiple dimensions to enhance classification discriminative capabilities.
[0020] This embodiment extracts features from three complementary dimensions: spectral reflectance features directly preserve the physical meaning of the original spectral curve; spatial neighborhood features mine the texture structure information around the pixel; and transform domain features eliminate noise interference and enhance the effective signal through mathematical transformation.
[0021] Extracting spectral reflectance features: The spectral curve of each pixel in the hyperspectral image is directly used as the spectral reflectance feature. Specifically, it is assumed that the hyperspectral image contains... Each band represents a pixel position. Spectral reflectance eigenvector at [location] ,in Indicates the first The reflectance values of each band. Spectral reflectance features preserve the radiation response characteristics of ground objects at different wavelengths and are a fundamental information carrier for distinguishing different ground object categories. For example, in the IndianPines dataset containing 224 bands, the spectral reflectance feature of each pixel is a 224-dimensional vector; in the PaviaUniversity dataset containing 103 bands, the spectral reflectance feature of each pixel is a 103-dimensional vector.
[0022] Extracting spatial neighborhood features: Constructing multiple local windows of different sizes centered on pixels to extract gray-level co-occurrence matrix parameters as spatial neighborhood features. The gray-level co-occurrence matrix can effectively describe the texture characteristics of an image and reflect the spatial relationships between pixels.
[0023] Preferably, in some possible implementations of the embodiments of the present invention, the method for extracting spatial neighborhood features includes: simultaneously constructing a three-layer nested square local window with a side length of 2k+1, centered on a pixel, where k takes values of 1, 2, and 3, corresponding to window sizes of 3×3, 5×5, and 7×7, respectively. The design basis for choosing a three-layer nested window is that the 3×3 window captures fine texture structures, the 5×5 window takes into account both local and mesoscale information, and the 7×7 window extracts texture patterns over a larger range. The three complement each other and can comprehensively characterize spatial structural features at different scales. The pixel spectral vectors within each layer of the window are first subjected to a minimum noise separation transform (MNF), and the three components with the highest signal-to-noise ratio after the transform are selected to form three single-band grayscale images. The MNF transform can sort the data from high to low signal-to-noise ratio, and selecting the first three components can effectively suppress noise interference while retaining the main information. Gray-level co-occurrence matrices with step sizes of 1 and 2 are calculated along two orthogonal directions, horizontal and vertical, respectively, on each single-band grayscale image. The selection of two orthogonal directions is based on the fact that horizontal and vertical directions are the most common texture orientations in remote sensing images, capable of capturing key spatial structural information. The selection of two step sizes is based on the fact that a step size of 1 reflects fine-grained relationships between adjacent pixels, while a step size of 2 reflects coarser-grained relationships between pixels at intervals; combining both can characterize multi-scale texture characteristics. Three texture statistics—contrast, entropy, and inverse difference moment—are extracted from each co-occurrence matrix. Contrast reflects the drasticness of image grayscale changes; entropy reflects the complexity of image texture; and inverse difference moment reflects the uniformity of image texture. These three statistics characterize texture features from different perspectives and have good complementarity.
[0024] Redundancy is suppressed by differential calculation of identical texture statistics with the same component, direction, but different lengths. The specific calculation method is as follows: ;in and These represent the same texture statistics extracted with step sizes of 1 and 2, respectively. The difference operation can eliminate shared redundant information under both step sizes, while preserving their respective unique contributions.
[0025] Max pooling is used to retain the most salient spatial patterns for identical texture statistics in the same component but different directions. The specific calculation method is as follows: ;in Represents the most significant spatial statistics, and These represent the same texture statistics extracted in the horizontal and vertical directions, respectively. Max pooling adaptively selects the most significant directional information, avoiding human bias in direction selection.
[0026] The processed texture statistics are weighted and fused using variable weighting coefficients based on local label density adaptation. The weights decrease non-linearly as the number of known label samples around the current pixel increases. Let the number of known label samples around the current pixel be... The formula for calculating the variable weighting coefficient is: ;in, Represents the variable weighting coefficient. To adjust the parameters, this embodiment uses an empirical value of 0.5, with a range of [0.3, 0.8]. Using a logarithmic function for non-linear decreasing makes the weight changes smoother, avoiding drastic fluctuations that could affect classification stability. When there are sufficient surrounding labeled samples, the classifier can directly learn the discrimination boundary from the label information, reducing its dependence on spatial texture features; therefore, the weights should be appropriately reduced. When surrounding labeled samples are scarce, spatial texture features contribute more significantly to classification, and the weights should be kept relatively large. Finally, the fused compact texture feature vector is appended with a spatial scale identifier and a component source identifier to serve as the spatial neighborhood feature. The spatial scale identifier is used in subsequent steps to distinguish the feature contributions of different window sizes, and the component source identifier is used to trace the MNF component index corresponding to the feature.
[0027] Extracting transform domain features: After performing minimum noise separation transform on the pixel spectral vector, the top few principal components are taken as transform domain features. The MNF transform estimates the noise covariance matrix and whitens the data, transforming the original spectral data into a new coordinate system sorted by signal-to-noise ratio, so that high signal-to-noise ratio components are arranged first.
[0028] In this embodiment, the top principal components whose cumulative information content after transformation reaches a preset contribution threshold are selected as transform domain features. The preset contribution threshold is set to 99%, with a value range of [95%, 99.5%]. The rationale for this setting is that a contribution threshold that is too low will lead to the loss of effective information, while a contribution threshold that is too high will retain too many noise components. The 99% setting achieves a good balance between information preservation and noise suppression. The above three types of features are combined to form an initial heterogeneous feature set, where each type of feature contains feature vectors for all pixel locations.
[0029] It should be noted that this embodiment uses MNF transform to extract transform domain features; other embodiments may also use principal component analysis (PCA) transform, independent component analysis (ICA) transform, or wavelet transform, etc., which are not limited here.
[0030] Step S2: Construct a dedicated feature saliency mapping space for each type of feature, and generate a subset of candidate features by suppressing band correlation and enhancing inter-class dispersion.
[0031] Hyperspectral data suffers from severe band redundancy, with adjacent bands often exhibiting high correlations. Directly using all bands for classification increases computational burden and introduces redundant interference. Furthermore, not all bands contribute equally to the classification task; features with high discriminative power for class differentiation need to be selected. Therefore, this step eliminates redundancy by suppressing band correlations and retains highly discriminative features by enhancing inter-class dispersion, generating a concise subset of candidate features.
[0032] Specifically, within each feature class, the Pearson correlation coefficient matrix between bands is calculated and a dynamic correlation threshold is set. Band pairs that exceed the dynamic correlation threshold are then clustered and merged.
[0033] Suppose a certain feature class contains D bands, then the Pearson correlation coefficient matrix is a D×D symmetric matrix, and the matrix elements are calculated using the following formula: ;in Indicates the first The band and the first The correlation coefficient between each band Indicates the first The sample at the th Characteristic values of each band, Indicates the first The mean of all samples in each band. The total number of samples, Indicates the first The sample at the th Characteristic values of each band, Indicates the first The mean of all samples in each band.
[0034] The dynamic correlation threshold is determined by the product of the logarithm of the current feature dimension and the inverse of the number of labeled samples. When the feature dimension is high, the correlation threshold should be appropriately lowered to allow for more aggressive band merging and mitigate the curse of dimensionality. When there are sufficient labeled samples, the correlation threshold can be appropriately increased to retain more detailed information. The dynamic threshold enables band correlation suppression to adapt to different data characteristics, avoiding the adaptability problem of a fixed threshold in different scenarios.
[0035] Band pairs with correlation coefficients higher than the dynamic correlation threshold are merged using a hierarchical clustering method to form several band clusters. The merging strategy is as follows: for each band cluster, the weighted average of all bands within the cluster is calculated as the representative feature of the cluster, and the weights are proportional to the signal-to-noise ratio of each band.
[0036] Within the merged band clusters, the improved Fisher inter-class dispersion to intra-class clustering ratio is calculated as the significance score. The traditional Fisher criterion is defined as the ratio of inter-class divergence to intra-class divergence; this embodiment improves upon it to better suit the characteristics of hyperspectral data. Let there be a total... The first category of land cover, the Class of samples in the first The set of eigenvalues on each band cluster is ,in For the first If the number of samples in a class is given, then the formula for calculating the inter-class scatter is: ;in, Represents the inter-class dispersion. Representing the Class in The characteristic mean of each band cluster For all categories in the The overall mean across each band cluster This is the category weighting factor, used to balance the differences in the number of samples in each category. The calculation formula is: ;in This represents the total number of samples. The class weighting factor is designed to give higher weights to smaller sample classes, preventing larger sample classes from dominating feature selection results. The formula for calculating intra-class clustering is: ;in, Intra-class cohesion Representing the Class of samples in the first The first band cluster on the 1 eigenvalue, As a smoothing constant, this embodiment uses an empirical value. This is used to prevent the denominator from being zero. The improved Fisher significance score calculation formula is as follows: ;in, The significance scores are used to rank the features from highest to lowest and retain a fixed percentage, forming a subset of candidate features. The retention percentage is set at 30%. Too low a retention percentage leads to the loss of effective features, while too high a retention percentage introduces redundant features; the 30% setting strikes a good balance between feature simplification and information preservation, aligning with empirical values for practical remote sensing classification tasks. The above steps are performed on the three types of features respectively, yielding candidate subsets for spectral reflectance features, spatial neighborhood features, and transform domain features.
[0037] This step eliminates interference from redundant bands by suppressing band correlation and preserves features with high discriminative power for classification by enhancing inter-class dispersion. It effectively solves the problem of severe band redundancy in hyperspectral data and provides high-quality candidate features for subsequent feature fusion.
[0038] Step S3: Dynamically calculate the inter-source adaptation weights based on the signal-to-noise ratio statistical characteristics and label sample density of each data source, perform weighted cross-projection on the candidate feature subsets, and obtain the global fusion feature candidate pool.
[0039] Since the three types of heterogeneous features originate from different extraction methods, their data quality and discriminative capabilities vary. If a simple feature concatenation method is used for fusion, the quality differences between the data sources will be ignored, leading to low-quality features interfering with the discriminative performance of high-quality features. Therefore, it is necessary to dynamically calculate weights based on the quality characteristics of each data source and use orthogonal projection to eliminate redundant correlations between features, thereby obtaining a high-quality global fusion feature candidate pool.
[0040] Specifically, the inter-source adaptation weights are calculated, and the mean signal-to-noise ratio of pixels corresponding to each type of feature and the proportion of available labeled samples are statistically analyzed to form a quality evaluation vector.
[0041] The average pixel signal-to-noise ratio (SNR) reflects the overall quality level of this type of feature. It is calculated as follows: for each pixel location, the SNR (the ratio of signal power to noise power) of this type of feature is calculated, and then the average of all pixels is taken. Let the nth pixel be an example of this. The mean pixel signal-to-noise ratio of the class feature is .
[0042] The proportion of labeled samples reflects the usability of this type of feature in supervised learning. It is calculated as the ratio of labeled samples to the total number of samples of that feature type. Let the first... The percentage of available labeled samples for class features is .
[0043] The quality evaluation vector is defined as follows: The quality evaluation vector, after being normalized, is used as the inter-source adaptation weight. The calculation formula is as follows: ;in, Representing the Inter-source adaptation weights for class features This is the maximum-minimum normalized result of the mean signal-to-noise ratio. The maximum-min normalized result represents the proportion of available labeled samples; and As balance coefficients, this embodiment uses empirical values of 0.6 and 0.4, with value ranges of [0.4, 0.7] and [0.3, 0.6] respectively, satisfying the balance coefficient constraint condition: The signal-to-noise ratio (SNR) directly reflects data quality and plays a more important role in feature fusion, generally being assigned a higher weight. The proportion of labeled samples reflects the degree of supervised learning, and as an auxiliary factor, it is generally assigned a lower weight. Each candidate feature subset is orthogonally projected and stitched together under the guidance of inter-source adaptation weights. Redundant features are then diagonalized and removed from the stitched features to obtain a single global fusion feature candidate pool.
[0044] Preferably, in some possible implementations of the embodiments of the present invention, the specific steps of weighted cross projection include: denoting the candidate feature subsets of spectral reflectance features, spatial neighborhood features, and transform domain features as follows: , , The corresponding inter-source adaptation weight is denoted as , , For each candidate feature subset within a source, a weighted principal component analysis is performed with the inter-source adaptation weights as diagonal elements, resulting in three weighted orthogonal projection matrices. , , The formula for calculating the covariance matrix in weighted principal component analysis is: ;in for The column mean matrix, The values are 1, 2, and 3. For the sample size, This represents the transpose of a vector. The weighted covariance matrix is decomposed into eigenvalues, and the corresponding eigenvectors, arranged in descending order of eigenvalues, form the projection matrix. .
[0045] Will , , Projecting each component onto its respective weighted principal component space, the principal components whose cumulative variance contribution rate reaches a preset contribution value are retained to form the dimensionality-reduced exclusive feature matrix. , , The preset contribution value is set at 99.5%, with a range of [98%, 99.9%]. The rationale for this setting is that a contribution rate of 99.5% can retain almost all effective information while eliminating a very small amount of noise components, achieving the optimal balance between information integrity and dimensionality reduction effect.
[0046] Using the current inter-source adaptation weights as a scaling factor, amplitude modulation is applied to each dimension-reduced specific feature matrix: ;in, This represents the modulated feature matrix; amplitude modulation allows features from high-quality data sources to occupy a larger numerical range after fusion, thus contributing greater weight in subsequent classification. Gram-Schmidt orthogonalization is performed sequentially between the modulated feature matrices to ensure that the later-added feature matrices are orthogonal to the concatenated parts. The specific steps are: first, using... Based on this; then right Perform orthogonalization to obtain the same as orthogonal Finally, right and Orthogonalization yields Orthogonalization ensures linear independence between the fused features, eliminating redundant correlations between cross-source features. The cubically orthogonalized feature matrices are horizontally concatenated in a fixed order of spectral-spatial-transform domain. Highly correlated feature pairs are then eliminated through quadratic orthogonalization. A highly correlated feature pair is defined as one with an absolute correlation coefficient greater than 0.98. The threshold of 0.98 is set because a correlation coefficient higher than 0.98 indicates a high degree of overlap in the information carried by the two features, requiring only one to be retained. For each pair of highly correlated features, the feature with the higher signal-to-noise ratio is retained, while the other is eliminated, ultimately outputting a single global fusion feature candidate pool.
[0047] This step achieves differentiated processing of data sources of different quality through inter-source adaptation weights, and eliminates redundant correlations of cross-source features through orthogonal projection, effectively solving the problems of redundant interference and information conflict in the process of heterogeneous feature fusion, and improving the discrimination efficiency of fused features.
[0048] Step S4: Build a multi-classifier prototype library and assign a feature-sensitive fingerprint to each classifier. Generate a classifier-specific feature subset by matching the feature-sensitive fingerprint with the fit of each feature in the global fusion feature candidate pool.
[0049] Different types of classifiers have different learning mechanisms and feature preferences, and the contribution of the same feature to different classifiers varies significantly. Using a uniform feature subset for all classifiers ignores individual classifier differences, leading to insufficient fit between features and classifiers. Therefore, it is necessary to establish a feature sensitivity description for each classifier and, based on this, customize a unique feature subset for each classifier to achieve adaptive matching between features and classifiers.
[0050] A pre-built prototype library of multi-classifiers includes Support Vector Machines, Random Forests (RF), Extreme Learning Machines, and one-dimensional convolutional networks.
[0051] The design rationale for choosing these four classifiers is as follows: SVM has good generalization ability and is suitable for handling small sample high-dimensional data; RF is robust to feature loss and noise and can automatically evaluate feature importance; ELM has extremely fast training speed and good generalization performance; 1D-CNN can automatically learn local feature patterns of spectral sequences. The four classifiers are complementary in their learning mechanisms and feature preferences, and their integration can improve the overall stability and accuracy of classification.
[0052] The hyperparameters of each classifier are set as follows: SVM uses a radial basis function (RBF) kernel, the penalty parameter is set to 100, and the value range is [10, 1000]; the kernel parameter is set to 0.01, and the value range is [0.001, 0.1].
[0053] The number of decision trees in RF is set to 200, with a value range of [100, 500]; the maximum depth is set to unlimited; and the number of features considered in each split is set to the square root of the total number of features.
[0054] The number of hidden layer nodes in ELM is set to 1000, with a value range of [500, 2000]; the activation function is the Sigmoid function.
[0055] The 1D-CNN contains two convolutional layers with kernel sizes of 3 and 5, and the number of kernels are 32 and 64, respectively. Batch normalization and ReLU activation functions are used. The number of nodes in the fully connected layer is 128. The optimizer is Adam, and the learning rate is set to 0.001.
[0056] For each classifier, feature perturbation experiments were conducted on multiple public datasets. The sensitivity coefficients of the impact of various feature changes on classification accuracy were recorded to form a feature sensitivity fingerprint.
[0057] This embodiment uses three publicly available hyperspectral datasets—IndianPines, PaviaUniversity, and Salinas—for feature perturbation experiments. These datasets have different spatial resolutions, number of bands, and number of land cover categories, which can comprehensively reflect the classifier's performance under different data characteristics.
[0058] The method of conducting feature perturbation experiments includes: adding Gaussian noise to each feature in the training set and recording the decrease in classification accuracy, and using the normalized reciprocal of the decrease as the sensitivity coefficient of the classifier to that feature.
[0059] The specific steps are as follows: For the first The classifier and the first First, using 3D features, a classifier is trained normally on the training set to obtain the baseline classification accuracy. Then for the first Add Gaussian noise with a standard deviation of 10% of the feature's standard deviation to the dimensional feature, retrain the classifier, and obtain the perturbed classification accuracy. ; Calculation accuracy decrease The normalized decrease rate is obtained by performing maximum-min normalization on the decrease in precision of all features; the sensitivity coefficient is calculated using the following formula: ;in, This represents the sensitivity coefficient, which ranges from 0 to 1. A larger value indicates that the classifier is more sensitive to that feature, and that the feature has a greater impact on the classifier's classification performance. Represents the normalized decrease magnitude; a vector is formed from the sensitivity coefficients corresponding to all features, serving as the first... The feature-sensitive fingerprint of a classifier.
[0060] Calculate the cosine similarity between each feature in the global fusion feature candidate pool and the feature sensitivity fingerprint of each classifier; dynamically select a fixed proportion of features for each classifier according to the similarity from high to low to form a classifier’s exclusive feature subset.
[0061] The specific steps are as follows: For the first The classifier will select the th classifier from the global fusion feature candidate pool. The vector representation of a feature (a vector formed by the values of that feature across all samples) is used to calculate its cosine similarity to the feature-sensitive fingerprint. All features are sorted from high to low according to their cosine similarity, and the top fixed proportion of features are selected to form the 1st feature. Each classifier has its own subset of features. The fixed percentage is set at 50%, with a range of [30%, 70%]. Too low a percentage may miss important features, while too high a percentage will introduce redundant features; the 50% setting achieves a good balance between feature coverage and conciseness.
[0062] This step quantifies the classifier's preference for different features through feature-sensitive fingerprinting, and achieves adaptive matching between feature subsets and classifiers through fit matching, effectively solving the problem of insufficient fit between heterogeneous features and classifiers, and improving the classification performance of each classifier.
[0063] Step S5: Input the feature subsets specific to each classifier into the corresponding classifier for independent training and output the class probability. Adaptively weight and integrate the class probabilities of all classifiers through inter-source adaptation weights to obtain the final land cover category determination result.
[0064] The classification results of a single classifier may be biased and unstable. Integrating the outputs of multiple classifiers can combine the discriminative information of each classifier, improving the overall classification accuracy and robustness. Furthermore, the reliability of different classifiers is related to the quality of the data source of the features they use, requiring differentiated confidence allocation based on inter-source adaptation weights.
[0065] Each classifier's unique feature subset is input into its corresponding classifier for forward inference, resulting in an independent probability output for each type of land cover.
[0066] During the training phase, the four classifiers are trained independently using labeled samples. The training data is divided into two parts: 70% for training and 30% for validation. This division is based on the principle that a 70% training ratio ensures a sufficient number of training samples, while a 30% validation ratio effectively evaluates the model's generalization ability.
[0067] During the inference phase, the specific feature subset of the pixel to be classified is input into the corresponding classifier, and the probability distribution of the pixel belonging to each category is output. The probability distribution satisfies the category probability condition, that is, the sum of the probability distributions is 1.
[0068] Next, the ensemble trust score is calculated, with the inter-source adaptation weights of each classifier used as the ensemble trust score for that classifier. Since each classifier's unique feature subset mainly comes from one or a few classes of features, the trustworthiness of the classifier is closely related to the quality of the data source from which it uses features. The inter-source adaptation weights have comprehensively considered the signal-to-noise ratio and label sample density of each data source, effectively reflecting the quality level of the data source.
[0069] Let the first The feature subset specific to the classifier comes from the first classifier. The proportion of class features is Then the first The formula for calculating the ensemble trust of a classifier is: ;in, Represents the level of integrated trust. For the first Inter-source adaptation weights for class features The value can be 1, 2, or 3, corresponding to spectral reflectance features, spatial neighborhood features, and transform domain features, respectively. The probability outputs of all classifiers are weighted, summed, and normalized according to the ensemble confidence level to obtain the final class probability distribution, i.e., pixel location. Final category probability distribution at [location] ;in, Representing the A classifier for pixel location The output category probability distribution is used to determine the final category, which is the category with the highest probability.
[0070] It should be noted that this embodiment uses a weighted probability fusion integration method; other embodiments may also use voting methods, stacking generalization, etc., and are not limited here. This step adaptively weights the outputs of multiple classifiers by adapting weights between sources, so that classifiers corresponding to high-quality data sources receive higher integration weights. This effectively solves the problem of traditional integration methods ignoring the differences in classifier reliability, and improves the accuracy and stability of the final classification result.
[0071] This embodiment effectively solves the problem of insufficient adaptability between heterogeneous features and classifiers in hyperspectral image classification by constructing a dedicated feature saliency mapping space to filter high-discrimination features, achieving high-quality fusion of heterogeneous features through inter-source adaptation weights, establishing an adaptive matching mechanism between features and classifiers through feature-sensitive fingerprints, and integrating the discriminative information of multiple classifiers through weighted integration. This improves classification accuracy, model stability, and cross-scenario generalization ability, providing reliable technical support for practical applications such as precision agriculture monitoring and ecological environment assessment.
[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the parameter scheme may be modified, or some technical features may be equivalently replaced, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation, characterized in that, This includes: Step S1: Simultaneously extracting spectral reflectance features, spatial neighborhood features, and transform domain features from the hyperspectral image to form an initial heterogeneous feature set; step S2: Construct a dedicated feature saliency mapping space for each type of feature, and generate a subset of candidate features by suppressing band correlation and enhancing inter-class dispersion; Step S3: Calculate the inter-source adaptation weights dynamically based on the signal-to-noise ratio statistical characteristics and label sample density of each data source, and perform weighted cross-projection on the subset of candidate features to obtain a global fusion feature candidate pool. Step S4: Build a multi-classifier prototype library and assign a feature-sensitive fingerprint to each classifier. Generate a classifier-specific feature subset by matching the feature-sensitive fingerprint with the fit of each feature in the global fusion feature candidate pool. Step S5: Input the feature subsets specific to each classifier into the corresponding classifier for independent training and output the class probability. Adaptively weight and integrate the class probabilities of all classifiers through inter-source adaptation weights to obtain the final land cover category determination result.
2. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 1, characterized in that, Step S1 includes: directly using the spectral curve of each pixel in the hyperspectral image as the spectral reflectance feature; constructing local windows of multiple sizes centered on the pixel to extract the gray-level co-occurrence matrix parameters as spatial neighborhood features; and taking the first few principal components as transform domain features after performing minimum noise separation transformation on the pixel spectral vector.
3. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 2, characterized in that, Step S2 includes: calculating the inter-band Pearson correlation coefficient matrix within each feature class and setting a dynamic correlation threshold; clustering and merging band pairs that are higher than the dynamic correlation threshold to obtain band clusters; calculating the improved Fisher inter-class dispersion to intra-class clustering ratio as a significance score within the merged band clusters; sorting the significance scores from high to low and retaining them in a fixed proportion to form a unique subset of candidate features.
4. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 1, characterized in that, Step S3 includes: calculating the mean pixel signal-to-noise ratio and the proportion of available label samples corresponding to each type of feature to form a quality evaluation vector; normalizing the quality evaluation vector and using it as the inter-source adaptation weight; performing orthogonal projection stitching on each candidate feature subset under the guidance of the inter-source adaptation weight; performing redundant feature diagonalization removal on the stitched features to obtain a single global fusion feature candidate pool.
5. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 4, characterized in that, Step S4 includes: pre-constructing a multi-classifier prototype library containing support vector machines, random forests, extreme learning machines, and one-dimensional convolutional networks; conducting feature perturbation experiments on multiple public datasets for each classifier, recording the sensitivity coefficients of various feature changes on classification accuracy, and forming feature sensitivity fingerprints; calculating the cosine similarity between each feature in the global fusion feature candidate pool and the feature sensitivity fingerprints of each classifier; dynamically selecting a fixed proportion of features for each classifier according to the similarity from high to low, forming a unique feature subset for each classifier.
6. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 5, characterized in that, Step S5 includes: inputting the specific feature subsets of each classifier into the corresponding classifier for forward inference. Forward inference includes a training phase and an inference phase. In the training phase, the four classifiers are trained independently using labeled samples. The training data is divided into training data and validation data according to the proportion of labeled samples. In the inference phase, the specific feature subsets of the pixels to be classified are input into the corresponding classifier, and the probability distribution of the pixel belonging to each category is output as the independent probability of each land cover category, and the probability distribution satisfies the category probability condition. The source-to-source adaptation weights of each classifier are used as the ensemble confidence of the classifier. The probability outputs of all classifiers are weighted, summed, and normalized according to the ensemble confidence, and the category with the highest category probability is taken as the final land cover category determination result.
7. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 6, characterized in that, The method of conducting feature perturbation experiments includes: adding Gaussian noise to each feature in the training set and recording the decrease in classification accuracy, and using the normalized reciprocal of the decrease as the sensitivity coefficient of the classifier to that feature.
8. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 7, characterized in that, The dynamic correlation threshold is determined by the product of the logarithmic function of the current feature dimension and the inverse of the number of labeled samples.
9. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 8, characterized in that, Multiple local windows of different sizes are constructed around pixels to extract gray-level co-occurrence matrix parameters as spatial neighborhood features. This includes: simultaneously constructing three nested square local windows with a side length of 2k+1, where k takes values of 1, 2, and 3, centered on the pixel; performing minimum noise separation transformation on the pixel spectral vector within each window, and selecting the three components with the highest signal-to-noise ratio after the transformation to form three single-band gray-level images; calculating gray-level co-occurrence matrices with step sizes of 1 and 2 along two orthogonal directions (horizontal and vertical) on each single-band gray-level image; extracting contrast, entropy, and inverse difference moment texture statistics from each co-occurrence matrix; performing differential redundancy suppression on identical texture statistics with the same component, direction, and different step lengths, and performing max pooling on identical texture statistics with the same component but different directions to retain the most significant spatial pattern; weighting and fusion of the processed texture statistics using adaptive variable weighting coefficients based on local label density, with the weights decreasing non-linearly as the number of known label samples around the current pixel increases; and finally, attaching spatial scale and component source labels to the fused compact texture feature vector as spatial neighborhood features.
10. The hyperspectral adaptive ensemble classification method based on heterogeneous feature adaptation according to claim 9, characterized in that, Each candidate feature subset is orthogonally projected and stitched together under the guidance of inter-source adaptation weights. Redundant features are then eliminated by diagonalization of the stitched features to obtain a single global fusion feature candidate pool. This process includes: assigning corresponding inter-source adaptation weights to candidate feature subsets of spectral reflectance, spatial neighborhood, and transform domain features; performing weighted principal component analysis (PCA) with the inter-source adaptation weights as diagonal elements on each candidate feature subset to obtain three weighted orthogonal projection matrices; and projecting each candidate feature subset onto its respective weighted principal component space, retaining the cumulative variance contribution. The principal components whose contribution rates reach the preset contribution values form the dimensionality-reduced exclusive feature matrix; using the current inter-source adaptation weight as the scale factor, the amplitude of each dimensionality-reduced exclusive feature matrix is modulated, and Gram-Schmidt orthogonalization is performed sequentially between the modulated feature matrices to make the later-accessed feature matrices orthogonal to the concatenated parts; the cubically orthogonalized feature matrices are horizontally concatenated in a fixed order of spectral-spatial-transform domain, and the highly correlated feature pairs that appear after concatenation are removed by quadratic orthogonalization, and finally a single global fusion feature candidate pool is output.